1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
13 // Note that this pass does the value numbering itself; it does not use the
14 // ValueNumbering analysis passes.
16 //===----------------------------------------------------------------------===//
18 #include "llvm/Transforms/Scalar.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/Hashing.h"
22 #include "llvm/ADT/MapVector.h"
23 #include "llvm/ADT/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/Local.h"
49 #include "llvm/Transforms/Utils/SSAUpdater.h"
52 using namespace PatternMatch;
54 #define DEBUG_TYPE "gvn"
56 STATISTIC(NumGVNInstr, "Number of instructions deleted");
57 STATISTIC(NumGVNLoad, "Number of loads deleted");
58 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
59 STATISTIC(NumGVNBlocks, "Number of blocks merged");
60 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
61 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
62 STATISTIC(NumPRELoad, "Number of loads PRE'd");
64 static cl::opt<bool> EnablePRE("enable-pre",
65 cl::init(true), cl::Hidden);
66 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
68 // Maximum allowed recursion depth.
69 static cl::opt<uint32_t>
70 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
71 cl::desc("Max recurse depth (default = 1000)"));
73 //===----------------------------------------------------------------------===//
75 //===----------------------------------------------------------------------===//
77 /// This class holds the mapping between values and value numbers. It is used
78 /// as an efficient mechanism to determine the expression-wise equivalence of
84 SmallVector<uint32_t, 4> varargs;
86 Expression(uint32_t o = ~2U) : opcode(o) { }
88 bool operator==(const Expression &other) const {
89 if (opcode != other.opcode)
91 if (opcode == ~0U || opcode == ~1U)
93 if (type != other.type)
95 if (varargs != other.varargs)
100 friend hash_code hash_value(const Expression &Value) {
101 return hash_combine(Value.opcode, Value.type,
102 hash_combine_range(Value.varargs.begin(),
103 Value.varargs.end()));
108 DenseMap<Value*, uint32_t> valueNumbering;
109 DenseMap<Expression, uint32_t> expressionNumbering;
111 MemoryDependenceAnalysis *MD;
114 uint32_t nextValueNumber;
116 Expression create_expression(Instruction* I);
117 Expression create_cmp_expression(unsigned Opcode,
118 CmpInst::Predicate Predicate,
119 Value *LHS, Value *RHS);
120 Expression create_extractvalue_expression(ExtractValueInst* EI);
121 uint32_t lookup_or_add_call(CallInst* C);
123 ValueTable() : nextValueNumber(1) { }
124 uint32_t lookup_or_add(Value *V);
125 uint32_t lookup(Value *V) const;
126 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
127 Value *LHS, Value *RHS);
128 void add(Value *V, uint32_t num);
130 void erase(Value *v);
131 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
132 AliasAnalysis *getAliasAnalysis() const { return AA; }
133 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
134 void setDomTree(DominatorTree* D) { DT = D; }
135 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
136 void verifyRemoved(const Value *) const;
141 template <> struct DenseMapInfo<Expression> {
142 static inline Expression getEmptyKey() {
146 static inline Expression getTombstoneKey() {
150 static unsigned getHashValue(const Expression e) {
151 using llvm::hash_value;
152 return static_cast<unsigned>(hash_value(e));
154 static bool isEqual(const Expression &LHS, const Expression &RHS) {
161 //===----------------------------------------------------------------------===//
162 // ValueTable Internal Functions
163 //===----------------------------------------------------------------------===//
165 Expression ValueTable::create_expression(Instruction *I) {
167 e.type = I->getType();
168 e.opcode = I->getOpcode();
169 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
171 e.varargs.push_back(lookup_or_add(*OI));
172 if (I->isCommutative()) {
173 // Ensure that commutative instructions that only differ by a permutation
174 // of their operands get the same value number by sorting the operand value
175 // numbers. Since all commutative instructions have two operands it is more
176 // efficient to sort by hand rather than using, say, std::sort.
177 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
178 if (e.varargs[0] > e.varargs[1])
179 std::swap(e.varargs[0], e.varargs[1]);
182 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
183 // Sort the operand value numbers so x<y and y>x get the same value number.
184 CmpInst::Predicate Predicate = C->getPredicate();
185 if (e.varargs[0] > e.varargs[1]) {
186 std::swap(e.varargs[0], e.varargs[1]);
187 Predicate = CmpInst::getSwappedPredicate(Predicate);
189 e.opcode = (C->getOpcode() << 8) | Predicate;
190 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
191 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
193 e.varargs.push_back(*II);
199 Expression ValueTable::create_cmp_expression(unsigned Opcode,
200 CmpInst::Predicate Predicate,
201 Value *LHS, Value *RHS) {
202 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
203 "Not a comparison!");
205 e.type = CmpInst::makeCmpResultType(LHS->getType());
206 e.varargs.push_back(lookup_or_add(LHS));
207 e.varargs.push_back(lookup_or_add(RHS));
209 // Sort the operand value numbers so x<y and y>x get the same value number.
210 if (e.varargs[0] > e.varargs[1]) {
211 std::swap(e.varargs[0], e.varargs[1]);
212 Predicate = CmpInst::getSwappedPredicate(Predicate);
214 e.opcode = (Opcode << 8) | Predicate;
218 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
219 assert(EI && "Not an ExtractValueInst?");
221 e.type = EI->getType();
224 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
225 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
226 // EI might be an extract from one of our recognised intrinsics. If it
227 // is we'll synthesize a semantically equivalent expression instead on
228 // an extract value expression.
229 switch (I->getIntrinsicID()) {
230 case Intrinsic::sadd_with_overflow:
231 case Intrinsic::uadd_with_overflow:
232 e.opcode = Instruction::Add;
234 case Intrinsic::ssub_with_overflow:
235 case Intrinsic::usub_with_overflow:
236 e.opcode = Instruction::Sub;
238 case Intrinsic::smul_with_overflow:
239 case Intrinsic::umul_with_overflow:
240 e.opcode = Instruction::Mul;
247 // Intrinsic recognized. Grab its args to finish building the expression.
248 assert(I->getNumArgOperands() == 2 &&
249 "Expect two args for recognised intrinsics.");
250 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
251 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
256 // Not a recognised intrinsic. Fall back to producing an extract value
258 e.opcode = EI->getOpcode();
259 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
261 e.varargs.push_back(lookup_or_add(*OI));
263 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
265 e.varargs.push_back(*II);
270 //===----------------------------------------------------------------------===//
271 // ValueTable External Functions
272 //===----------------------------------------------------------------------===//
274 /// add - Insert a value into the table with a specified value number.
275 void ValueTable::add(Value *V, uint32_t num) {
276 valueNumbering.insert(std::make_pair(V, num));
279 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
280 if (AA->doesNotAccessMemory(C)) {
281 Expression exp = create_expression(C);
282 uint32_t &e = expressionNumbering[exp];
283 if (!e) e = nextValueNumber++;
284 valueNumbering[C] = e;
286 } else if (AA->onlyReadsMemory(C)) {
287 Expression exp = create_expression(C);
288 uint32_t &e = expressionNumbering[exp];
290 e = nextValueNumber++;
291 valueNumbering[C] = e;
295 e = nextValueNumber++;
296 valueNumbering[C] = e;
300 MemDepResult local_dep = MD->getDependency(C);
302 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
303 valueNumbering[C] = nextValueNumber;
304 return nextValueNumber++;
307 if (local_dep.isDef()) {
308 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
310 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
311 valueNumbering[C] = nextValueNumber;
312 return nextValueNumber++;
315 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
316 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
317 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
319 valueNumbering[C] = nextValueNumber;
320 return nextValueNumber++;
324 uint32_t v = lookup_or_add(local_cdep);
325 valueNumbering[C] = v;
330 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
331 MD->getNonLocalCallDependency(CallSite(C));
332 // FIXME: Move the checking logic to MemDep!
333 CallInst* cdep = nullptr;
335 // Check to see if we have a single dominating call instruction that is
337 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
338 const NonLocalDepEntry *I = &deps[i];
339 if (I->getResult().isNonLocal())
342 // We don't handle non-definitions. If we already have a call, reject
343 // instruction dependencies.
344 if (!I->getResult().isDef() || cdep != nullptr) {
349 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
350 // FIXME: All duplicated with non-local case.
351 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
352 cdep = NonLocalDepCall;
361 valueNumbering[C] = nextValueNumber;
362 return nextValueNumber++;
365 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
366 valueNumbering[C] = nextValueNumber;
367 return nextValueNumber++;
369 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
370 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
371 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
373 valueNumbering[C] = nextValueNumber;
374 return nextValueNumber++;
378 uint32_t v = lookup_or_add(cdep);
379 valueNumbering[C] = v;
383 valueNumbering[C] = nextValueNumber;
384 return nextValueNumber++;
388 /// lookup_or_add - Returns the value number for the specified value, assigning
389 /// it a new number if it did not have one before.
