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/SSAUpdater.h"
51 using namespace PatternMatch;
53 #define DEBUG_TYPE "gvn"
55 STATISTIC(NumGVNInstr, "Number of instructions deleted");
56 STATISTIC(NumGVNLoad, "Number of loads deleted");
57 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
58 STATISTIC(NumGVNBlocks, "Number of blocks merged");
59 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
60 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
61 STATISTIC(NumPRELoad, "Number of loads PRE'd");
63 static cl::opt<bool> EnablePRE("enable-pre",
64 cl::init(true), cl::Hidden);
65 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
67 // Maximum allowed recursion depth.
68 static cl::opt<uint32_t>
69 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
70 cl::desc("Max recurse depth (default = 1000)"));
72 //===----------------------------------------------------------------------===//
74 //===----------------------------------------------------------------------===//
76 /// This class holds the mapping between values and value numbers. It is used
77 /// as an efficient mechanism to determine the expression-wise equivalence of
83 SmallVector<uint32_t, 4> varargs;
85 Expression(uint32_t o = ~2U) : opcode(o) { }
87 bool operator==(const Expression &other) const {
88 if (opcode != other.opcode)
90 if (opcode == ~0U || opcode == ~1U)
92 if (type != other.type)
94 if (varargs != other.varargs)
99 friend hash_code hash_value(const Expression &Value) {
100 return hash_combine(Value.opcode, Value.type,
101 hash_combine_range(Value.varargs.begin(),
102 Value.varargs.end()));
107 DenseMap<Value*, uint32_t> valueNumbering;
108 DenseMap<Expression, uint32_t> expressionNumbering;
110 MemoryDependenceAnalysis *MD;
113 uint32_t nextValueNumber;
115 Expression create_expression(Instruction* I);
116 Expression create_cmp_expression(unsigned Opcode,
117 CmpInst::Predicate Predicate,
118 Value *LHS, Value *RHS);
119 Expression create_extractvalue_expression(ExtractValueInst* EI);
120 uint32_t lookup_or_add_call(CallInst* C);
122 ValueTable() : nextValueNumber(1) { }
123 uint32_t lookup_or_add(Value *V);
124 uint32_t lookup(Value *V) const;
125 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
126 Value *LHS, Value *RHS);
127 void add(Value *V, uint32_t num);
129 void erase(Value *v);
130 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
131 AliasAnalysis *getAliasAnalysis() const { return AA; }
132 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
133 void setDomTree(DominatorTree* D) { DT = D; }
134 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
135 void verifyRemoved(const Value *) const;
140 template <> struct DenseMapInfo<Expression> {
141 static inline Expression getEmptyKey() {
145 static inline Expression getTombstoneKey() {
149 static unsigned getHashValue(const Expression e) {
150 using llvm::hash_value;
151 return static_cast<unsigned>(hash_value(e));
153 static bool isEqual(const Expression &LHS, const Expression &RHS) {
160 //===----------------------------------------------------------------------===//
161 // ValueTable Internal Functions
162 //===----------------------------------------------------------------------===//
164 Expression ValueTable::create_expression(Instruction *I) {
166 e.type = I->getType();
167 e.opcode = I->getOpcode();
168 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
170 e.varargs.push_back(lookup_or_add(*OI));
171 if (I->isCommutative()) {
172 // Ensure that commutative instructions that only differ by a permutation
173 // of their operands get the same value number by sorting the operand value
174 // numbers. Since all commutative instructions have two operands it is more
175 // efficient to sort by hand rather than using, say, std::sort.
176 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
177 if (e.varargs[0] > e.varargs[1])
178 std::swap(e.varargs[0], e.varargs[1]);
181 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
182 // Sort the operand value numbers so x<y and y>x get the same value number.
183 CmpInst::Predicate Predicate = C->getPredicate();
184 if (e.varargs[0] > e.varargs[1]) {
185 std::swap(e.varargs[0], e.varargs[1]);
186 Predicate = CmpInst::getSwappedPredicate(Predicate);
188 e.opcode = (C->getOpcode() << 8) | Predicate;
189 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
190 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
192 e.varargs.push_back(*II);
198 Expression ValueTable::create_cmp_expression(unsigned Opcode,
199 CmpInst::Predicate Predicate,
200 Value *LHS, Value *RHS) {
201 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
202 "Not a comparison!");
204 e.type = CmpInst::makeCmpResultType(LHS->getType());
205 e.varargs.push_back(lookup_or_add(LHS));
206 e.varargs.push_back(lookup_or_add(RHS));
208 // Sort the operand value numbers so x<y and y>x get the same value number.
209 if (e.varargs[0] > e.varargs[1]) {
210 std::swap(e.varargs[0], e.varargs[1]);
211 Predicate = CmpInst::getSwappedPredicate(Predicate);
213 e.opcode = (Opcode << 8) | Predicate;
217 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
218 assert(EI && "Not an ExtractValueInst?");
220 e.type = EI->getType();
223 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
224 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
225 // EI might be an extract from one of our recognised intrinsics. If it
226 // is we'll synthesize a semantically equivalent expression instead on
227 // an extract value expression.
228 switch (I->getIntrinsicID()) {
229 case Intrinsic::sadd_with_overflow:
230 case Intrinsic::uadd_with_overflow:
231 e.opcode = Instruction::Add;
233 case Intrinsic::ssub_with_overflow:
234 case Intrinsic::usub_with_overflow:
235 e.opcode = Instruction::Sub;
237 case Intrinsic::smul_with_overflow:
238 case Intrinsic::umul_with_overflow:
239 e.opcode = Instruction::Mul;
246 // Intrinsic recognized. Grab its args to finish building the expression.
247 assert(I->getNumArgOperands() == 2 &&
248 "Expect two args for recognised intrinsics.");
249 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
250 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
255 // Not a recognised intrinsic. Fall back to producing an extract value
257 e.opcode = EI->getOpcode();
258 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
260 e.varargs.push_back(lookup_or_add(*OI));
262 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
264 e.varargs.push_back(*II);
269 //===----------------------------------------------------------------------===//
270 // ValueTable External Functions
271 //===----------------------------------------------------------------------===//
273 /// add - Insert a value into the table with a specified value number.
274 void ValueTable::add(Value *V, uint32_t num) {
275 valueNumbering.insert(std::make_pair(V, num));
278 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
279 if (AA->doesNotAccessMemory(C)) {
280 Expression exp = create_expression(C);
281 uint32_t &e = expressionNumbering[exp];
282 if (!e) e = nextValueNumber++;
283 valueNumbering[C] = e;
285 } else if (AA->onlyReadsMemory(C)) {
286 Expression exp = create_expression(C);
287 uint32_t &e = expressionNumbering[exp];
289 e = nextValueNumber++;
290 valueNumbering[C] = e;
294 e = nextValueNumber++;
295 valueNumbering[C] = e;
299 MemDepResult local_dep = MD->getDependency(C);
301 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
302 valueNumbering[C] = nextValueNumber;
303 return nextValueNumber++;
306 if (local_dep.isDef()) {
307 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
309 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
310 valueNumbering[C] = nextValueNumber;
311 return nextValueNumber++;
314 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
315 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
316 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
318 valueNumbering[C] = nextValueNumber;
319 return nextValueNumber++;
323 uint32_t v = lookup_or_add(local_cdep);
324 valueNumbering[C] = v;
329 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
330 MD->getNonLocalCallDependency(CallSite(C));
331 // FIXME: Move the checking logic to MemDep!
332 CallInst* cdep = nullptr;
334 // Check to see if we have a single dominating call instruction that is
336 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
337 const NonLocalDepEntry *I = &deps[i];
338 if (I->getResult().isNonLocal())
341 // We don't handle non-definitions. If we already have a call, reject
342 // instruction dependencies.
343 if (!I->getResult().isDef() || cdep != nullptr) {
348 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
349 // FIXME: All duplicated with non-local case.
350 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
351 cdep = NonLocalDepCall;
360 valueNumbering[C] = nextValueNumber;
361 return nextValueNumber++;
364 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
365 valueNumbering[C] = nextValueNumber;
366 return nextValueNumber++;
368 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
369 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
370 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
372 valueNumbering[C] = nextValueNumber;
373 return nextValueNumber++;
377 uint32_t v = lookup_or_add(cdep);
378 valueNumbering[C] = v;
382 valueNumbering[C] = nextValueNumber;
383 return nextValueNumber++;
387 /// lookup_or_add - Returns the value number for the specified value, assigning
388 /// it a new number if it did not have one before.