390 uint32_t ValueTable::lookup_or_add(Value *V) {
391 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
392 if (VI != valueNumbering.end())
395 if (!isa<Instruction>(V)) {
396 valueNumbering[V] = nextValueNumber;
397 return nextValueNumber++;
400 Instruction* I = cast<Instruction>(V);
402 switch (I->getOpcode()) {
403 case Instruction::Call:
404 return lookup_or_add_call(cast<CallInst>(I));
405 case Instruction::Add:
406 case Instruction::FAdd:
407 case Instruction::Sub:
408 case Instruction::FSub:
409 case Instruction::Mul:
410 case Instruction::FMul:
411 case Instruction::UDiv:
412 case Instruction::SDiv:
413 case Instruction::FDiv:
414 case Instruction::URem:
415 case Instruction::SRem:
416 case Instruction::FRem:
417 case Instruction::Shl:
418 case Instruction::LShr:
419 case Instruction::AShr:
420 case Instruction::And:
421 case Instruction::Or:
422 case Instruction::Xor:
423 case Instruction::ICmp:
424 case Instruction::FCmp:
425 case Instruction::Trunc:
426 case Instruction::ZExt:
427 case Instruction::SExt:
428 case Instruction::FPToUI:
429 case Instruction::FPToSI:
430 case Instruction::UIToFP:
431 case Instruction::SIToFP:
432 case Instruction::FPTrunc:
433 case Instruction::FPExt:
434 case Instruction::PtrToInt:
435 case Instruction::IntToPtr:
436 case Instruction::BitCast:
437 case Instruction::Select:
438 case Instruction::ExtractElement:
439 case Instruction::InsertElement:
440 case Instruction::ShuffleVector:
441 case Instruction::InsertValue:
442 case Instruction::GetElementPtr:
443 exp = create_expression(I);
445 case Instruction::ExtractValue:
446 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
449 valueNumbering[V] = nextValueNumber;
450 return nextValueNumber++;
453 uint32_t& e = expressionNumbering[exp];
454 if (!e) e = nextValueNumber++;
455 valueNumbering[V] = e;
459 /// lookup - Returns the value number of the specified value. Fails if
460 /// the value has not yet been numbered.
461 uint32_t ValueTable::lookup(Value *V) const {
462 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
463 assert(VI != valueNumbering.end() && "Value not numbered?");
467 /// lookup_or_add_cmp - Returns the value number of the given comparison,
468 /// assigning it a new number if it did not have one before. Useful when
469 /// we deduced the result of a comparison, but don't immediately have an
470 /// instruction realizing that comparison to hand.
471 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
472 CmpInst::Predicate Predicate,
473 Value *LHS, Value *RHS) {
474 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
475 uint32_t& e = expressionNumbering[exp];
476 if (!e) e = nextValueNumber++;
480 /// clear - Remove all entries from the ValueTable.
481 void ValueTable::clear() {
482 valueNumbering.clear();
483 expressionNumbering.clear();
487 /// erase - Remove a value from the value numbering.
488 void ValueTable::erase(Value *V) {
489 valueNumbering.erase(V);
492 /// verifyRemoved - Verify that the value is removed from all internal data
494 void ValueTable::verifyRemoved(const Value *V) const {
495 for (DenseMap<Value*, uint32_t>::const_iterator
496 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
497 assert(I->first != V && "Inst still occurs in value numbering map!");
501 //===----------------------------------------------------------------------===//
503 //===----------------------------------------------------------------------===//
507 struct AvailableValueInBlock {
508 /// BB - The basic block in question.
511 SimpleVal, // A simple offsetted value that is accessed.
512 LoadVal, // A value produced by a load.
513 MemIntrin, // A memory intrinsic which is loaded from.
514 UndefVal // A UndefValue representing a value from dead block (which
515 // is not yet physically removed from the CFG).
518 /// V - The value that is live out of the block.
519 PointerIntPair<Value *, 2, ValType> Val;
521 /// Offset - The byte offset in Val that is interesting for the load query.
524 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
525 unsigned Offset = 0) {
526 AvailableValueInBlock Res;
528 Res.Val.setPointer(V);
529 Res.Val.setInt(SimpleVal);
534 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
535 unsigned Offset = 0) {
536 AvailableValueInBlock Res;
538 Res.Val.setPointer(MI);
539 Res.Val.setInt(MemIntrin);
544 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
545 unsigned Offset = 0) {
546 AvailableValueInBlock Res;
548 Res.Val.setPointer(LI);
549 Res.Val.setInt(LoadVal);
554 static AvailableValueInBlock getUndef(BasicBlock *BB) {
555 AvailableValueInBlock Res;
557 Res.Val.setPointer(nullptr);
558 Res.Val.setInt(UndefVal);
563 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
564 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
565 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
566 bool isUndefValue() const { return Val.getInt() == UndefVal; }
568 Value *getSimpleValue() const {
569 assert(isSimpleValue() && "Wrong accessor");
570 return Val.getPointer();
573 LoadInst *getCoercedLoadValue() const {
574 assert(isCoercedLoadValue() && "Wrong accessor");
575 return cast<LoadInst>(Val.getPointer());
578 MemIntrinsic *getMemIntrinValue() const {
579 assert(isMemIntrinValue() && "Wrong accessor");
580 return cast<MemIntrinsic>(Val.getPointer());
583 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
584 /// defined here to the specified type. This handles various coercion cases.
585 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const;
588 class GVN : public FunctionPass {
590 MemoryDependenceAnalysis *MD;
592 const DataLayout *DL;
593 const TargetLibraryInfo *TLI;
594 SetVector<BasicBlock *> DeadBlocks;
598 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
599 /// have that value number. Use findLeader to query it.
600 struct LeaderTableEntry {
602 const BasicBlock *BB;
603 LeaderTableEntry *Next;
605 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
606 BumpPtrAllocator TableAllocator;
608 SmallVector<Instruction*, 8> InstrsToErase;
610 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
611 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
612 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
615 static char ID; // Pass identification, replacement for typeid
616 explicit GVN(bool noloads = false)
617 : FunctionPass(ID), NoLoads(noloads), MD(nullptr) {
618 initializeGVNPass(*PassRegistry::getPassRegistry());
621 bool runOnFunction(Function &F) override;
623 /// markInstructionForDeletion - This removes the specified instruction from
624 /// our various maps and marks it for deletion.
625 void markInstructionForDeletion(Instruction *I) {
627 InstrsToErase.push_back(I);
630 const DataLayout *getDataLayout() const { return DL; }
631 DominatorTree &getDominatorTree() const { return *DT; }
632 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
633 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
635 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
636 /// its value number.
637 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
638 LeaderTableEntry &Curr = LeaderTable[N];
645 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
648 Node->Next = Curr.Next;
652 /// removeFromLeaderTable - Scan the list of values corresponding to a given
653 /// value number, and remove the given instruction if encountered.
654 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
655 LeaderTableEntry* Prev = nullptr;
656 LeaderTableEntry* Curr = &LeaderTable[N];
658 while (Curr->Val != I || Curr->BB != BB) {
664 Prev->Next = Curr->Next;
670 LeaderTableEntry* Next = Curr->Next;
671 Curr->Val = Next->Val;
673 Curr->Next = Next->Next;
678 // List of critical edges to be split between iterations.
679 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
681 // This transformation requires dominator postdominator info
682 void getAnalysisUsage(AnalysisUsage &AU) const override {
683 AU.addRequired<DominatorTreeWrapperPass>();
684 AU.addRequired<TargetLibraryInfo>();
686 AU.addRequired<MemoryDependenceAnalysis>();
687 AU.addRequired<AliasAnalysis>();
689 AU.addPreserved<DominatorTreeWrapperPass>();
690 AU.addPreserved<AliasAnalysis>();
694 // Helper fuctions of redundant load elimination
695 bool processLoad(LoadInst *L);
696 bool processNonLocalLoad(LoadInst *L);
697 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
698 AvailValInBlkVect &ValuesPerBlock,
699 UnavailBlkVect &UnavailableBlocks);
700 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
701 UnavailBlkVect &UnavailableBlocks);
703 // Other helper routines
704 bool processInstruction(Instruction *I);
705 bool processBlock(BasicBlock *BB);
706 void dump(DenseMap<uint32_t, Value*> &d);
707 bool iterateOnFunction(Function &F);
708 bool performPRE(Function &F);
709 Value *findLeader(const BasicBlock *BB, uint32_t num);
710 void cleanupGlobalSets();
711 void verifyRemoved(const Instruction *I) const;
712 bool splitCriticalEdges();
713 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
714 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
715 const BasicBlockEdge &Root);
716 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
717 bool processFoldableCondBr(BranchInst *BI);
718 void addDeadBlock(BasicBlock *BB);
719 void assignValNumForDeadCode();
725 // createGVNPass - The public interface to this file...
726 FunctionPass *llvm::createGVNPass(bool NoLoads) {
727 return new GVN(NoLoads);
730 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
731 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
732 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
733 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
734 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
735 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
737 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
738 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
740 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
741 E = d.end(); I != E; ++I) {
742 errs() << I->first << "\n";
749 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
750 /// we're analyzing is fully available in the specified block. As we go, keep
751 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
752 /// map is actually a tri-state map with the following values:
753 /// 0) we know the block *is not* fully available.
754 /// 1) we know the block *is* fully available.
755 /// 2) we do not know whether the block is fully available or not, but we are
756 /// currently speculating that it will be.
757 /// 3) we are speculating for this block and have used that to speculate for
759 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
760 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
761 uint32_t RecurseDepth) {
762 if (RecurseDepth > MaxRecurseDepth)
765 // Optimistically assume that the block is fully available and check to see
766 // if we already know about this block in one lookup.
767 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
768 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
770 // If the entry already existed for this block, return the precomputed value.
772 // If this is a speculative "available" value, mark it as being used for
773 // speculation of other blocks.
774 if (IV.first->second == 2)
775 IV.first->second = 3;
776 return IV.first->second != 0;
779 // Otherwise, see if it is fully available in all predecessors.
780 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
782 // If this block has no predecessors, it isn't live-in here.
784 goto SpeculationFailure;
786 for (; PI != PE; ++PI)
787 // If the value isn't fully available in one of our predecessors, then it
788 // isn't fully available in this block either. Undo our previous
789 // optimistic assumption and bail out.
790 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
791 goto SpeculationFailure;
795 // SpeculationFailure - If we get here, we found out that this is not, after
796 // all, a fully-available block. We have a problem if we speculated on this and
797 // used the speculation to mark other blocks as available.
799 char &BBVal = FullyAvailableBlocks[BB];
801 // If we didn't speculate on this, just return with it set to false.
807 // If we did speculate on this value, we could have blocks set to 1 that are
808 // incorrect. Walk the (transitive) successors of this block and mark them as
810 SmallVector<BasicBlock*, 32> BBWorklist;
811 BBWorklist.push_back(BB);
814 BasicBlock *Entry = BBWorklist.pop_back_val();
815 // Note that this sets blocks to 0 (unavailable) if they happen to not
816 // already be in FullyAvailableBlocks. This is safe.
817 char &EntryVal = FullyAvailableBlocks[Entry];
818 if (EntryVal == 0) continue; // Already unavailable.
820 // Mark as unavailable.
823 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
824 } while (!BBWorklist.empty());
830 /// CanCoerceMustAliasedValueToLoad - Return true if
831 /// CoerceAvailableValueToLoadType will succeed.
832 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
834 const DataLayout &DL) {
835 // If the loaded or stored value is an first class array or struct, don't try
836 // to transform them. We need to be able to bitcast to integer.
837 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
838 StoredVal->getType()->isStructTy() ||
839 StoredVal->getType()->isArrayTy())
842 // The store has to be at least as big as the load.
843 if (DL.getTypeSizeInBits(StoredVal->getType()) <
844 DL.getTypeSizeInBits(LoadTy))
850 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
851 /// then a load from a must-aliased pointer of a different type, try to coerce
852 /// the stored value. LoadedTy is the type of the load we want to replace and
853 /// InsertPt is the place to insert new instructions.
855 /// If we can't do it, return null.
856 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
858 Instruction *InsertPt,
859 const DataLayout &DL) {
860 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
863 // If this is already the right type, just return it.
864 Type *StoredValTy = StoredVal->getType();
866 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
867 uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
869 // If the store and reload are the same size, we can always reuse it.
870 if (StoreSize == LoadSize) {
871 // Pointer to Pointer -> use bitcast.
872 if (StoredValTy->getScalarType()->isPointerTy() &&
873 LoadedTy->getScalarType()->isPointerTy())
874 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
876 // Convert source pointers to integers, which can be bitcast.
877 if (StoredValTy->getScalarType()->isPointerTy()) {
878 StoredValTy = DL.getIntPtrType(StoredValTy);
879 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
882 Type *TypeToCastTo = LoadedTy;
883 if (TypeToCastTo->getScalarType()->isPointerTy())
884 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
886 if (StoredValTy != TypeToCastTo)
887 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
889 // Cast to pointer if the load needs a pointer type.
890 if (LoadedTy->getScalarType()->isPointerTy())
891 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
896 // If the loaded value is smaller than the available value, then we can
897 // extract out a piece from it. If the available value is too small, then we
898 // can't do anything.
899 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
901 // Convert source pointers to integers, which can be manipulated.
902 if (StoredValTy->getScalarType()->isPointerTy()) {
903 StoredValTy = DL.getIntPtrType(StoredValTy);
904 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
907 // Convert vectors and fp to integer, which can be manipulated.
908 if (!StoredValTy->isIntegerTy()) {
909 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
910 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
913 // If this is a big-endian system, we need to shift the value down to the low
914 // bits so that a truncate will work.
915 if (DL.isBigEndian()) {
916 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
917 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
920 // Truncate the integer to the right size now.
921 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
922 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
924 if (LoadedTy == NewIntTy)
927 // If the result is a pointer, inttoptr.
928 if (LoadedTy->getScalarType()->isPointerTy())
929 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
931 // Otherwise, bitcast.
932 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
935 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
936 /// memdep query of a load that ends up being a clobbering memory write (store,
937 /// memset, memcpy, memmove). This means that the write *may* provide bits used
938 /// by the load but we can't be sure because the pointers don't mustalias.
940 /// Check this case to see if there is anything more we can do before we give
941 /// up. This returns -1 if we have to give up, or a byte number in the stored
942 /// value of the piece that feeds the load.
943 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
945 uint64_t WriteSizeInBits,
946 const DataLayout &DL) {
947 // If the loaded or stored value is a first class array or struct, don't try
948 // to transform them. We need to be able to bitcast to integer.
949 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
952 int64_t StoreOffset = 0, LoadOffset = 0;
953 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&DL);
954 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &DL);
955 if (StoreBase != LoadBase)
958 // If the load and store are to the exact same address, they should have been
959 // a must alias. AA must have gotten confused.
960 // FIXME: Study to see if/when this happens. One case is forwarding a memset
961 // to a load from the base of the memset.
963 if (LoadOffset == StoreOffset) {
964 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
965 << "Base = " << *StoreBase << "\n"
966 << "Store Ptr = " << *WritePtr << "\n"
967 << "Store Offs = " << StoreOffset << "\n"
968 << "Load Ptr = " << *LoadPtr << "\n";
973 // If the load and store don't overlap at all, the store doesn't provide
974 // anything to the load. In this case, they really don't alias at all, AA
975 // must have gotten confused.
976 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
978 if ((WriteSizeInBits & 7) | (LoadSize & 7))
980 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
984 bool isAAFailure = false;
985 if (StoreOffset < LoadOffset)
986 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
988 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
992 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
993 << "Base = " << *StoreBase << "\n"
994 << "Store Ptr = " << *WritePtr << "\n"
995 << "Store Offs = " << StoreOffset << "\n"
996 << "Load Ptr = " << *LoadPtr << "\n";
1002 // If the Load isn't completely contained within the stored bits, we don't
1003 // have all the bits to feed it. We could do something crazy in the future
1004 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1006 if (StoreOffset > LoadOffset ||
1007 StoreOffset+StoreSize < LoadOffset+LoadSize)
1010 // Okay, we can do this transformation. Return the number of bytes into the
1011 // store that the load is.
1012 return LoadOffset-StoreOffset;
1015 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1016 /// memdep query of a load that ends up being a clobbering store.
1017 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1019 const DataLayout &DL) {
1020 // Cannot handle reading from store of first-class aggregate yet.
1021 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1022 DepSI->getValueOperand()->getType()->isArrayTy())
1025 Value *StorePtr = DepSI->getPointerOperand();
1026 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1027 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1028 StorePtr, StoreSize, DL);
1031 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
1032 /// memdep query of a load that ends up being clobbered by another load. See if
1033 /// the other load can feed into the second load.
1034 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1035 LoadInst *DepLI, const DataLayout &DL){
1036 // Cannot handle reading from store of first-class aggregate yet.
1037 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1040 Value *DepPtr = DepLI->getPointerOperand();
1041 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
1042 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
1043 if (R != -1) return R;
1045 // If we have a load/load clobber an DepLI can be widened to cover this load,
1046 // then we should widen it!
1047 int64_t LoadOffs = 0;
1048 const Value *LoadBase =
1049 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &DL);
1050 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1052 unsigned Size = MemoryDependenceAnalysis::
1053 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, DL);
1054 if (Size == 0) return -1;
1056 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
1061 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1063 const DataLayout &DL) {
1064 // If the mem operation is a non-constant size, we can't handle it.
1065 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1066 if (!SizeCst) return -1;
1067 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1069 // If this is memset, we just need to see if the offset is valid in the size
1071 if (MI->getIntrinsicID() == Intrinsic::memset)
1072 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1075 // If we have a memcpy/memmove, the only case we can handle is if this is a
1076 // copy from constant memory. In that case, we can read directly from the
1078 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1080 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1081 if (!Src) return -1;
1083 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &DL));
1084 if (!GV || !GV->isConstant()) return -1;
1086 // See if the access is within the bounds of the transfer.
1087 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1088 MI->getDest(), MemSizeInBits, DL);
1092 unsigned AS = Src->getType()->getPointerAddressSpace();
1093 // Otherwise, see if we can constant fold a load from the constant with the
1094 // offset applied as appropriate.