389 uint32_t ValueTable::lookup_or_add(Value *V) {
390 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
391 if (VI != valueNumbering.end())
394 if (!isa<Instruction>(V)) {
395 valueNumbering[V] = nextValueNumber;
396 return nextValueNumber++;
399 Instruction* I = cast<Instruction>(V);
401 switch (I->getOpcode()) {
402 case Instruction::Call:
403 return lookup_or_add_call(cast<CallInst>(I));
404 case Instruction::Add:
405 case Instruction::FAdd:
406 case Instruction::Sub:
407 case Instruction::FSub:
408 case Instruction::Mul:
409 case Instruction::FMul:
410 case Instruction::UDiv:
411 case Instruction::SDiv:
412 case Instruction::FDiv:
413 case Instruction::URem:
414 case Instruction::SRem:
415 case Instruction::FRem:
416 case Instruction::Shl:
417 case Instruction::LShr:
418 case Instruction::AShr:
419 case Instruction::And:
420 case Instruction::Or:
421 case Instruction::Xor:
422 case Instruction::ICmp:
423 case Instruction::FCmp:
424 case Instruction::Trunc:
425 case Instruction::ZExt:
426 case Instruction::SExt:
427 case Instruction::FPToUI:
428 case Instruction::FPToSI:
429 case Instruction::UIToFP:
430 case Instruction::SIToFP:
431 case Instruction::FPTrunc:
432 case Instruction::FPExt:
433 case Instruction::PtrToInt:
434 case Instruction::IntToPtr:
435 case Instruction::BitCast:
436 case Instruction::Select:
437 case Instruction::ExtractElement:
438 case Instruction::InsertElement:
439 case Instruction::ShuffleVector:
440 case Instruction::InsertValue:
441 case Instruction::GetElementPtr:
442 exp = create_expression(I);
444 case Instruction::ExtractValue:
445 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
448 valueNumbering[V] = nextValueNumber;
449 return nextValueNumber++;
452 uint32_t& e = expressionNumbering[exp];
453 if (!e) e = nextValueNumber++;
454 valueNumbering[V] = e;
458 /// lookup - Returns the value number of the specified value. Fails if
459 /// the value has not yet been numbered.
460 uint32_t ValueTable::lookup(Value *V) const {
461 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
462 assert(VI != valueNumbering.end() && "Value not numbered?");
466 /// lookup_or_add_cmp - Returns the value number of the given comparison,
467 /// assigning it a new number if it did not have one before. Useful when
468 /// we deduced the result of a comparison, but don't immediately have an
469 /// instruction realizing that comparison to hand.
470 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
471 CmpInst::Predicate Predicate,
472 Value *LHS, Value *RHS) {
473 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
474 uint32_t& e = expressionNumbering[exp];
475 if (!e) e = nextValueNumber++;
479 /// clear - Remove all entries from the ValueTable.
480 void ValueTable::clear() {
481 valueNumbering.clear();
482 expressionNumbering.clear();
486 /// erase - Remove a value from the value numbering.
487 void ValueTable::erase(Value *V) {
488 valueNumbering.erase(V);
491 /// verifyRemoved - Verify that the value is removed from all internal data
493 void ValueTable::verifyRemoved(const Value *V) const {
494 for (DenseMap<Value*, uint32_t>::const_iterator
495 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
496 assert(I->first != V && "Inst still occurs in value numbering map!");
500 //===----------------------------------------------------------------------===//
502 //===----------------------------------------------------------------------===//
506 struct AvailableValueInBlock {
507 /// BB - The basic block in question.
510 SimpleVal, // A simple offsetted value that is accessed.
511 LoadVal, // A value produced by a load.
512 MemIntrin, // A memory intrinsic which is loaded from.
513 UndefVal // A UndefValue representing a value from dead block (which
514 // is not yet physically removed from the CFG).
517 /// V - The value that is live out of the block.
518 PointerIntPair<Value *, 2, ValType> Val;
520 /// Offset - The byte offset in Val that is interesting for the load query.
523 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
524 unsigned Offset = 0) {
525 AvailableValueInBlock Res;
527 Res.Val.setPointer(V);
528 Res.Val.setInt(SimpleVal);
533 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
534 unsigned Offset = 0) {
535 AvailableValueInBlock Res;
537 Res.Val.setPointer(MI);
538 Res.Val.setInt(MemIntrin);
543 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
544 unsigned Offset = 0) {
545 AvailableValueInBlock Res;
547 Res.Val.setPointer(LI);
548 Res.Val.setInt(LoadVal);
553 static AvailableValueInBlock getUndef(BasicBlock *BB) {
554 AvailableValueInBlock Res;
556 Res.Val.setPointer(nullptr);
557 Res.Val.setInt(UndefVal);
562 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
563 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
564 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
565 bool isUndefValue() const { return Val.getInt() == UndefVal; }
567 Value *getSimpleValue() const {
568 assert(isSimpleValue() && "Wrong accessor");
569 return Val.getPointer();
572 LoadInst *getCoercedLoadValue() const {
573 assert(isCoercedLoadValue() && "Wrong accessor");
574 return cast<LoadInst>(Val.getPointer());
577 MemIntrinsic *getMemIntrinValue() const {
578 assert(isMemIntrinValue() && "Wrong accessor");
579 return cast<MemIntrinsic>(Val.getPointer());
582 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
583 /// defined here to the specified type. This handles various coercion cases.
584 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const;
587 class GVN : public FunctionPass {
589 MemoryDependenceAnalysis *MD;
591 const DataLayout *DL;
592 const TargetLibraryInfo *TLI;
593 SetVector<BasicBlock *> DeadBlocks;
597 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
598 /// have that value number. Use findLeader to query it.
599 struct LeaderTableEntry {
601 const BasicBlock *BB;
602 LeaderTableEntry *Next;
604 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
605 BumpPtrAllocator TableAllocator;
607 SmallVector<Instruction*, 8> InstrsToErase;
609 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
610 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
611 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
614 static char ID; // Pass identification, replacement for typeid
615 explicit GVN(bool noloads = false)
616 : FunctionPass(ID), NoLoads(noloads), MD(nullptr) {
617 initializeGVNPass(*PassRegistry::getPassRegistry());
620 bool runOnFunction(Function &F) override;
622 /// markInstructionForDeletion - This removes the specified instruction from
623 /// our various maps and marks it for deletion.
624 void markInstructionForDeletion(Instruction *I) {
626 InstrsToErase.push_back(I);
629 const DataLayout *getDataLayout() const { return DL; }
630 DominatorTree &getDominatorTree() const { return *DT; }
631 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
632 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
634 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
635 /// its value number.
636 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
637 LeaderTableEntry &Curr = LeaderTable[N];
644 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
647 Node->Next = Curr.Next;
651 /// removeFromLeaderTable - Scan the list of values corresponding to a given
652 /// value number, and remove the given instruction if encountered.
653 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
654 LeaderTableEntry* Prev = nullptr;
655 LeaderTableEntry* Curr = &LeaderTable[N];
657 while (Curr->Val != I || Curr->BB != BB) {
663 Prev->Next = Curr->Next;
669 LeaderTableEntry* Next = Curr->Next;
670 Curr->Val = Next->Val;
672 Curr->Next = Next->Next;
677 // List of critical edges to be split between iterations.
678 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
680 // This transformation requires dominator postdominator info
681 void getAnalysisUsage(AnalysisUsage &AU) const override {
682 AU.addRequired<DominatorTreeWrapperPass>();
683 AU.addRequired<TargetLibraryInfo>();
685 AU.addRequired<MemoryDependenceAnalysis>();
686 AU.addRequired<AliasAnalysis>();
688 AU.addPreserved<DominatorTreeWrapperPass>();
689 AU.addPreserved<AliasAnalysis>();
693 // Helper fuctions of redundant load elimination
694 bool processLoad(LoadInst *L);
695 bool processNonLocalLoad(LoadInst *L);
696 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
697 AvailValInBlkVect &ValuesPerBlock,
698 UnavailBlkVect &UnavailableBlocks);
699 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
700 UnavailBlkVect &UnavailableBlocks);
702 // Other helper routines
703 bool processInstruction(Instruction *I);
704 bool processBlock(BasicBlock *BB);
705 void dump(DenseMap<uint32_t, Value*> &d);
706 bool iterateOnFunction(Function &F);
707 bool performPRE(Function &F);
708 Value *findLeader(const BasicBlock *BB, uint32_t num);
709 void cleanupGlobalSets();
710 void verifyRemoved(const Instruction *I) const;
711 bool splitCriticalEdges();
712 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
713 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
714 const BasicBlockEdge &Root);
715 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
716 bool processFoldableCondBr(BranchInst *BI);
717 void addDeadBlock(BasicBlock *BB);
718 void assignValNumForDeadCode();
724 // createGVNPass - The public interface to this file...
725 FunctionPass *llvm::createGVNPass(bool NoLoads) {
726 return new GVN(NoLoads);
729 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
730 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
731 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
732 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
733 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
734 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
736 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
737 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
739 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
740 E = d.end(); I != E; ++I) {
741 errs() << I->first << "\n";
748 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
749 /// we're analyzing is fully available in the specified block. As we go, keep
750 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
751 /// map is actually a tri-state map with the following values:
752 /// 0) we know the block *is not* fully available.
753 /// 1) we know the block *is* fully available.
754 /// 2) we do not know whether the block is fully available or not, but we are
755 /// currently speculating that it will be.
756 /// 3) we are speculating for this block and have used that to speculate for
758 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
759 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
760 uint32_t RecurseDepth) {
761 if (RecurseDepth > MaxRecurseDepth)
764 // Optimistically assume that the block is fully available and check to see
765 // if we already know about this block in one lookup.
766 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
767 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
769 // If the entry already existed for this block, return the precomputed value.
771 // If this is a speculative "available" value, mark it as being used for
772 // speculation of other blocks.
773 if (IV.first->second == 2)
774 IV.first->second = 3;
775 return IV.first->second != 0;
778 // Otherwise, see if it is fully available in all predecessors.
779 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
781 // If this block has no predecessors, it isn't live-in here.
783 goto SpeculationFailure;
785 for (; PI != PE; ++PI)
786 // If the value isn't fully available in one of our predecessors, then it
787 // isn't fully available in this block either. Undo our previous
788 // optimistic assumption and bail out.
789 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
790 goto SpeculationFailure;
794 // SpeculationFailure - If we get here, we found out that this is not, after
795 // all, a fully-available block. We have a problem if we speculated on this and
796 // used the speculation to mark other blocks as available.
798 char &BBVal = FullyAvailableBlocks[BB];
800 // If we didn't speculate on this, just return with it set to false.
806 // If we did speculate on this value, we could have blocks set to 1 that are
807 // incorrect. Walk the (transitive) successors of this block and mark them as
809 SmallVector<BasicBlock*, 32> BBWorklist;
810 BBWorklist.push_back(BB);
813 BasicBlock *Entry = BBWorklist.pop_back_val();
814 // Note that this sets blocks to 0 (unavailable) if they happen to not
815 // already be in FullyAvailableBlocks. This is safe.
816 char &EntryVal = FullyAvailableBlocks[Entry];
817 if (EntryVal == 0) continue; // Already unavailable.
819 // Mark as unavailable.
822 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
823 } while (!BBWorklist.empty());
829 /// CanCoerceMustAliasedValueToLoad - Return true if
830 /// CoerceAvailableValueToLoadType will succeed.
831 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
833 const DataLayout &DL) {
834 // If the loaded or stored value is an first class array or struct, don't try
835 // to transform them. We need to be able to bitcast to integer.
836 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
837 StoredVal->getType()->isStructTy() ||
838 StoredVal->getType()->isArrayTy())
841 // The store has to be at least as big as the load.
842 if (DL.getTypeSizeInBits(StoredVal->getType()) <
843 DL.getTypeSizeInBits(LoadTy))
849 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
850 /// then a load from a must-aliased pointer of a different type, try to coerce
851 /// the stored value. LoadedTy is the type of the load we want to replace and
852 /// InsertPt is the place to insert new instructions.
854 /// If we can't do it, return null.
855 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
857 Instruction *InsertPt,
858 const DataLayout &DL) {
859 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
862 // If this is already the right type, just return it.
863 Type *StoredValTy = StoredVal->getType();
865 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
866 uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
868 // If the store and reload are the same size, we can always reuse it.
869 if (StoreSize == LoadSize) {
870 // Pointer to Pointer -> use bitcast.
871 if (StoredValTy->getScalarType()->isPointerTy() &&
872 LoadedTy->getScalarType()->isPointerTy())
873 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
875 // Convert source pointers to integers, which can be bitcast.
876 if (StoredValTy->getScalarType()->isPointerTy()) {
877 StoredValTy = DL.getIntPtrType(StoredValTy);
878 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
881 Type *TypeToCastTo = LoadedTy;
882 if (TypeToCastTo->getScalarType()->isPointerTy())
883 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
885 if (StoredValTy != TypeToCastTo)
886 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
888 // Cast to pointer if the load needs a pointer type.
889 if (LoadedTy->getScalarType()->isPointerTy())
890 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
895 // If the loaded value is smaller than the available value, then we can
896 // extract out a piece from it. If the available value is too small, then we
897 // can't do anything.
898 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
900 // Convert source pointers to integers, which can be manipulated.
901 if (StoredValTy->getScalarType()->isPointerTy()) {
902 StoredValTy = DL.getIntPtrType(StoredValTy);
903 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
906 // Convert vectors and fp to integer, which can be manipulated.
907 if (!StoredValTy->isIntegerTy()) {
908 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
909 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
912 // If this is a big-endian system, we need to shift the value down to the low
913 // bits so that a truncate will work.
914 if (DL.isBigEndian()) {
915 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
916 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
919 // Truncate the integer to the right size now.
920 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
921 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
923 if (LoadedTy == NewIntTy)
926 // If the result is a pointer, inttoptr.
927 if (LoadedTy->getScalarType()->isPointerTy())
928 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
930 // Otherwise, bitcast.
931 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
934 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
935 /// memdep query of a load that ends up being a clobbering memory write (store,
936 /// memset, memcpy, memmove). This means that the write *may* provide bits used
937 /// by the load but we can't be sure because the pointers don't mustalias.
939 /// Check this case to see if there is anything more we can do before we give
940 /// up. This returns -1 if we have to give up, or a byte number in the stored
941 /// value of the piece that feeds the load.
942 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
944 uint64_t WriteSizeInBits,
945 const DataLayout &DL) {
946 // If the loaded or stored value is a first class array or struct, don't try
947 // to transform them. We need to be able to bitcast to integer.
948 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
951 int64_t StoreOffset = 0, LoadOffset = 0;
952 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&DL);
953 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &DL);
954 if (StoreBase != LoadBase)
957 // If the load and store are to the exact same address, they should have been
958 // a must alias. AA must have gotten confused.
959 // FIXME: Study to see if/when this happens. One case is forwarding a memset
960 // to a load from the base of the memset.
962 if (LoadOffset == StoreOffset) {
963 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
964 << "Base = " << *StoreBase << "\n"
965 << "Store Ptr = " << *WritePtr << "\n"
966 << "Store Offs = " << StoreOffset << "\n"
967 << "Load Ptr = " << *LoadPtr << "\n";
972 // If the load and store don't overlap at all, the store doesn't provide
973 // anything to the load. In this case, they really don't alias at all, AA
974 // must have gotten confused.
975 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
977 if ((WriteSizeInBits & 7) | (LoadSize & 7))
979 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
983 bool isAAFailure = false;
984 if (StoreOffset < LoadOffset)
985 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
987 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
991 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
992 << "Base = " << *StoreBase << "\n"
993 << "Store Ptr = " << *WritePtr << "\n"
994 << "Store Offs = " << StoreOffset << "\n"
995 << "Load Ptr = " << *LoadPtr << "\n";
1001 // If the Load isn't completely contained within the stored bits, we don't
1002 // have all the bits to feed it. We could do something crazy in the future
1003 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1005 if (StoreOffset > LoadOffset ||
1006 StoreOffset+StoreSize < LoadOffset+LoadSize)
1009 // Okay, we can do this transformation. Return the number of bytes into the
1010 // store that the load is.
1011 return LoadOffset-StoreOffset;
1014 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1015 /// memdep query of a load that ends up being a clobbering store.
1016 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1018 const DataLayout &DL) {
1019 // Cannot handle reading from store of first-class aggregate yet.
1020 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1021 DepSI->getValueOperand()->getType()->isArrayTy())
1024 Value *StorePtr = DepSI->getPointerOperand();
1025 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1026 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1027 StorePtr, StoreSize, DL);
1030 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
1031 /// memdep query of a load that ends up being clobbered by another load. See if
1032 /// the other load can feed into the second load.
1033 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1034 LoadInst *DepLI, const DataLayout &DL){
1035 // Cannot handle reading from store of first-class aggregate yet.
1036 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1039 Value *DepPtr = DepLI->getPointerOperand();
1040 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
1041 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
1042 if (R != -1) return R;
1044 // If we have a load/load clobber an DepLI can be widened to cover this load,
1045 // then we should widen it!
1046 int64_t LoadOffs = 0;
1047 const Value *LoadBase =
1048 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &DL);
1049 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1051 unsigned Size = MemoryDependenceAnalysis::
1052 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, DL);
1053 if (Size == 0) return -1;
1055 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
1060 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1062 const DataLayout &DL) {
1063 // If the mem operation is a non-constant size, we can't handle it.
1064 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1065 if (!SizeCst) return -1;
1066 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1068 // If this is memset, we just need to see if the offset is valid in the size
1070 if (MI->getIntrinsicID() == Intrinsic::memset)
1071 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1074 // If we have a memcpy/memmove, the only case we can handle is if this is a
1075 // copy from constant memory. In that case, we can read directly from the
1077 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1079 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1080 if (!Src) return -1;
1082 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &DL));
1083 if (!GV || !GV->isConstant()) return -1;
1085 // See if the access is within the bounds of the transfer.
1086 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1087 MI->getDest(), MemSizeInBits, DL);
1091 unsigned AS = Src->getType()->getPointerAddressSpace();
1092 // Otherwise, see if we can constant fold a load from the constant with the
1093 // offset applied as appropriate.
1094 Src = ConstantExpr::getBitCast(Src,
1095 Type::getInt8PtrTy(Src->getContext(), AS));
1096 Constant *OffsetCst =
1097 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1098 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1099 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1100 if (ConstantFoldLoadFromConstPtr(Src, &DL))
1106 /// GetStoreValueForLoad - This function is called when we have a
1107 /// memdep query of a load that ends up being a clobbering store. This means
1108 /// that the store provides bits used by the load but we the pointers don't
1109 /// mustalias. Check this case to see if there is anything more we can do
1110 /// before we give up.