1095 Src = ConstantExpr::getBitCast(Src,
1096 Type::getInt8PtrTy(Src->getContext(), AS));
1097 Constant *OffsetCst =
1098 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1099 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1100 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1101 if (ConstantFoldLoadFromConstPtr(Src, &DL))
1107 /// GetStoreValueForLoad - This function is called when we have a
1108 /// memdep query of a load that ends up being a clobbering store. This means
1109 /// that the store provides bits used by the load but we the pointers don't
1110 /// mustalias. Check this case to see if there is anything more we can do
1111 /// before we give up.
1112 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1114 Instruction *InsertPt, const DataLayout &DL){
1115 LLVMContext &Ctx = SrcVal->getType()->getContext();
1117 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1118 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
1120 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1122 // Compute which bits of the stored value are being used by the load. Convert
1123 // to an integer type to start with.
1124 if (SrcVal->getType()->getScalarType()->isPointerTy())
1125 SrcVal = Builder.CreatePtrToInt(SrcVal,
1126 DL.getIntPtrType(SrcVal->getType()));
1127 if (!SrcVal->getType()->isIntegerTy())
1128 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1130 // Shift the bits to the least significant depending on endianness.
1132 if (DL.isLittleEndian())
1133 ShiftAmt = Offset*8;
1135 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1138 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1140 if (LoadSize != StoreSize)
1141 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1143 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, DL);
1146 /// GetLoadValueForLoad - This function is called when we have a
1147 /// memdep query of a load that ends up being a clobbering load. This means
1148 /// that the load *may* provide bits used by the load but we can't be sure
1149 /// because the pointers don't mustalias. Check this case to see if there is
1150 /// anything more we can do before we give up.
1151 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1152 Type *LoadTy, Instruction *InsertPt,
1154 const DataLayout &DL = *gvn.getDataLayout();
1155 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1156 // widen SrcVal out to a larger load.
1157 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
1158 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1159 if (Offset+LoadSize > SrcValSize) {
1160 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1161 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1162 // If we have a load/load clobber an DepLI can be widened to cover this
1163 // load, then we should widen it to the next power of 2 size big enough!
1164 unsigned NewLoadSize = Offset+LoadSize;
1165 if (!isPowerOf2_32(NewLoadSize))
1166 NewLoadSize = NextPowerOf2(NewLoadSize);
1168 Value *PtrVal = SrcVal->getPointerOperand();
1170 // Insert the new load after the old load. This ensures that subsequent
1171 // memdep queries will find the new load. We can't easily remove the old
1172 // load completely because it is already in the value numbering table.
1173 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1175 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1176 DestPTy = PointerType::get(DestPTy,
1177 PtrVal->getType()->getPointerAddressSpace());
1178 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1179 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1180 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1181 NewLoad->takeName(SrcVal);
1182 NewLoad->setAlignment(SrcVal->getAlignment());
1184 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1185 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1187 // Replace uses of the original load with the wider load. On a big endian
1188 // system, we need to shift down to get the relevant bits.
1189 Value *RV = NewLoad;
1190 if (DL.isBigEndian())
1191 RV = Builder.CreateLShr(RV,
1192 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1193 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1194 SrcVal->replaceAllUsesWith(RV);
1196 // We would like to use gvn.markInstructionForDeletion here, but we can't
1197 // because the load is already memoized into the leader map table that GVN
1198 // tracks. It is potentially possible to remove the load from the table,
1199 // but then there all of the operations based on it would need to be
1200 // rehashed. Just leave the dead load around.
1201 gvn.getMemDep().removeInstruction(SrcVal);
1205 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1209 /// GetMemInstValueForLoad - This function is called when we have a
1210 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1211 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1212 Type *LoadTy, Instruction *InsertPt,
1213 const DataLayout &DL){
1214 LLVMContext &Ctx = LoadTy->getContext();
1215 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1217 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1219 // We know that this method is only called when the mem transfer fully
1220 // provides the bits for the load.
1221 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1222 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1223 // independently of what the offset is.
1224 Value *Val = MSI->getValue();
1226 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1228 Value *OneElt = Val;
1230 // Splat the value out to the right number of bits.
1231 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1232 // If we can double the number of bytes set, do it.
1233 if (NumBytesSet*2 <= LoadSize) {
1234 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1235 Val = Builder.CreateOr(Val, ShVal);
1240 // Otherwise insert one byte at a time.
1241 Value *ShVal = Builder.CreateShl(Val, 1*8);
1242 Val = Builder.CreateOr(OneElt, ShVal);
1246 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, DL);
1249 // Otherwise, this is a memcpy/memmove from a constant global.
1250 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1251 Constant *Src = cast<Constant>(MTI->getSource());
1252 unsigned AS = Src->getType()->getPointerAddressSpace();
1254 // Otherwise, see if we can constant fold a load from the constant with the
1255 // offset applied as appropriate.
1256 Src = ConstantExpr::getBitCast(Src,
1257 Type::getInt8PtrTy(Src->getContext(), AS));
1258 Constant *OffsetCst =
1259 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1260 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1261 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1262 return ConstantFoldLoadFromConstPtr(Src, &DL);
1266 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1267 /// construct SSA form, allowing us to eliminate LI. This returns the value
1268 /// that should be used at LI's definition site.
1269 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1270 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1272 // Check for the fully redundant, dominating load case. In this case, we can
1273 // just use the dominating value directly.
1274 if (ValuesPerBlock.size() == 1 &&
1275 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1277 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1278 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1281 // Otherwise, we have to construct SSA form.
1282 SmallVector<PHINode*, 8> NewPHIs;
1283 SSAUpdater SSAUpdate(&NewPHIs);
1284 SSAUpdate.Initialize(LI->getType(), LI->getName());
1286 Type *LoadTy = LI->getType();
1288 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1289 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1290 BasicBlock *BB = AV.BB;
1292 if (SSAUpdate.HasValueForBlock(BB))
1295 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1298 // Perform PHI construction.
1299 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1301 // If new PHI nodes were created, notify alias analysis.
1302 if (V->getType()->getScalarType()->isPointerTy()) {
1303 AliasAnalysis *AA = gvn.getAliasAnalysis();
1305 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1306 AA->copyValue(LI, NewPHIs[i]);
1308 // Now that we've copied information to the new PHIs, scan through
1309 // them again and inform alias analysis that we've added potentially
1310 // escaping uses to any values that are operands to these PHIs.
1311 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1312 PHINode *P = NewPHIs[i];
1313 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1314 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1315 AA->addEscapingUse(P->getOperandUse(jj));
1323 Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1325 if (isSimpleValue()) {
1326 Res = getSimpleValue();
1327 if (Res->getType() != LoadTy) {
1328 const DataLayout *DL = gvn.getDataLayout();
1329 assert(DL && "Need target data to handle type mismatch case");
1330 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1333 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1334 << *getSimpleValue() << '\n'
1335 << *Res << '\n' << "\n\n\n");
1337 } else if (isCoercedLoadValue()) {
1338 LoadInst *Load = getCoercedLoadValue();
1339 if (Load->getType() == LoadTy && Offset == 0) {
1342 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1345 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1346 << *getCoercedLoadValue() << '\n'
1347 << *Res << '\n' << "\n\n\n");
1349 } else if (isMemIntrinValue()) {
1350 const DataLayout *DL = gvn.getDataLayout();
1351 assert(DL && "Need target data to handle type mismatch case");
1352 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1353 LoadTy, BB->getTerminator(), *DL);
1354 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1355 << " " << *getMemIntrinValue() << '\n'
1356 << *Res << '\n' << "\n\n\n");
1358 assert(isUndefValue() && "Should be UndefVal");
1359 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1360 return UndefValue::get(LoadTy);
1365 static bool isLifetimeStart(const Instruction *Inst) {
1366 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1367 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1371 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1372 AvailValInBlkVect &ValuesPerBlock,
1373 UnavailBlkVect &UnavailableBlocks) {
1375 // Filter out useless results (non-locals, etc). Keep track of the blocks
1376 // where we have a value available in repl, also keep track of whether we see
1377 // dependencies that produce an unknown value for the load (such as a call
1378 // that could potentially clobber the load).
1379 unsigned NumDeps = Deps.size();
1380 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1381 BasicBlock *DepBB = Deps[i].getBB();
1382 MemDepResult DepInfo = Deps[i].getResult();
1384 if (DeadBlocks.count(DepBB)) {
1385 // Dead dependent mem-op disguise as a load evaluating the same value
1386 // as the load in question.
1387 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1391 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1392 UnavailableBlocks.push_back(DepBB);
1396 if (DepInfo.isClobber()) {
1397 // The address being loaded in this non-local block may not be the same as
1398 // the pointer operand of the load if PHI translation occurs. Make sure
1399 // to consider the right address.
1400 Value *Address = Deps[i].getAddress();
1402 // If the dependence is to a store that writes to a superset of the bits
1403 // read by the load, we can extract the bits we need for the load from the
1405 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1406 if (DL && Address) {
1407 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1410 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1411 DepSI->getValueOperand(),
1418 // Check to see if we have something like this:
1421 // if we have this, replace the later with an extraction from the former.
1422 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1423 // If this is a clobber and L is the first instruction in its block, then
1424 // we have the first instruction in the entry block.