1111 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1113 Instruction *InsertPt, const DataLayout &DL){
1114 LLVMContext &Ctx = SrcVal->getType()->getContext();
1116 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1117 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
1119 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1121 // Compute which bits of the stored value are being used by the load. Convert
1122 // to an integer type to start with.
1123 if (SrcVal->getType()->getScalarType()->isPointerTy())
1124 SrcVal = Builder.CreatePtrToInt(SrcVal,
1125 DL.getIntPtrType(SrcVal->getType()));
1126 if (!SrcVal->getType()->isIntegerTy())
1127 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1129 // Shift the bits to the least significant depending on endianness.
1131 if (DL.isLittleEndian())
1132 ShiftAmt = Offset*8;
1134 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1137 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1139 if (LoadSize != StoreSize)
1140 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1142 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, DL);
1145 /// GetLoadValueForLoad - This function is called when we have a
1146 /// memdep query of a load that ends up being a clobbering load. This means
1147 /// that the load *may* provide bits used by the load but we can't be sure
1148 /// because the pointers don't mustalias. Check this case to see if there is
1149 /// anything more we can do before we give up.
1150 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1151 Type *LoadTy, Instruction *InsertPt,
1153 const DataLayout &DL = *gvn.getDataLayout();
1154 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1155 // widen SrcVal out to a larger load.
1156 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
1157 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1158 if (Offset+LoadSize > SrcValSize) {
1159 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1160 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1161 // If we have a load/load clobber an DepLI can be widened to cover this
1162 // load, then we should widen it to the next power of 2 size big enough!
1163 unsigned NewLoadSize = Offset+LoadSize;
1164 if (!isPowerOf2_32(NewLoadSize))
1165 NewLoadSize = NextPowerOf2(NewLoadSize);
1167 Value *PtrVal = SrcVal->getPointerOperand();
1169 // Insert the new load after the old load. This ensures that subsequent
1170 // memdep queries will find the new load. We can't easily remove the old
1171 // load completely because it is already in the value numbering table.
1172 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1174 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1175 DestPTy = PointerType::get(DestPTy,
1176 PtrVal->getType()->getPointerAddressSpace());
1177 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1178 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1179 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1180 NewLoad->takeName(SrcVal);
1181 NewLoad->setAlignment(SrcVal->getAlignment());
1183 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1184 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1186 // Replace uses of the original load with the wider load. On a big endian
1187 // system, we need to shift down to get the relevant bits.
1188 Value *RV = NewLoad;
1189 if (DL.isBigEndian())
1190 RV = Builder.CreateLShr(RV,
1191 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1192 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1193 SrcVal->replaceAllUsesWith(RV);
1195 // We would like to use gvn.markInstructionForDeletion here, but we can't
1196 // because the load is already memoized into the leader map table that GVN
1197 // tracks. It is potentially possible to remove the load from the table,
1198 // but then there all of the operations based on it would need to be
1199 // rehashed. Just leave the dead load around.
1200 gvn.getMemDep().removeInstruction(SrcVal);
1204 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1208 /// GetMemInstValueForLoad - This function is called when we have a
1209 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1210 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1211 Type *LoadTy, Instruction *InsertPt,
1212 const DataLayout &DL){
1213 LLVMContext &Ctx = LoadTy->getContext();
1214 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1216 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1218 // We know that this method is only called when the mem transfer fully
1219 // provides the bits for the load.
1220 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1221 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1222 // independently of what the offset is.
1223 Value *Val = MSI->getValue();
1225 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1227 Value *OneElt = Val;
1229 // Splat the value out to the right number of bits.
1230 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1231 // If we can double the number of bytes set, do it.
1232 if (NumBytesSet*2 <= LoadSize) {
1233 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1234 Val = Builder.CreateOr(Val, ShVal);
1239 // Otherwise insert one byte at a time.
1240 Value *ShVal = Builder.CreateShl(Val, 1*8);
1241 Val = Builder.CreateOr(OneElt, ShVal);
1245 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, DL);
1248 // Otherwise, this is a memcpy/memmove from a constant global.
1249 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1250 Constant *Src = cast<Constant>(MTI->getSource());
1251 unsigned AS = Src->getType()->getPointerAddressSpace();
1253 // Otherwise, see if we can constant fold a load from the constant with the
1254 // offset applied as appropriate.
1255 Src = ConstantExpr::getBitCast(Src,
1256 Type::getInt8PtrTy(Src->getContext(), AS));
1257 Constant *OffsetCst =
1258 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1259 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1260 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1261 return ConstantFoldLoadFromConstPtr(Src, &DL);
1265 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1266 /// construct SSA form, allowing us to eliminate LI. This returns the value
1267 /// that should be used at LI's definition site.
1268 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1269 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1271 // Check for the fully redundant, dominating load case. In this case, we can
1272 // just use the dominating value directly.
1273 if (ValuesPerBlock.size() == 1 &&
1274 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1276 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1277 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1280 // Otherwise, we have to construct SSA form.
1281 SmallVector<PHINode*, 8> NewPHIs;
1282 SSAUpdater SSAUpdate(&NewPHIs);
1283 SSAUpdate.Initialize(LI->getType(), LI->getName());
1285 Type *LoadTy = LI->getType();
1287 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1288 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1289 BasicBlock *BB = AV.BB;
1291 if (SSAUpdate.HasValueForBlock(BB))
1294 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1297 // Perform PHI construction.
1298 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1300 // If new PHI nodes were created, notify alias analysis.
1301 if (V->getType()->getScalarType()->isPointerTy()) {
1302 AliasAnalysis *AA = gvn.getAliasAnalysis();
1304 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1305 AA->copyValue(LI, NewPHIs[i]);
1307 // Now that we've copied information to the new PHIs, scan through
1308 // them again and inform alias analysis that we've added potentially
1309 // escaping uses to any values that are operands to these PHIs.
1310 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1311 PHINode *P = NewPHIs[i];
1312 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1313 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1314 AA->addEscapingUse(P->getOperandUse(jj));
1322 Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1324 if (isSimpleValue()) {
1325 Res = getSimpleValue();
1326 if (Res->getType() != LoadTy) {
1327 const DataLayout *DL = gvn.getDataLayout();
1328 assert(DL && "Need target data to handle type mismatch case");
1329 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1332 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1333 << *getSimpleValue() << '\n'
1334 << *Res << '\n' << "\n\n\n");
1336 } else if (isCoercedLoadValue()) {
1337 LoadInst *Load = getCoercedLoadValue();
1338 if (Load->getType() == LoadTy && Offset == 0) {
1341 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1344 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1345 << *getCoercedLoadValue() << '\n'
1346 << *Res << '\n' << "\n\n\n");
1348 } else if (isMemIntrinValue()) {
1349 const DataLayout *DL = gvn.getDataLayout();
1350 assert(DL && "Need target data to handle type mismatch case");
1351 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1352 LoadTy, BB->getTerminator(), *DL);
1353 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1354 << " " << *getMemIntrinValue() << '\n'
1355 << *Res << '\n' << "\n\n\n");
1357 assert(isUndefValue() && "Should be UndefVal");
1358 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1359 return UndefValue::get(LoadTy);
1364 static bool isLifetimeStart(const Instruction *Inst) {
1365 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1366 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1370 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1371 AvailValInBlkVect &ValuesPerBlock,
1372 UnavailBlkVect &UnavailableBlocks) {
1374 // Filter out useless results (non-locals, etc). Keep track of the blocks
1375 // where we have a value available in repl, also keep track of whether we see
1376 // dependencies that produce an unknown value for the load (such as a call
1377 // that could potentially clobber the load).
1378 unsigned NumDeps = Deps.size();
1379 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1380 BasicBlock *DepBB = Deps[i].getBB();
1381 MemDepResult DepInfo = Deps[i].getResult();
1383 if (DeadBlocks.count(DepBB)) {
1384 // Dead dependent mem-op disguise as a load evaluating the same value
1385 // as the load in question.
1386 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1390 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1391 UnavailableBlocks.push_back(DepBB);
1395 if (DepInfo.isClobber()) {
1396 // The address being loaded in this non-local block may not be the same as
1397 // the pointer operand of the load if PHI translation occurs. Make sure
1398 // to consider the right address.
1399 Value *Address = Deps[i].getAddress();
1401 // If the dependence is to a store that writes to a superset of the bits
1402 // read by the load, we can extract the bits we need for the load from the
1404 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1405 if (DL && Address) {
1406 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1409 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1410 DepSI->getValueOperand(),
1417 // Check to see if we have something like this:
1420 // if we have this, replace the later with an extraction from the former.
1421 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1422 // If this is a clobber and L is the first instruction in its block, then
1423 // we have the first instruction in the entry block.
1424 if (DepLI != LI && Address && DL) {
1425 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), Address,
1429 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1436 // If the clobbering value is a memset/memcpy/memmove, see if we can
1437 // forward a value on from it.
1438 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1439 if (DL && Address) {
1440 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1443 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1450 UnavailableBlocks.push_back(DepBB);
1454 // DepInfo.isDef() here
1456 Instruction *DepInst = DepInfo.getInst();
1458 // Loading the allocation -> undef.
1459 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1460 // Loading immediately after lifetime begin -> undef.
1461 isLifetimeStart(DepInst)) {
1462 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1463 UndefValue::get(LI->getType())));
1467 // Loading from calloc (which zero initializes memory) -> zero
1468 if (isCallocLikeFn(DepInst, TLI)) {
1469 ValuesPerBlock.push_back(AvailableValueInBlock::get(
1470 DepBB, Constant::getNullValue(LI->getType())));
1474 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1475 // Reject loads and stores that are to the same address but are of
1476 // different types if we have to.
1477 if (S->getValueOperand()->getType() != LI->getType()) {
1478 // If the stored value is larger or equal to the loaded value, we can
1480 if (!DL || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1481 LI->getType(), *DL)) {
1482 UnavailableBlocks.push_back(DepBB);
1487 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1488 S->getValueOperand()));
1492 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1493 // If the types mismatch and we can't handle it, reject reuse of the load.
1494 if (LD->getType() != LI->getType()) {
1495 // If the stored value is larger or equal to the loaded value, we can
1497 if (!DL || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*DL)) {
1498 UnavailableBlocks.push_back(DepBB);
1502 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1506 UnavailableBlocks.push_back(DepBB);
1510 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1511 UnavailBlkVect &UnavailableBlocks) {
1512 // Okay, we have *some* definitions of the value. This means that the value
1513 // is available in some of our (transitive) predecessors. Lets think about
1514 // doing PRE of this load. This will involve inserting a new load into the
1515 // predecessor when it's not available. We could do this in general, but
1516 // prefer to not increase code size. As such, we only do this when we know
1517 // that we only have to insert *one* load (which means we're basically moving
1518 // the load, not inserting a new one).
1520 SmallPtrSet<BasicBlock *, 4> Blockers;
1521 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1522 Blockers.insert(UnavailableBlocks[i]);
1524 // Let's find the first basic block with more than one predecessor. Walk
1525 // backwards through predecessors if needed.
1526 BasicBlock *LoadBB = LI->getParent();
1527 BasicBlock *TmpBB = LoadBB;
1529 while (TmpBB->getSinglePredecessor()) {
1530 TmpBB = TmpBB->getSinglePredecessor();
1531 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1533 if (Blockers.count(TmpBB))
1536 // If any of these blocks has more than one successor (i.e. if the edge we
1537 // just traversed was critical), then there are other paths through this
1538 // block along which the load may not be anticipated. Hoisting the load
1539 // above this block would be adding the load to execution paths along
1540 // which it was not previously executed.
1541 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1548 // Check to see how many predecessors have the loaded value fully
1550 MapVector<BasicBlock *, Value *> PredLoads;
1551 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1552 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1553 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1554 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1555 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1557 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1558 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1560 BasicBlock *Pred = *PI;
1561 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1565 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1566 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1567 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1568 << Pred->getName() << "': " << *LI << '\n');
1572 if (LoadBB->isLandingPad()) {
1574 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1575 << Pred->getName() << "': " << *LI << '\n');
1579 CriticalEdgePred.push_back(Pred);
1581 // Only add the predecessors that will not be split for now.
1582 PredLoads[Pred] = nullptr;
1586 // Decide whether PRE is profitable for this load.
1587 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1588 assert(NumUnavailablePreds != 0 &&
1589 "Fully available value should already be eliminated!");
1591 // If this load is unavailable in multiple predecessors, reject it.
1592 // FIXME: If we could restructure the CFG, we could make a common pred with
1593 // all the preds that don't have an available LI and insert a new load into
1595 if (NumUnavailablePreds != 1)
1598 // Split critical edges, and update the unavailable predecessors accordingly.
1599 for (BasicBlock *OrigPred : CriticalEdgePred) {
1600 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1601 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1602 PredLoads[NewPred] = nullptr;
1603 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1604 << LoadBB->getName() << '\n');
1607 // Check if the load can safely be moved to all the unavailable predecessors.
1608 bool CanDoPRE = true;
1609 SmallVector<Instruction*, 8> NewInsts;
1610 for (auto &PredLoad : PredLoads) {
1611 BasicBlock *UnavailablePred = PredLoad.first;
1613 // Do PHI translation to get its value in the predecessor if necessary. The
1614 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1616 // If all preds have a single successor, then we know it is safe to insert
1617 // the load on the pred (?!?), so we can insert code to materialize the
1618 // pointer if it is not available.
1619 PHITransAddr Address(LI->getPointerOperand(), DL);
1620 Value *LoadPtr = nullptr;
1621 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1624 // If we couldn't find or insert a computation of this phi translated value,
1627 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1628 << *LI->getPointerOperand() << "\n");
1633 PredLoad.second = LoadPtr;
1637 while (!NewInsts.empty()) {
1638 Instruction *I = NewInsts.pop_back_val();
1639 if (MD) MD->removeInstruction(I);
1640 I->eraseFromParent();
1642 // HINT: Don't revert the edge-splitting as following transformation may
1643 // also need to split these critical edges.
1644 return !CriticalEdgePred.empty();
1647 // Okay, we can eliminate this load by inserting a reload in the predecessor
1648 // and using PHI construction to get the value in the other predecessors, do
1650 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1651 DEBUG(if (!NewInsts.empty())
1652 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1653 << *NewInsts.back() << '\n');
1655 // Assign value numbers to the new instructions.
1656 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1657 // FIXME: We really _ought_ to insert these value numbers into their
1658 // parent's availability map. However, in doing so, we risk getting into
1659 // ordering issues. If a block hasn't been processed yet, we would be
1660 // marking a value as AVAIL-IN, which isn't what we intend.
1661 VN.lookup_or_add(NewInsts[i]);
1664 for (const auto &PredLoad : PredLoads) {
1665 BasicBlock *UnavailablePred = PredLoad.first;
1666 Value *LoadPtr = PredLoad.second;
1668 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1670 UnavailablePred->getTerminator());
1672 // Transfer the old load's TBAA tag to the new load.
1673 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1674 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1676 // Transfer DebugLoc.
1677 NewLoad->setDebugLoc(LI->getDebugLoc());
1679 // Add the newly created load.
1680 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1682 MD->invalidateCachedPointerInfo(LoadPtr);
1683 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1686 // Perform PHI construction.
1687 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1688 LI->replaceAllUsesWith(V);
1689 if (isa<PHINode>(V))
1691 if (V->getType()->getScalarType()->isPointerTy())
1692 MD->invalidateCachedPointerInfo(V);
1693 markInstructionForDeletion(LI);
1698 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1699 /// non-local by performing PHI construction.
1700 bool GVN::processNonLocalLoad(LoadInst *LI) {
1701 // Step 1: Find the non-local dependencies of the load.
1703 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1704 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1706 // If we had to process more than one hundred blocks to find the
1707 // dependencies, this load isn't worth worrying about. Optimizing
1708 // it will be too expensive.
1709 unsigned NumDeps = Deps.size();
1713 // If we had a phi translation failure, we'll have a single entry which is a
1714 // clobber in the current block. Reject this early.
1716 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1718 dbgs() << "GVN: non-local load ";
1719 LI->printAsOperand(dbgs());
1720 dbgs() << " has unknown dependencies\n";
1725 // Step 2: Analyze the availability of the load
1726 AvailValInBlkVect ValuesPerBlock;
1727 UnavailBlkVect UnavailableBlocks;
1728 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1730 // If we have no predecessors that produce a known value for this load, exit
1732 if (ValuesPerBlock.empty())
1735 // Step 3: Eliminate fully redundancy.
1737 // If all of the instructions we depend on produce a known value for this
1738 // load, then it is fully redundant and we can use PHI insertion to compute
1739 // its value. Insert PHIs and remove the fully redundant value now.
1740 if (UnavailableBlocks.empty()) {
1741 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1743 // Perform PHI construction.
1744 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1745 LI->replaceAllUsesWith(V);
1747 if (isa<PHINode>(V))
1749 if (V->getType()->getScalarType()->isPointerTy())
1750 MD->invalidateCachedPointerInfo(V);
1751 markInstructionForDeletion(LI);
1756 // Step 4: Eliminate partial redundancy.
1757 if (!EnablePRE || !EnableLoadPRE)
1760 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1764 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1765 // Patch the replacement so that it is not more restrictive than the value
1767 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1768 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1769 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) &&
1770 isa<OverflowingBinaryOperator>(ReplOp)) {
1771 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap())
1772 ReplOp->setHasNoSignedWrap(false);
1773 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap())
1774 ReplOp->setHasNoUnsignedWrap(false);
1776 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1777 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
1778 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata);
1779 for (int i = 0, n = Metadata.size(); i < n; ++i) {
1780 unsigned Kind = Metadata[i].first;
1781 MDNode *IMD = I->getMetadata(Kind);
1782 MDNode *ReplMD = Metadata[i].second;
1785 ReplInst->setMetadata(Kind, nullptr); // Remove unknown metadata
1787 case LLVMContext::MD_dbg:
1788 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1789 case LLVMContext::MD_tbaa:
1790 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD));
1792 case LLVMContext::MD_range:
1793 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD));
1795 case LLVMContext::MD_prof:
1796 llvm_unreachable("MD_prof in a non-terminator instruction");
1798 case LLVMContext::MD_fpmath:
1799 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD));
1806 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1807 patchReplacementInstruction(I, Repl);
1808 I->replaceAllUsesWith(Repl);
1811 /// processLoad - Attempt to eliminate a load, first by eliminating it
1812 /// locally, and then attempting non-local elimination if that fails.