1425 if (DepLI != LI && Address && DL) {
1426 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), Address,
1430 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1437 // If the clobbering value is a memset/memcpy/memmove, see if we can
1438 // forward a value on from it.
1439 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1440 if (DL && Address) {
1441 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1444 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1451 UnavailableBlocks.push_back(DepBB);
1455 // DepInfo.isDef() here
1457 Instruction *DepInst = DepInfo.getInst();
1459 // Loading the allocation -> undef.
1460 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1461 // Loading immediately after lifetime begin -> undef.
1462 isLifetimeStart(DepInst)) {
1463 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1464 UndefValue::get(LI->getType())));
1468 // Loading from calloc (which zero initializes memory) -> zero
1469 if (isCallocLikeFn(DepInst, TLI)) {
1470 ValuesPerBlock.push_back(AvailableValueInBlock::get(
1471 DepBB, Constant::getNullValue(LI->getType())));
1475 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1476 // Reject loads and stores that are to the same address but are of
1477 // different types if we have to.
1478 if (S->getValueOperand()->getType() != LI->getType()) {
1479 // If the stored value is larger or equal to the loaded value, we can
1481 if (!DL || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1482 LI->getType(), *DL)) {
1483 UnavailableBlocks.push_back(DepBB);
1488 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1489 S->getValueOperand()));
1493 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1494 // If the types mismatch and we can't handle it, reject reuse of the load.
1495 if (LD->getType() != LI->getType()) {
1496 // If the stored value is larger or equal to the loaded value, we can
1498 if (!DL || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*DL)) {
1499 UnavailableBlocks.push_back(DepBB);
1503 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1507 UnavailableBlocks.push_back(DepBB);
1511 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1512 UnavailBlkVect &UnavailableBlocks) {
1513 // Okay, we have *some* definitions of the value. This means that the value
1514 // is available in some of our (transitive) predecessors. Lets think about
1515 // doing PRE of this load. This will involve inserting a new load into the
1516 // predecessor when it's not available. We could do this in general, but
1517 // prefer to not increase code size. As such, we only do this when we know
1518 // that we only have to insert *one* load (which means we're basically moving
1519 // the load, not inserting a new one).
1521 SmallPtrSet<BasicBlock *, 4> Blockers;
1522 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1523 Blockers.insert(UnavailableBlocks[i]);
1525 // Let's find the first basic block with more than one predecessor. Walk
1526 // backwards through predecessors if needed.
1527 BasicBlock *LoadBB = LI->getParent();
1528 BasicBlock *TmpBB = LoadBB;
1530 while (TmpBB->getSinglePredecessor()) {
1531 TmpBB = TmpBB->getSinglePredecessor();
1532 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1534 if (Blockers.count(TmpBB))
1537 // If any of these blocks has more than one successor (i.e. if the edge we
1538 // just traversed was critical), then there are other paths through this
1539 // block along which the load may not be anticipated. Hoisting the load
1540 // above this block would be adding the load to execution paths along
1541 // which it was not previously executed.
1542 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1549 // Check to see how many predecessors have the loaded value fully
1551 MapVector<BasicBlock *, Value *> PredLoads;
1552 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1553 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1554 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1555 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1556 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1558 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1559 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1561 BasicBlock *Pred = *PI;
1562 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1566 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1567 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1568 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1569 << Pred->getName() << "': " << *LI << '\n');
1573 if (LoadBB->isLandingPad()) {
1575 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1576 << Pred->getName() << "': " << *LI << '\n');
1580 CriticalEdgePred.push_back(Pred);
1582 // Only add the predecessors that will not be split for now.
1583 PredLoads[Pred] = nullptr;
1587 // Decide whether PRE is profitable for this load.
1588 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1589 assert(NumUnavailablePreds != 0 &&
1590 "Fully available value should already be eliminated!");
1592 // If this load is unavailable in multiple predecessors, reject it.
1593 // FIXME: If we could restructure the CFG, we could make a common pred with
1594 // all the preds that don't have an available LI and insert a new load into
1596 if (NumUnavailablePreds != 1)
1599 // Split critical edges, and update the unavailable predecessors accordingly.
1600 for (BasicBlock *OrigPred : CriticalEdgePred) {
1601 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1602 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1603 PredLoads[NewPred] = nullptr;
1604 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1605 << LoadBB->getName() << '\n');
1608 // Check if the load can safely be moved to all the unavailable predecessors.
1609 bool CanDoPRE = true;
1610 SmallVector<Instruction*, 8> NewInsts;
1611 for (auto &PredLoad : PredLoads) {
1612 BasicBlock *UnavailablePred = PredLoad.first;
1614 // Do PHI translation to get its value in the predecessor if necessary. The
1615 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1617 // If all preds have a single successor, then we know it is safe to insert
1618 // the load on the pred (?!?), so we can insert code to materialize the
1619 // pointer if it is not available.
1620 PHITransAddr Address(LI->getPointerOperand(), DL);
1621 Value *LoadPtr = nullptr;
1622 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1625 // If we couldn't find or insert a computation of this phi translated value,
1628 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1629 << *LI->getPointerOperand() << "\n");
1634 PredLoad.second = LoadPtr;
1638 while (!NewInsts.empty()) {
1639 Instruction *I = NewInsts.pop_back_val();
1640 if (MD) MD->removeInstruction(I);
1641 I->eraseFromParent();
1643 // HINT: Don't revert the edge-splitting as following transformation may
1644 // also need to split these critical edges.
1645 return !CriticalEdgePred.empty();
1648 // Okay, we can eliminate this load by inserting a reload in the predecessor
1649 // and using PHI construction to get the value in the other predecessors, do
1651 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1652 DEBUG(if (!NewInsts.empty())
1653 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1654 << *NewInsts.back() << '\n');
1656 // Assign value numbers to the new instructions.
1657 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1658 // FIXME: We really _ought_ to insert these value numbers into their
1659 // parent's availability map. However, in doing so, we risk getting into
1660 // ordering issues. If a block hasn't been processed yet, we would be
1661 // marking a value as AVAIL-IN, which isn't what we intend.
1662 VN.lookup_or_add(NewInsts[i]);
1665 for (const auto &PredLoad : PredLoads) {
1666 BasicBlock *UnavailablePred = PredLoad.first;
1667 Value *LoadPtr = PredLoad.second;
1669 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1671 UnavailablePred->getTerminator());
1673 // Transfer the old load's AA tags to the new load.
1675 LI->getAAMetadata(Tags);
1677 NewLoad->setAAMetadata(Tags);
1679 // Transfer DebugLoc.
1680 NewLoad->setDebugLoc(LI->getDebugLoc());
1682 // Add the newly created load.
1683 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1685 MD->invalidateCachedPointerInfo(LoadPtr);
1686 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1689 // Perform PHI construction.
1690 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1691 LI->replaceAllUsesWith(V);
1692 if (isa<PHINode>(V))
1694 if (V->getType()->getScalarType()->isPointerTy())
1695 MD->invalidateCachedPointerInfo(V);
1696 markInstructionForDeletion(LI);
1701 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1702 /// non-local by performing PHI construction.
1703 bool GVN::processNonLocalLoad(LoadInst *LI) {
1704 // Step 1: Find the non-local dependencies of the load.
1706 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1707 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1709 // If we had to process more than one hundred blocks to find the
1710 // dependencies, this load isn't worth worrying about. Optimizing
1711 // it will be too expensive.
1712 unsigned NumDeps = Deps.size();
1716 // If we had a phi translation failure, we'll have a single entry which is a
1717 // clobber in the current block. Reject this early.
1719 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1721 dbgs() << "GVN: non-local load ";
1722 LI->printAsOperand(dbgs());
1723 dbgs() << " has unknown dependencies\n";
1728 // Step 2: Analyze the availability of the load
1729 AvailValInBlkVect ValuesPerBlock;
1730 UnavailBlkVect UnavailableBlocks;
1731 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1733 // If we have no predecessors that produce a known value for this load, exit
1735 if (ValuesPerBlock.empty())
1738 // Step 3: Eliminate fully redundancy.
1740 // If all of the instructions we depend on produce a known value for this
1741 // load, then it is fully redundant and we can use PHI insertion to compute
1742 // its value. Insert PHIs and remove the fully redundant value now.
1743 if (UnavailableBlocks.empty()) {
1744 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1746 // Perform PHI construction.
1747 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1748 LI->replaceAllUsesWith(V);
1750 if (isa<PHINode>(V))
1752 if (V->getType()->getScalarType()->isPointerTy())
1753 MD->invalidateCachedPointerInfo(V);
1754 markInstructionForDeletion(LI);
1759 // Step 4: Eliminate partial redundancy.
1760 if (!EnablePRE || !EnableLoadPRE)
1763 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1767 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1768 // Patch the replacement so that it is not more restrictive than the value
1770 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1771 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1772 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) &&
1773 isa<OverflowingBinaryOperator>(ReplOp)) {
1774 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap())
1775 ReplOp->setHasNoSignedWrap(false);
1776 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap())
1777 ReplOp->setHasNoUnsignedWrap(false);
1779 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1780 // FIXME: If both the original and replacement value are part of the
1781 // same control-flow region (meaning that the execution of one
1782 // guarentees the executation of the other), then we can combine the
1783 // noalias scopes here and do better than the general conservative
1784 // answer used in combineMetadata().