1813 bool GVN::processLoad(LoadInst *L) {
1820 if (L->use_empty()) {
1821 markInstructionForDeletion(L);
1825 // ... to a pointer that has been loaded from before...
1826 MemDepResult Dep = MD->getDependency(L);
1828 // If we have a clobber and target data is around, see if this is a clobber
1829 // that we can fix up through code synthesis.
1830 if (Dep.isClobber() && DL) {
1831 // Check to see if we have something like this:
1832 // store i32 123, i32* %P
1833 // %A = bitcast i32* %P to i8*
1834 // %B = gep i8* %A, i32 1
1837 // We could do that by recognizing if the clobber instructions are obviously
1838 // a common base + constant offset, and if the previous store (or memset)
1839 // completely covers this load. This sort of thing can happen in bitfield
1841 Value *AvailVal = nullptr;
1842 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1843 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1844 L->getPointerOperand(),
1847 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1848 L->getType(), L, *DL);
1851 // Check to see if we have something like this:
1854 // if we have this, replace the later with an extraction from the former.
1855 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1856 // If this is a clobber and L is the first instruction in its block, then
1857 // we have the first instruction in the entry block.
1861 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1862 L->getPointerOperand(),
1865 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1868 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1869 // a value on from it.
1870 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1871 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1872 L->getPointerOperand(),
1875 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *DL);
1879 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1880 << *AvailVal << '\n' << *L << "\n\n\n");
1882 // Replace the load!
1883 L->replaceAllUsesWith(AvailVal);
1884 if (AvailVal->getType()->getScalarType()->isPointerTy())
1885 MD->invalidateCachedPointerInfo(AvailVal);
1886 markInstructionForDeletion(L);
1892 // If the value isn't available, don't do anything!
1893 if (Dep.isClobber()) {
1895 // fast print dep, using operator<< on instruction is too slow.
1896 dbgs() << "GVN: load ";
1897 L->printAsOperand(dbgs());
1898 Instruction *I = Dep.getInst();
1899 dbgs() << " is clobbered by " << *I << '\n';
1904 // If it is defined in another block, try harder.
1905 if (Dep.isNonLocal())
1906 return processNonLocalLoad(L);
1910 // fast print dep, using operator<< on instruction is too slow.
1911 dbgs() << "GVN: load ";
1912 L->printAsOperand(dbgs());
1913 dbgs() << " has unknown dependence\n";
1918 Instruction *DepInst = Dep.getInst();
1919 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1920 Value *StoredVal = DepSI->getValueOperand();
1922 // The store and load are to a must-aliased pointer, but they may not
1923 // actually have the same type. See if we know how to reuse the stored
1924 // value (depending on its type).
1925 if (StoredVal->getType() != L->getType()) {
1927 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1932 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1933 << '\n' << *L << "\n\n\n");
1940 L->replaceAllUsesWith(StoredVal);
1941 if (StoredVal->getType()->getScalarType()->isPointerTy())
1942 MD->invalidateCachedPointerInfo(StoredVal);
1943 markInstructionForDeletion(L);
1948 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1949 Value *AvailableVal = DepLI;
1951 // The loads are of a must-aliased pointer, but they may not actually have
1952 // the same type. See if we know how to reuse the previously loaded value
1953 // (depending on its type).
1954 if (DepLI->getType() != L->getType()) {
1956 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1961 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1962 << "\n" << *L << "\n\n\n");
1969 patchAndReplaceAllUsesWith(L, AvailableVal);
1970 if (DepLI->getType()->getScalarType()->isPointerTy())
1971 MD->invalidateCachedPointerInfo(DepLI);
1972 markInstructionForDeletion(L);
1977 // If this load really doesn't depend on anything, then we must be loading an
1978 // undef value. This can happen when loading for a fresh allocation with no
1979 // intervening stores, for example.
1980 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1981 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1982 markInstructionForDeletion(L);
1987 // If this load occurs either right after a lifetime begin,
1988 // then the loaded value is undefined.
1989 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1990 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1991 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1992 markInstructionForDeletion(L);
1998 // If this load follows a calloc (which zero initializes memory),
1999 // then the loaded value is zero
2000 if (isCallocLikeFn(DepInst, TLI)) {
2001 L->replaceAllUsesWith(Constant::getNullValue(L->getType()));
2002 markInstructionForDeletion(L);
2010 // findLeader - In order to find a leader for a given value number at a
2011 // specific basic block, we first obtain the list of all Values for that number,
2012 // and then scan the list to find one whose block dominates the block in
2013 // question. This is fast because dominator tree queries consist of only
2014 // a few comparisons of DFS numbers.
2015 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
2016 LeaderTableEntry Vals = LeaderTable[num];
2017 if (!Vals.Val) return nullptr;
2019 Value *Val = nullptr;
2020 if (DT->dominates(Vals.BB, BB)) {
2022 if (isa<Constant>(Val)) return Val;
2025 LeaderTableEntry* Next = Vals.Next;
2027 if (DT->dominates(Next->BB, BB)) {
2028 if (isa<Constant>(Next->Val)) return Next->Val;
2029 if (!Val) Val = Next->Val;
2038 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
2039 /// use is dominated by the given basic block. Returns the number of uses that
2041 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
2042 const BasicBlockEdge &Root) {
2044 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2048 if (DT->dominates(Root, U)) {
2056 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
2057 /// true if every path from the entry block to 'Dst' passes via this edge. In
2058 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2059 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2060 DominatorTree *DT) {
2061 // While in theory it is interesting to consider the case in which Dst has
2062 // more than one predecessor, because Dst might be part of a loop which is
2063 // only reachable from Src, in practice it is pointless since at the time
2064 // GVN runs all such loops have preheaders, which means that Dst will have
2065 // been changed to have only one predecessor, namely Src.
2066 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2067 const BasicBlock *Src = E.getStart();
2068 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2070 return Pred != nullptr;
2073 /// propagateEquality - The given values are known to be equal in every block
2074 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2075 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2076 bool GVN::propagateEquality(Value *LHS, Value *RHS,
2077 const BasicBlockEdge &Root) {
2078 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2079 Worklist.push_back(std::make_pair(LHS, RHS));
2080 bool Changed = false;
2081 // For speed, compute a conservative fast approximation to
2082 // DT->dominates(Root, Root.getEnd());
2083 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2085 while (!Worklist.empty()) {
2086 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2087 LHS = Item.first; RHS = Item.second;
2089 if (LHS == RHS) continue;
2090 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2092 // Don't try to propagate equalities between constants.
2093 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2095 // Prefer a constant on the right-hand side, or an Argument if no constants.
2096 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2097 std::swap(LHS, RHS);
2098 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2100 // If there is no obvious reason to prefer the left-hand side over the right-
2101 // hand side, ensure the longest lived term is on the right-hand side, so the
2102 // shortest lived term will be replaced by the longest lived. This tends to
2103 // expose more simplifications.
2104 uint32_t LVN = VN.lookup_or_add(LHS);
2105 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2106 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2107 // Move the 'oldest' value to the right-hand side, using the value number as
2109 uint32_t RVN = VN.lookup_or_add(RHS);
2111 std::swap(LHS, RHS);
2116 // If value numbering later sees that an instruction in the scope is equal
2117 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2118 // the invariant that instructions only occur in the leader table for their
2119 // own value number (this is used by removeFromLeaderTable), do not do this
2120 // if RHS is an instruction (if an instruction in the scope is morphed into
2121 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2122 // using the leader table is about compiling faster, not optimizing better).
2123 // The leader table only tracks basic blocks, not edges. Only add to if we
2124 // have the simple case where the edge dominates the end.
2125 if (RootDominatesEnd && !isa<Instruction>(RHS))
2126 addToLeaderTable(LVN, RHS, Root.getEnd());
2128 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2129 // LHS always has at least one use that is not dominated by Root, this will
2130 // never do anything if LHS has only one use.
2131 if (!LHS->hasOneUse()) {
2132 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2133 Changed |= NumReplacements > 0;
2134 NumGVNEqProp += NumReplacements;
2137 // Now try to deduce additional equalities from this one. For example, if the
2138 // known equality was "(A != B)" == "false" then it follows that A and B are
2139 // equal in the scope. Only boolean equalities with an explicit true or false
2140 // RHS are currently supported.
2141 if (!RHS->getType()->isIntegerTy(1))
2142 // Not a boolean equality - bail out.
2144 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2146 // RHS neither 'true' nor 'false' - bail out.
2148 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2149 bool isKnownTrue = CI->isAllOnesValue();
2150 bool isKnownFalse = !isKnownTrue;
2152 // If "A && B" is known true then both A and B are known true. If "A || B"
2153 // is known false then both A and B are known false.
2155 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2156 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2157 Worklist.push_back(std::make_pair(A, RHS));
2158 Worklist.push_back(std::make_pair(B, RHS));
2162 // If we are propagating an equality like "(A == B)" == "true" then also
2163 // propagate the equality A == B. When propagating a comparison such as
2164 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2165 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
2166 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2168 // If "A == B" is known true, or "A != B" is known false, then replace
2169 // A with B everywhere in the scope.
2170 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2171 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2172 Worklist.push_back(std::make_pair(Op0, Op1));
2174 // If "A >= B" is known true, replace "A < B" with false everywhere.
2175 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2176 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2177 // Since we don't have the instruction "A < B" immediately to hand, work out
2178 // the value number that it would have and use that to find an appropriate
2179 // instruction (if any).
2180 uint32_t NextNum = VN.getNextUnusedValueNumber();
2181 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2182 // If the number we were assigned was brand new then there is no point in
2183 // looking for an instruction realizing it: there cannot be one!
2184 if (Num < NextNum) {
2185 Value *NotCmp = findLeader(Root.getEnd(), Num);
2186 if (NotCmp && isa<Instruction>(NotCmp)) {
2187 unsigned NumReplacements =
2188 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2189 Changed |= NumReplacements > 0;
2190 NumGVNEqProp += NumReplacements;
2193 // Ensure that any instruction in scope that gets the "A < B" value number
2194 // is replaced with false.
2195 // The leader table only tracks basic blocks, not edges. Only add to if we
2196 // have the simple case where the edge dominates the end.
2197 if (RootDominatesEnd)
2198 addToLeaderTable(Num, NotVal, Root.getEnd());
2207 /// processInstruction - When calculating availability, handle an instruction
2208 /// by inserting it into the appropriate sets
2209 bool GVN::processInstruction(Instruction *I) {
2210 // Ignore dbg info intrinsics.
2211 if (isa<DbgInfoIntrinsic>(I))
2214 // If the instruction can be easily simplified then do so now in preference
2215 // to value numbering it. Value numbering often exposes redundancies, for
2216 // example if it determines that %y is equal to %x then the instruction
2217 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2218 if (Value *V = SimplifyInstruction(I, DL, TLI, DT)) {
2219 I->replaceAllUsesWith(V);
2220 if (MD && V->getType()->getScalarType()->isPointerTy())
2221 MD->invalidateCachedPointerInfo(V);
2222 markInstructionForDeletion(I);
2227 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2228 if (processLoad(LI))
2231 unsigned Num = VN.lookup_or_add(LI);
2232 addToLeaderTable(Num, LI, LI->getParent());
2236 // For conditional branches, we can perform simple conditional propagation on
2237 // the condition value itself.
2238 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2239 if (!BI->isConditional())
2242 if (isa<Constant>(BI->getCondition()))
2243 return processFoldableCondBr(BI);
2245 Value *BranchCond = BI->getCondition();
2246 BasicBlock *TrueSucc = BI->getSuccessor(0);
2247 BasicBlock *FalseSucc = BI->getSuccessor(1);
2248 // Avoid multiple edges early.
2249 if (TrueSucc == FalseSucc)
2252 BasicBlock *Parent = BI->getParent();
2253 bool Changed = false;
2255 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2256 BasicBlockEdge TrueE(Parent, TrueSucc);
2257 Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
2259 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2260 BasicBlockEdge FalseE(Parent, FalseSucc);
2261 Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
2266 // For switches, propagate the case values into the case destinations.
2267 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2268 Value *SwitchCond = SI->getCondition();
2269 BasicBlock *Parent = SI->getParent();
2270 bool Changed = false;
2272 // Remember how many outgoing edges there are to every successor.
2273 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2274 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2275 ++SwitchEdges[SI->getSuccessor(i)];
2277 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2279 BasicBlock *Dst = i.getCaseSuccessor();
2280 // If there is only a single edge, propagate the case value into it.
2281 if (SwitchEdges.lookup(Dst) == 1) {
2282 BasicBlockEdge E(Parent, Dst);
2283 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
2289 // Instructions with void type don't return a value, so there's
2290 // no point in trying to find redundancies in them.
2291 if (I->getType()->isVoidTy()) return false;
2293 uint32_t NextNum = VN.getNextUnusedValueNumber();
2294 unsigned Num = VN.lookup_or_add(I);
2296 // Allocations are always uniquely numbered, so we can save time and memory
2297 // by fast failing them.
2298 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2299 addToLeaderTable(Num, I, I->getParent());
2303 // If the number we were assigned was a brand new VN, then we don't
2304 // need to do a lookup to see if the number already exists
2305 // somewhere in the domtree: it can't!
2306 if (Num >= NextNum) {
2307 addToLeaderTable(Num, I, I->getParent());
2311 // Perform fast-path value-number based elimination of values inherited from
2313 Value *repl = findLeader(I->getParent(), Num);
2315 // Failure, just remember this instance for future use.
2316 addToLeaderTable(Num, I, I->getParent());
2321 patchAndReplaceAllUsesWith(I, repl);
2322 if (MD && repl->getType()->getScalarType()->isPointerTy())
2323 MD->invalidateCachedPointerInfo(repl);
2324 markInstructionForDeletion(I);
2328 /// runOnFunction - This is the main transformation entry point for a function.
2329 bool GVN::runOnFunction(Function& F) {
2330 if (skipOptnoneFunction(F))
2334 MD = &getAnalysis<MemoryDependenceAnalysis>();
2335 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2336 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
2337 DL = DLP ? &DLP->getDataLayout() : nullptr;
2338 TLI = &getAnalysis<TargetLibraryInfo>();
2339 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2343 bool Changed = false;
2344 bool ShouldContinue = true;
2346 // Merge unconditional branches, allowing PRE to catch more
2347 // optimization opportunities.
2348 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2349 BasicBlock *BB = FI++;
2351 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2352 if (removedBlock) ++NumGVNBlocks;
2354 Changed |= removedBlock;
2357 unsigned Iteration = 0;
2358 while (ShouldContinue) {
2359 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2360 ShouldContinue = iterateOnFunction(F);
2361 Changed |= ShouldContinue;
2366 // Fabricate val-num for dead-code in order to suppress assertion in
2368 assignValNumForDeadCode();
2369 bool PREChanged = true;
2370 while (PREChanged) {
2371 PREChanged = performPRE(F);
2372 Changed |= PREChanged;
2376 // FIXME: Should perform GVN again after PRE does something. PRE can move
2377 // computations into blocks where they become fully redundant. Note that
2378 // we can't do this until PRE's critical edge splitting updates memdep.
2379 // Actually, when this happens, we should just fully integrate PRE into GVN.
2381 cleanupGlobalSets();
2382 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2390 bool GVN::processBlock(BasicBlock *BB) {
2391 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2392 // (and incrementing BI before processing an instruction).
2393 assert(InstrsToErase.empty() &&
2394 "We expect InstrsToErase to be empty across iterations");
2395 if (DeadBlocks.count(BB))
2398 bool ChangedFunction = false;
2400 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2402 ChangedFunction |= processInstruction(BI);
2403 if (InstrsToErase.empty()) {
2408 // If we need some instructions deleted, do it now.
2409 NumGVNInstr += InstrsToErase.size();
2411 // Avoid iterator invalidation.
2412 bool AtStart = BI == BB->begin();
2416 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2417 E = InstrsToErase.end(); I != E; ++I) {
2418 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2419 if (MD) MD->removeInstruction(*I);
2420 DEBUG(verifyRemoved(*I));
2421 (*I)->eraseFromParent();
2423 InstrsToErase.clear();
2431 return ChangedFunction;
2434 /// performPRE - Perform a purely local form of PRE that looks for diamond
2435 /// control flow patterns and attempts to perform simple PRE at the join point.
2436 bool GVN::performPRE(Function &F) {
2437 bool Changed = false;
2438 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2439 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2440 // Nothing to PRE in the entry block.
2441 if (CurrentBlock == &F.getEntryBlock()) continue;
2443 // Don't perform PRE on a landing pad.
2444 if (CurrentBlock->isLandingPad()) continue;
2446 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2447 BE = CurrentBlock->end(); BI != BE; ) {
2448 Instruction *CurInst = BI++;
2450 if (isa<AllocaInst>(CurInst) ||
2451 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2452 CurInst->getType()->isVoidTy() ||
2453 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2454 isa<DbgInfoIntrinsic>(CurInst))
2457 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2458 // sinking the compare again, and it would force the code generator to
2459 // move the i1 from processor flags or predicate registers into a general
2460 // purpose register.
2461 if (isa<CmpInst>(CurInst))
2464 // We don't currently value number ANY inline asm calls.