1786 // In general, GVN unifies expressions over different control-flow
1787 // regions, and so we need a conservative combination of the noalias
1789 unsigned KnownIDs[] = {
1790 LLVMContext::MD_tbaa,
1791 LLVMContext::MD_alias_scope,
1792 LLVMContext::MD_noalias,
1793 LLVMContext::MD_range,
1794 LLVMContext::MD_fpmath,
1795 LLVMContext::MD_invariant_load,
1797 combineMetadata(ReplInst, I, KnownIDs);
1801 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1802 patchReplacementInstruction(I, Repl);
1803 I->replaceAllUsesWith(Repl);
1806 /// processLoad - Attempt to eliminate a load, first by eliminating it
1807 /// locally, and then attempting non-local elimination if that fails.
1808 bool GVN::processLoad(LoadInst *L) {
1815 if (L->use_empty()) {
1816 markInstructionForDeletion(L);
1820 // ... to a pointer that has been loaded from before...
1821 MemDepResult Dep = MD->getDependency(L);
1823 // If we have a clobber and target data is around, see if this is a clobber
1824 // that we can fix up through code synthesis.
1825 if (Dep.isClobber() && DL) {
1826 // Check to see if we have something like this:
1827 // store i32 123, i32* %P
1828 // %A = bitcast i32* %P to i8*
1829 // %B = gep i8* %A, i32 1
1832 // We could do that by recognizing if the clobber instructions are obviously
1833 // a common base + constant offset, and if the previous store (or memset)
1834 // completely covers this load. This sort of thing can happen in bitfield
1836 Value *AvailVal = nullptr;
1837 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1838 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1839 L->getPointerOperand(),
1842 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1843 L->getType(), L, *DL);
1846 // Check to see if we have something like this:
1849 // if we have this, replace the later with an extraction from the former.
1850 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1851 // If this is a clobber and L is the first instruction in its block, then
1852 // we have the first instruction in the entry block.
1856 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1857 L->getPointerOperand(),
1860 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1863 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1864 // a value on from it.
1865 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1866 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1867 L->getPointerOperand(),
1870 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *DL);
1874 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1875 << *AvailVal << '\n' << *L << "\n\n\n");
1877 // Replace the load!
1878 L->replaceAllUsesWith(AvailVal);
1879 if (AvailVal->getType()->getScalarType()->isPointerTy())
1880 MD->invalidateCachedPointerInfo(AvailVal);
1881 markInstructionForDeletion(L);
1887 // If the value isn't available, don't do anything!
1888 if (Dep.isClobber()) {
1890 // fast print dep, using operator<< on instruction is too slow.
1891 dbgs() << "GVN: load ";
1892 L->printAsOperand(dbgs());
1893 Instruction *I = Dep.getInst();
1894 dbgs() << " is clobbered by " << *I << '\n';
1899 // If it is defined in another block, try harder.
1900 if (Dep.isNonLocal())
1901 return processNonLocalLoad(L);
1905 // fast print dep, using operator<< on instruction is too slow.
1906 dbgs() << "GVN: load ";
1907 L->printAsOperand(dbgs());
1908 dbgs() << " has unknown dependence\n";
1913 Instruction *DepInst = Dep.getInst();
1914 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1915 Value *StoredVal = DepSI->getValueOperand();
1917 // The store and load are to a must-aliased pointer, but they may not
1918 // actually have the same type. See if we know how to reuse the stored
1919 // value (depending on its type).
1920 if (StoredVal->getType() != L->getType()) {
1922 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1927 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1928 << '\n' << *L << "\n\n\n");
1935 L->replaceAllUsesWith(StoredVal);
1936 if (StoredVal->getType()->getScalarType()->isPointerTy())
1937 MD->invalidateCachedPointerInfo(StoredVal);
1938 markInstructionForDeletion(L);
1943 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1944 Value *AvailableVal = DepLI;
1946 // The loads are of a must-aliased pointer, but they may not actually have
1947 // the same type. See if we know how to reuse the previously loaded value
1948 // (depending on its type).
1949 if (DepLI->getType() != L->getType()) {
1951 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1956 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1957 << "\n" << *L << "\n\n\n");
1964 patchAndReplaceAllUsesWith(L, AvailableVal);
1965 if (DepLI->getType()->getScalarType()->isPointerTy())
1966 MD->invalidateCachedPointerInfo(DepLI);
1967 markInstructionForDeletion(L);
1972 // If this load really doesn't depend on anything, then we must be loading an
1973 // undef value. This can happen when loading for a fresh allocation with no
1974 // intervening stores, for example.
1975 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1976 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1977 markInstructionForDeletion(L);
1982 // If this load occurs either right after a lifetime begin,
1983 // then the loaded value is undefined.
1984 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1985 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1986 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1987 markInstructionForDeletion(L);
1993 // If this load follows a calloc (which zero initializes memory),
1994 // then the loaded value is zero
1995 if (isCallocLikeFn(DepInst, TLI)) {
1996 L->replaceAllUsesWith(Constant::getNullValue(L->getType()));
1997 markInstructionForDeletion(L);
2005 // findLeader - In order to find a leader for a given value number at a
2006 // specific basic block, we first obtain the list of all Values for that number,
2007 // and then scan the list to find one whose block dominates the block in
2008 // question. This is fast because dominator tree queries consist of only
2009 // a few comparisons of DFS numbers.
2010 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
2011 LeaderTableEntry Vals = LeaderTable[num];
2012 if (!Vals.Val) return nullptr;
2014 Value *Val = nullptr;
2015 if (DT->dominates(Vals.BB, BB)) {
2017 if (isa<Constant>(Val)) return Val;
2020 LeaderTableEntry* Next = Vals.Next;
2022 if (DT->dominates(Next->BB, BB)) {
2023 if (isa<Constant>(Next->Val)) return Next->Val;
2024 if (!Val) Val = Next->Val;
2033 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
2034 /// use is dominated by the given basic block. Returns the number of uses that
2036 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
2037 const BasicBlockEdge &Root) {
2039 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2043 if (DT->dominates(Root, U)) {
2051 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
2052 /// true if every path from the entry block to 'Dst' passes via this edge. In
2053 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2054 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2055 DominatorTree *DT) {
2056 // While in theory it is interesting to consider the case in which Dst has
2057 // more than one predecessor, because Dst might be part of a loop which is
2058 // only reachable from Src, in practice it is pointless since at the time
2059 // GVN runs all such loops have preheaders, which means that Dst will have
2060 // been changed to have only one predecessor, namely Src.
2061 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2062 const BasicBlock *Src = E.getStart();
2063 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2065 return Pred != nullptr;
2068 /// propagateEquality - The given values are known to be equal in every block
2069 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2070 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2071 bool GVN::propagateEquality(Value *LHS, Value *RHS,
2072 const BasicBlockEdge &Root) {
2073 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2074 Worklist.push_back(std::make_pair(LHS, RHS));
2075 bool Changed = false;
2076 // For speed, compute a conservative fast approximation to
2077 // DT->dominates(Root, Root.getEnd());
2078 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2080 while (!Worklist.empty()) {
2081 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2082 LHS = Item.first; RHS = Item.second;
2084 if (LHS == RHS) continue;
2085 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2087 // Don't try to propagate equalities between constants.
2088 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2090 // Prefer a constant on the right-hand side, or an Argument if no constants.
2091 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2092 std::swap(LHS, RHS);
2093 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2095 // If there is no obvious reason to prefer the left-hand side over the right-
2096 // hand side, ensure the longest lived term is on the right-hand side, so the
2097 // shortest lived term will be replaced by the longest lived. This tends to
2098 // expose more simplifications.
2099 uint32_t LVN = VN.lookup_or_add(LHS);
2100 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2101 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2102 // Move the 'oldest' value to the right-hand side, using the value number as
2104 uint32_t RVN = VN.lookup_or_add(RHS);
2106 std::swap(LHS, RHS);
2111 // If value numbering later sees that an instruction in the scope is equal
2112 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2113 // the invariant that instructions only occur in the leader table for their
2114 // own value number (this is used by removeFromLeaderTable), do not do this
2115 // if RHS is an instruction (if an instruction in the scope is morphed into
2116 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2117 // using the leader table is about compiling faster, not optimizing better).
2118 // The leader table only tracks basic blocks, not edges. Only add to if we
2119 // have the simple case where the edge dominates the end.
2120 if (RootDominatesEnd && !isa<Instruction>(RHS))
2121 addToLeaderTable(LVN, RHS, Root.getEnd());
2123 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2124 // LHS always has at least one use that is not dominated by Root, this will
2125 // never do anything if LHS has only one use.
2126 if (!LHS->hasOneUse()) {
2127 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2128 Changed |= NumReplacements > 0;
2129 NumGVNEqProp += NumReplacements;
2132 // Now try to deduce additional equalities from this one. For example, if the
2133 // known equality was "(A != B)" == "false" then it follows that A and B are
2134 // equal in the scope. Only boolean equalities with an explicit true or false
2135 // RHS are currently supported.