2465 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2466 if (CallI->isInlineAsm())
2469 uint32_t ValNo = VN.lookup(CurInst);
2471 // Look for the predecessors for PRE opportunities. We're
2472 // only trying to solve the basic diamond case, where
2473 // a value is computed in the successor and one predecessor,
2474 // but not the other. We also explicitly disallow cases
2475 // where the successor is its own predecessor, because they're
2476 // more complicated to get right.
2477 unsigned NumWith = 0;
2478 unsigned NumWithout = 0;
2479 BasicBlock *PREPred = nullptr;
2482 for (pred_iterator PI = pred_begin(CurrentBlock),
2483 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2484 BasicBlock *P = *PI;
2485 // We're not interested in PRE where the block is its
2486 // own predecessor, or in blocks with predecessors
2487 // that are not reachable.
2488 if (P == CurrentBlock) {
2491 } else if (!DT->isReachableFromEntry(P)) {
2496 Value* predV = findLeader(P, ValNo);
2498 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2501 } else if (predV == CurInst) {
2502 /* CurInst dominates this predecessor. */
2506 predMap.push_back(std::make_pair(predV, P));
2511 // Don't do PRE when it might increase code size, i.e. when
2512 // we would need to insert instructions in more than one pred.
2513 if (NumWithout != 1 || NumWith == 0)
2516 // Don't do PRE across indirect branch.
2517 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2520 // We can't do PRE safely on a critical edge, so instead we schedule
2521 // the edge to be split and perform the PRE the next time we iterate
2523 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2524 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2525 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2529 // Instantiate the expression in the predecessor that lacked it.
2530 // Because we are going top-down through the block, all value numbers
2531 // will be available in the predecessor by the time we need them. Any
2532 // that weren't originally present will have been instantiated earlier
2534 Instruction *PREInstr = CurInst->clone();
2535 bool success = true;
2536 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2537 Value *Op = PREInstr->getOperand(i);
2538 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2541 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2542 PREInstr->setOperand(i, V);
2549 // Fail out if we encounter an operand that is not available in
2550 // the PRE predecessor. This is typically because of loads which
2551 // are not value numbered precisely.
2553 DEBUG(verifyRemoved(PREInstr));
2558 PREInstr->insertBefore(PREPred->getTerminator());
2559 PREInstr->setName(CurInst->getName() + ".pre");
2560 PREInstr->setDebugLoc(CurInst->getDebugLoc());
2561 VN.add(PREInstr, ValNo);
2564 // Update the availability map to include the new instruction.
2565 addToLeaderTable(ValNo, PREInstr, PREPred);
2567 // Create a PHI to make the value available in this block.
2568 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(),
2569 CurInst->getName() + ".pre-phi",
2570 CurrentBlock->begin());
2571 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2572 if (Value *V = predMap[i].first)
2573 Phi->addIncoming(V, predMap[i].second);
2575 Phi->addIncoming(PREInstr, PREPred);
2579 addToLeaderTable(ValNo, Phi, CurrentBlock);
2580 Phi->setDebugLoc(CurInst->getDebugLoc());
2581 CurInst->replaceAllUsesWith(Phi);
2582 if (Phi->getType()->getScalarType()->isPointerTy()) {
2583 // Because we have added a PHI-use of the pointer value, it has now
2584 // "escaped" from alias analysis' perspective. We need to inform
2586 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
2588 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2589 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2593 MD->invalidateCachedPointerInfo(Phi);
2596 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2598 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2599 if (MD) MD->removeInstruction(CurInst);
2600 DEBUG(verifyRemoved(CurInst));
2601 CurInst->eraseFromParent();
2606 if (splitCriticalEdges())
2612 /// Split the critical edge connecting the given two blocks, and return
2613 /// the block inserted to the critical edge.
2614 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2615 BasicBlock *BB = SplitCriticalEdge(Pred, Succ, this);
2617 MD->invalidateCachedPredecessors();
2621 /// splitCriticalEdges - Split critical edges found during the previous
2622 /// iteration that may enable further optimization.
2623 bool GVN::splitCriticalEdges() {
2624 if (toSplit.empty())
2627 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2628 SplitCriticalEdge(Edge.first, Edge.second, this);
2629 } while (!toSplit.empty());
2630 if (MD) MD->invalidateCachedPredecessors();
2634 /// iterateOnFunction - Executes one iteration of GVN
2635 bool GVN::iterateOnFunction(Function &F) {
2636 cleanupGlobalSets();
2638 // Top-down walk of the dominator tree
2639 bool Changed = false;
2641 // Needed for value numbering with phi construction to work.
2642 ReversePostOrderTraversal<Function*> RPOT(&F);
2643 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2644 RE = RPOT.end(); RI != RE; ++RI)
2645 Changed |= processBlock(*RI);
2647 // Save the blocks this function have before transformation begins. GVN may
2648 // split critical edge, and hence may invalidate the RPO/DT iterator.
2650 std::vector<BasicBlock *> BBVect;
2651 BBVect.reserve(256);
2652 for (DomTreeNode *x : depth_first(DT->getRootNode()))
2653 BBVect.push_back(x->getBlock());
2655 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2657 Changed |= processBlock(*I);
2663 void GVN::cleanupGlobalSets() {
2665 LeaderTable.clear();
2666 TableAllocator.Reset();
2669 /// verifyRemoved - Verify that the specified instruction does not occur in our
2670 /// internal data structures.
2671 void GVN::verifyRemoved(const Instruction *Inst) const {
2672 VN.verifyRemoved(Inst);
2674 // Walk through the value number scope to make sure the instruction isn't
2675 // ferreted away in it.
2676 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2677 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2678 const LeaderTableEntry *Node = &I->second;
2679 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2681 while (Node->Next) {
2683 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2688 // BB is declared dead, which implied other blocks become dead as well. This
2689 // function is to add all these blocks to "DeadBlocks". For the dead blocks'
2690 // live successors, update their phi nodes by replacing the operands
2691 // corresponding to dead blocks with UndefVal.
2693 void GVN::addDeadBlock(BasicBlock *BB) {
2694 SmallVector<BasicBlock *, 4> NewDead;
2695 SmallSetVector<BasicBlock *, 4> DF;
2697 NewDead.push_back(BB);
2698 while (!NewDead.empty()) {
2699 BasicBlock *D = NewDead.pop_back_val();
2700 if (DeadBlocks.count(D))
2703 // All blocks dominated by D are dead.
2704 SmallVector<BasicBlock *, 8> Dom;
2705 DT->getDescendants(D, Dom);
2706 DeadBlocks.insert(Dom.begin(), Dom.end());
2708 // Figure out the dominance-frontier(D).
2709 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
2710 E = Dom.end(); I != E; I++) {
2712 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
2713 BasicBlock *S = *SI;
2714 if (DeadBlocks.count(S))
2717 bool AllPredDead = true;
2718 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
2719 if (!DeadBlocks.count(*PI)) {
2720 AllPredDead = false;
2725 // S could be proved dead later on. That is why we don't update phi
2726 // operands at this moment.
2729 // While S is not dominated by D, it is dead by now. This could take
2730 // place if S already have a dead predecessor before D is declared
2732 NewDead.push_back(S);
2738 // For the dead blocks' live successors, update their phi nodes by replacing
2739 // the operands corresponding to dead blocks with UndefVal.
2740 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2743 if (DeadBlocks.count(B))
2746 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2747 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
2748 PE = Preds.end(); PI != PE; PI++) {
2749 BasicBlock *P = *PI;
2751 if (!DeadBlocks.count(P))
2754 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2755 if (BasicBlock *S = splitCriticalEdges(P, B))
2756 DeadBlocks.insert(P = S);
2759 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2760 PHINode &Phi = cast<PHINode>(*II);
2761 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2762 UndefValue::get(Phi.getType()));
2768 // If the given branch is recognized as a foldable branch (i.e. conditional
2769 // branch with constant condition), it will perform following analyses and
2771 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2772 // R be the target of the dead out-coming edge.
2773 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2774 // edge. The result of this step will be {X| X is dominated by R}
2775 // 2) Identify those blocks which haves at least one dead prodecessor. The
2776 // result of this step will be dominance-frontier(R).
2777 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2778 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2780 // Return true iff *NEW* dead code are found.
2781 bool GVN::processFoldableCondBr(BranchInst *BI) {
2782 if (!BI || BI->isUnconditional())
2785 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2789 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2790 BI->getSuccessor(1) : BI->getSuccessor(0);
2791 if (DeadBlocks.count(DeadRoot))
2794 if (!DeadRoot->getSinglePredecessor())
2795 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2797 addDeadBlock(DeadRoot);
2801 // performPRE() will trigger assert if it come across an instruciton without
2802 // associated val-num. As it normally has far more live instructions than dead
2803 // instructions, it makes more sense just to "fabricate" a val-number for the
2804 // dead code than checking if instruction involved is dead or not.
2805 void GVN::assignValNumForDeadCode() {
2806 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
2807 E = DeadBlocks.end(); I != E; I++) {
2808 BasicBlock *BB = *I;
2809 for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
2811 Instruction *Inst = &*II;
2812 unsigned ValNum = VN.lookup_or_add(Inst);
2813 addToLeaderTable(ValNum, Inst, BB);