2136 if (!RHS->getType()->isIntegerTy(1))
2137 // Not a boolean equality - bail out.
2139 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2141 // RHS neither 'true' nor 'false' - bail out.
2143 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2144 bool isKnownTrue = CI->isAllOnesValue();
2145 bool isKnownFalse = !isKnownTrue;
2147 // If "A && B" is known true then both A and B are known true. If "A || B"
2148 // is known false then both A and B are known false.
2150 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2151 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2152 Worklist.push_back(std::make_pair(A, RHS));
2153 Worklist.push_back(std::make_pair(B, RHS));
2157 // If we are propagating an equality like "(A == B)" == "true" then also
2158 // propagate the equality A == B. When propagating a comparison such as
2159 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2160 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
2161 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2163 // If "A == B" is known true, or "A != B" is known false, then replace
2164 // A with B everywhere in the scope.
2165 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2166 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2167 Worklist.push_back(std::make_pair(Op0, Op1));
2169 // If "A >= B" is known true, replace "A < B" with false everywhere.
2170 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2171 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2172 // Since we don't have the instruction "A < B" immediately to hand, work out
2173 // the value number that it would have and use that to find an appropriate
2174 // instruction (if any).
2175 uint32_t NextNum = VN.getNextUnusedValueNumber();
2176 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2177 // If the number we were assigned was brand new then there is no point in
2178 // looking for an instruction realizing it: there cannot be one!
2179 if (Num < NextNum) {
2180 Value *NotCmp = findLeader(Root.getEnd(), Num);
2181 if (NotCmp && isa<Instruction>(NotCmp)) {
2182 unsigned NumReplacements =
2183 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2184 Changed |= NumReplacements > 0;
2185 NumGVNEqProp += NumReplacements;
2188 // Ensure that any instruction in scope that gets the "A < B" value number
2189 // is replaced with false.
2190 // The leader table only tracks basic blocks, not edges. Only add to if we
2191 // have the simple case where the edge dominates the end.
2192 if (RootDominatesEnd)
2193 addToLeaderTable(Num, NotVal, Root.getEnd());
2202 /// processInstruction - When calculating availability, handle an instruction
2203 /// by inserting it into the appropriate sets
2204 bool GVN::processInstruction(Instruction *I) {
2205 // Ignore dbg info intrinsics.
2206 if (isa<DbgInfoIntrinsic>(I))
2209 // If the instruction can be easily simplified then do so now in preference
2210 // to value numbering it. Value numbering often exposes redundancies, for
2211 // example if it determines that %y is equal to %x then the instruction
2212 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2213 if (Value *V = SimplifyInstruction(I, DL, TLI, DT)) {
2214 I->replaceAllUsesWith(V);
2215 if (MD && V->getType()->getScalarType()->isPointerTy())
2216 MD->invalidateCachedPointerInfo(V);
2217 markInstructionForDeletion(I);
2222 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2223 if (processLoad(LI))
2226 unsigned Num = VN.lookup_or_add(LI);
2227 addToLeaderTable(Num, LI, LI->getParent());
2231 // For conditional branches, we can perform simple conditional propagation on
2232 // the condition value itself.
2233 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2234 if (!BI->isConditional())
2237 if (isa<Constant>(BI->getCondition()))
2238 return processFoldableCondBr(BI);
2240 Value *BranchCond = BI->getCondition();
2241 BasicBlock *TrueSucc = BI->getSuccessor(0);
2242 BasicBlock *FalseSucc = BI->getSuccessor(1);
2243 // Avoid multiple edges early.
2244 if (TrueSucc == FalseSucc)
2247 BasicBlock *Parent = BI->getParent();
2248 bool Changed = false;
2250 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2251 BasicBlockEdge TrueE(Parent, TrueSucc);
2252 Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
2254 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2255 BasicBlockEdge FalseE(Parent, FalseSucc);
2256 Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
2261 // For switches, propagate the case values into the case destinations.
2262 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2263 Value *SwitchCond = SI->getCondition();
2264 BasicBlock *Parent = SI->getParent();
2265 bool Changed = false;
2267 // Remember how many outgoing edges there are to every successor.
2268 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2269 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2270 ++SwitchEdges[SI->getSuccessor(i)];
2272 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2274 BasicBlock *Dst = i.getCaseSuccessor();
2275 // If there is only a single edge, propagate the case value into it.
2276 if (SwitchEdges.lookup(Dst) == 1) {
2277 BasicBlockEdge E(Parent, Dst);
2278 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
2284 // Instructions with void type don't return a value, so there's
2285 // no point in trying to find redundancies in them.
2286 if (I->getType()->isVoidTy()) return false;
2288 uint32_t NextNum = VN.getNextUnusedValueNumber();
2289 unsigned Num = VN.lookup_or_add(I);
2291 // Allocations are always uniquely numbered, so we can save time and memory
2292 // by fast failing them.
2293 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2294 addToLeaderTable(Num, I, I->getParent());
2298 // If the number we were assigned was a brand new VN, then we don't
2299 // need to do a lookup to see if the number already exists
2300 // somewhere in the domtree: it can't!
2301 if (Num >= NextNum) {
2302 addToLeaderTable(Num, I, I->getParent());
2306 // Perform fast-path value-number based elimination of values inherited from
2308 Value *repl = findLeader(I->getParent(), Num);
2310 // Failure, just remember this instance for future use.
2311 addToLeaderTable(Num, I, I->getParent());
2316 patchAndReplaceAllUsesWith(I, repl);
2317 if (MD && repl->getType()->getScalarType()->isPointerTy())
2318 MD->invalidateCachedPointerInfo(repl);
2319 markInstructionForDeletion(I);
2323 /// runOnFunction - This is the main transformation entry point for a function.
2324 bool GVN::runOnFunction(Function& F) {
2325 if (skipOptnoneFunction(F))
2329 MD = &getAnalysis<MemoryDependenceAnalysis>();
2330 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2331 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
2332 DL = DLP ? &DLP->getDataLayout() : nullptr;
2333 TLI = &getAnalysis<TargetLibraryInfo>();
2334 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2338 bool Changed = false;
2339 bool ShouldContinue = true;
2341 // Merge unconditional branches, allowing PRE to catch more
2342 // optimization opportunities.
2343 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2344 BasicBlock *BB = FI++;
2346 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2347 if (removedBlock) ++NumGVNBlocks;
2349 Changed |= removedBlock;
2352 unsigned Iteration = 0;
2353 while (ShouldContinue) {
2354 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2355 ShouldContinue = iterateOnFunction(F);
2356 Changed |= ShouldContinue;
2361 // Fabricate val-num for dead-code in order to suppress assertion in
2363 assignValNumForDeadCode();
2364 bool PREChanged = true;
2365 while (PREChanged) {
2366 PREChanged = performPRE(F);
2367 Changed |= PREChanged;
2371 // FIXME: Should perform GVN again after PRE does something. PRE can move
2372 // computations into blocks where they become fully redundant. Note that
2373 // we can't do this until PRE's critical edge splitting updates memdep.
2374 // Actually, when this happens, we should just fully integrate PRE into GVN.
2376 cleanupGlobalSets();
2377 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2385 bool GVN::processBlock(BasicBlock *BB) {
2386 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2387 // (and incrementing BI before processing an instruction).
2388 assert(InstrsToErase.empty() &&
2389 "We expect InstrsToErase to be empty across iterations");
2390 if (DeadBlocks.count(BB))
2393 bool ChangedFunction = false;
2395 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2397 ChangedFunction |= processInstruction(BI);
2398 if (InstrsToErase.empty()) {
2403 // If we need some instructions deleted, do it now.
2404 NumGVNInstr += InstrsToErase.size();
2406 // Avoid iterator invalidation.
2407 bool AtStart = BI == BB->begin();
2411 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2412 E = InstrsToErase.end(); I != E; ++I) {
2413 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2414 if (MD) MD->removeInstruction(*I);
2415 DEBUG(verifyRemoved(*I));
2416 (*I)->eraseFromParent();
2418 InstrsToErase.clear();
2426 return ChangedFunction;
2429 /// performPRE - Perform a purely local form of PRE that looks for diamond
2430 /// control flow patterns and attempts to perform simple PRE at the join point.
2431 bool GVN::performPRE(Function &F) {
2432 bool Changed = false;
2433 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2434 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2435 // Nothing to PRE in the entry block.
2436 if (CurrentBlock == &F.getEntryBlock()) continue;
2438 // Don't perform PRE on a landing pad.
2439 if (CurrentBlock->isLandingPad()) continue;
2441 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2442 BE = CurrentBlock->end(); BI != BE; ) {
2443 Instruction *CurInst = BI++;
2445 if (isa<AllocaInst>(CurInst) ||
2446 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2447 CurInst->getType()->isVoidTy() ||
2448 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2449 isa<DbgInfoIntrinsic>(CurInst))
2452 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2453 // sinking the compare again, and it would force the code generator to
2454 // move the i1 from processor flags or predicate registers into a general
2455 // purpose register.
2456 if (isa<CmpInst>(CurInst))
2459 // We don't currently value number ANY inline asm calls.
2460 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2461 if (CallI->isInlineAsm())
2464 uint32_t ValNo = VN.lookup(CurInst);
2466 // Look for the predecessors for PRE opportunities. We're
2467 // only trying to solve the basic diamond case, where
2468 // a value is computed in the successor and one predecessor,
2469 // but not the other. We also explicitly disallow cases
2470 // where the successor is its own predecessor, because they're
2471 // more complicated to get right.
2472 unsigned NumWith = 0;
2473 unsigned NumWithout = 0;
2474 BasicBlock *PREPred = nullptr;
2477 for (pred_iterator PI = pred_begin(CurrentBlock),
2478 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2479 BasicBlock *P = *PI;
2480 // We're not interested in PRE where the block is its
2481 // own predecessor, or in blocks with predecessors
2482 // that are not reachable.
2483 if (P == CurrentBlock) {
2486 } else if (!DT->isReachableFromEntry(P)) {
2491 Value* predV = findLeader(P, ValNo);
2493 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2496 } else if (predV == CurInst) {
2497 /* CurInst dominates this predecessor. */
2501 predMap.push_back(std::make_pair(predV, P));
2506 // Don't do PRE when it might increase code size, i.e. when
2507 // we would need to insert instructions in more than one pred.
2508 if (NumWithout != 1 || NumWith == 0)
2511 // Don't do PRE across indirect branch.
2512 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2515 // We can't do PRE safely on a critical edge, so instead we schedule
2516 // the edge to be split and perform the PRE the next time we iterate
2518 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2519 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2520 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2524 // Instantiate the expression in the predecessor that lacked it.
2525 // Because we are going top-down through the block, all value numbers
2526 // will be available in the predecessor by the time we need them. Any
2527 // that weren't originally present will have been instantiated earlier
2529 Instruction *PREInstr = CurInst->clone();
2530 bool success = true;
2531 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2532 Value *Op = PREInstr->getOperand(i);
2533 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2536 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2537 PREInstr->setOperand(i, V);
2544 // Fail out if we encounter an operand that is not available in
2545 // the PRE predecessor. This is typically because of loads which
2546 // are not value numbered precisely.
2548 DEBUG(verifyRemoved(PREInstr));
2553 PREInstr->insertBefore(PREPred->getTerminator());
2554 PREInstr->setName(CurInst->getName() + ".pre");
2555 PREInstr->setDebugLoc(CurInst->getDebugLoc());
2556 VN.add(PREInstr, ValNo);
2559 // Update the availability map to include the new instruction.
2560 addToLeaderTable(ValNo, PREInstr, PREPred);
2562 // Create a PHI to make the value available in this block.
2563 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(),
2564 CurInst->getName() + ".pre-phi",
2565 CurrentBlock->begin());
2566 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2567 if (Value *V = predMap[i].first)
2568 Phi->addIncoming(V, predMap[i].second);
2570 Phi->addIncoming(PREInstr, PREPred);
2574 addToLeaderTable(ValNo, Phi, CurrentBlock);
2575 Phi->setDebugLoc(CurInst->getDebugLoc());
2576 CurInst->replaceAllUsesWith(Phi);
2577 if (Phi->getType()->getScalarType()->isPointerTy()) {
2578 // Because we have added a PHI-use of the pointer value, it has now
2579 // "escaped" from alias analysis' perspective. We need to inform
2581 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
2583 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2584 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2588 MD->invalidateCachedPointerInfo(Phi);
2591 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2593 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2594 if (MD) MD->removeInstruction(CurInst);
2595 DEBUG(verifyRemoved(CurInst));
2596 CurInst->eraseFromParent();
2601 if (splitCriticalEdges())
2607 /// Split the critical edge connecting the given two blocks, and return
2608 /// the block inserted to the critical edge.
2609 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2610 BasicBlock *BB = SplitCriticalEdge(Pred, Succ, this);
2612 MD->invalidateCachedPredecessors();
2616 /// splitCriticalEdges - Split critical edges found during the previous
2617 /// iteration that may enable further optimization.
2618 bool GVN::splitCriticalEdges() {
2619 if (toSplit.empty())
2622 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2623 SplitCriticalEdge(Edge.first, Edge.second, this);
2624 } while (!toSplit.empty());
2625 if (MD) MD->invalidateCachedPredecessors();
2629 /// iterateOnFunction - Executes one iteration of GVN
2630 bool GVN::iterateOnFunction(Function &F) {
2631 cleanupGlobalSets();
2633 // Top-down walk of the dominator tree
2634 bool Changed = false;
2636 // Needed for value numbering with phi construction to work.
2637 ReversePostOrderTraversal<Function*> RPOT(&F);
2638 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2639 RE = RPOT.end(); RI != RE; ++RI)
2640 Changed |= processBlock(*RI);
2642 // Save the blocks this function have before transformation begins. GVN may
2643 // split critical edge, and hence may invalidate the RPO/DT iterator.
2645 std::vector<BasicBlock *> BBVect;
2646 BBVect.reserve(256);
2647 for (DomTreeNode *X : depth_first(DT->getRootNode()))
2648 BBVect.push_back(X->getBlock());
2650 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2652 Changed |= processBlock(*I);
2658 void GVN::cleanupGlobalSets() {
2660 LeaderTable.clear();
2661 TableAllocator.Reset();
2664 /// verifyRemoved - Verify that the specified instruction does not occur in our
2665 /// internal data structures.
2666 void GVN::verifyRemoved(const Instruction *Inst) const {
2667 VN.verifyRemoved(Inst);
2669 // Walk through the value number scope to make sure the instruction isn't
2670 // ferreted away in it.
2671 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2672 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2673 const LeaderTableEntry *Node = &I->second;
2674 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2676 while (Node->Next) {
2678 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2683 // BB is declared dead, which implied other blocks become dead as well. This
2684 // function is to add all these blocks to "DeadBlocks". For the dead blocks'
2685 // live successors, update their phi nodes by replacing the operands
2686 // corresponding to dead blocks with UndefVal.
2688 void GVN::addDeadBlock(BasicBlock *BB) {
2689 SmallVector<BasicBlock *, 4> NewDead;
2690 SmallSetVector<BasicBlock *, 4> DF;
2692 NewDead.push_back(BB);
2693 while (!NewDead.empty()) {
2694 BasicBlock *D = NewDead.pop_back_val();
2695 if (DeadBlocks.count(D))
2698 // All blocks dominated by D are dead.
2699 SmallVector<BasicBlock *, 8> Dom;
2700 DT->getDescendants(D, Dom);
2701 DeadBlocks.insert(Dom.begin(), Dom.end());
2703 // Figure out the dominance-frontier(D).
2704 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
2705 E = Dom.end(); I != E; I++) {
2707 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
2708 BasicBlock *S = *SI;
2709 if (DeadBlocks.count(S))
2712 bool AllPredDead = true;
2713 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
2714 if (!DeadBlocks.count(*PI)) {
2715 AllPredDead = false;
2720 // S could be proved dead later on. That is why we don't update phi
2721 // operands at this moment.
2724 // While S is not dominated by D, it is dead by now. This could take
2725 // place if S already have a dead predecessor before D is declared
2727 NewDead.push_back(S);
2733 // For the dead blocks' live successors, update their phi nodes by replacing
2734 // the operands corresponding to dead blocks with UndefVal.
2735 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2738 if (DeadBlocks.count(B))
2741 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2742 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
2743 PE = Preds.end(); PI != PE; PI++) {
2744 BasicBlock *P = *PI;
2746 if (!DeadBlocks.count(P))
2749 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2750 if (BasicBlock *S = splitCriticalEdges(P, B))
2751 DeadBlocks.insert(P = S);
2754 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2755 PHINode &Phi = cast<PHINode>(*II);
2756 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2757 UndefValue::get(Phi.getType()));
2763 // If the given branch is recognized as a foldable branch (i.e. conditional
2764 // branch with constant condition), it will perform following analyses and
2766 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2767 // R be the target of the dead out-coming edge.
2768 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2769 // edge. The result of this step will be {X| X is dominated by R}
2770 // 2) Identify those blocks which haves at least one dead prodecessor. The
2771 // result of this step will be dominance-frontier(R).
2772 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2773 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2775 // Return true iff *NEW* dead code are found.
2776 bool GVN::processFoldableCondBr(BranchInst *BI) {
2777 if (!BI || BI->isUnconditional())
2780 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2784 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2785 BI->getSuccessor(1) : BI->getSuccessor(0);
2786 if (DeadBlocks.count(DeadRoot))
2789 if (!DeadRoot->getSinglePredecessor())
2790 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2792 addDeadBlock(DeadRoot);
2796 // performPRE() will trigger assert if it comes across an instruction without
2797 // associated val-num. As it normally has far more live instructions than dead
2798 // instructions, it makes more sense just to "fabricate" a val-number for the
2799 // dead code than checking if instruction involved is dead or not.
2800 void GVN::assignValNumForDeadCode() {
2801 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
2802 E = DeadBlocks.end(); I != E; I++) {
2803 BasicBlock *BB = *I;
2804 for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
2806 Instruction *Inst = &*II;
2807 unsigned ValNum = VN.lookup_or_add(Inst);
2808 addToLeaderTable(ValNum, Inst, BB);