1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
13 // Note that this pass does the value numbering itself; it does not use the
14 // ValueNumbering analysis passes.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "gvn"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/GlobalVariable.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/ConstantFolding.h"
25 #include "llvm/Analysis/Dominators.h"
26 #include "llvm/Analysis/InstructionSimplify.h"
27 #include "llvm/Analysis/Loads.h"
28 #include "llvm/Analysis/MemoryBuiltins.h"
29 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
30 #include "llvm/Analysis/PHITransAddr.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/Assembly/Writer.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/Target/TargetLibraryInfo.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/SSAUpdater.h"
37 #include "llvm/ADT/DenseMap.h"
38 #include "llvm/ADT/DepthFirstIterator.h"
39 #include "llvm/ADT/Hashing.h"
40 #include "llvm/ADT/SmallPtrSet.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/Support/Allocator.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/IRBuilder.h"
46 #include "llvm/Support/PatternMatch.h"
48 using namespace PatternMatch;
50 STATISTIC(NumGVNInstr, "Number of instructions deleted");
51 STATISTIC(NumGVNLoad, "Number of loads deleted");
52 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
53 STATISTIC(NumGVNBlocks, "Number of blocks merged");
54 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
55 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
56 STATISTIC(NumPRELoad, "Number of loads PRE'd");
58 static cl::opt<bool> EnablePRE("enable-pre",
59 cl::init(true), cl::Hidden);
60 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
62 //===----------------------------------------------------------------------===//
64 //===----------------------------------------------------------------------===//
66 /// This class holds the mapping between values and value numbers. It is used
67 /// as an efficient mechanism to determine the expression-wise equivalence of
73 SmallVector<uint32_t, 4> varargs;
75 Expression(uint32_t o = ~2U) : opcode(o) { }
77 bool operator==(const Expression &other) const {
78 if (opcode != other.opcode)
80 if (opcode == ~0U || opcode == ~1U)
82 if (type != other.type)
84 if (varargs != other.varargs)
89 friend hash_code hash_value(const Expression &Value) {
90 return hash_combine(Value.opcode, Value.type,
91 hash_combine_range(Value.varargs.begin(),
92 Value.varargs.end()));
97 DenseMap<Value*, uint32_t> valueNumbering;
98 DenseMap<Expression, uint32_t> expressionNumbering;
100 MemoryDependenceAnalysis *MD;
103 uint32_t nextValueNumber;
105 Expression create_expression(Instruction* I);
106 Expression create_cmp_expression(unsigned Opcode,
107 CmpInst::Predicate Predicate,
108 Value *LHS, Value *RHS);
109 Expression create_extractvalue_expression(ExtractValueInst* EI);
110 uint32_t lookup_or_add_call(CallInst* C);
112 ValueTable() : nextValueNumber(1) { }
113 uint32_t lookup_or_add(Value *V);
114 uint32_t lookup(Value *V) const;
115 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
116 Value *LHS, Value *RHS);
117 void add(Value *V, uint32_t num);
119 void erase(Value *v);
120 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
121 AliasAnalysis *getAliasAnalysis() const { return AA; }
122 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
123 void setDomTree(DominatorTree* D) { DT = D; }
124 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
125 void verifyRemoved(const Value *) const;
130 template <> struct DenseMapInfo<Expression> {
131 static inline Expression getEmptyKey() {
135 static inline Expression getTombstoneKey() {
139 static unsigned getHashValue(const Expression e) {
140 using llvm::hash_value;
141 return static_cast<unsigned>(hash_value(e));
143 static bool isEqual(const Expression &LHS, const Expression &RHS) {
150 //===----------------------------------------------------------------------===//
151 // ValueTable Internal Functions
152 //===----------------------------------------------------------------------===//
154 Expression ValueTable::create_expression(Instruction *I) {
156 e.type = I->getType();
157 e.opcode = I->getOpcode();
158 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
160 e.varargs.push_back(lookup_or_add(*OI));
161 if (I->isCommutative()) {
162 // Ensure that commutative instructions that only differ by a permutation
163 // of their operands get the same value number by sorting the operand value
164 // numbers. Since all commutative instructions have two operands it is more
165 // efficient to sort by hand rather than using, say, std::sort.
166 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
167 if (e.varargs[0] > e.varargs[1])
168 std::swap(e.varargs[0], e.varargs[1]);
171 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
172 // Sort the operand value numbers so x<y and y>x get the same value number.
173 CmpInst::Predicate Predicate = C->getPredicate();
174 if (e.varargs[0] > e.varargs[1]) {
175 std::swap(e.varargs[0], e.varargs[1]);
176 Predicate = CmpInst::getSwappedPredicate(Predicate);
178 e.opcode = (C->getOpcode() << 8) | Predicate;
179 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
180 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
182 e.varargs.push_back(*II);
188 Expression ValueTable::create_cmp_expression(unsigned Opcode,
189 CmpInst::Predicate Predicate,
190 Value *LHS, Value *RHS) {
191 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
192 "Not a comparison!");
194 e.type = CmpInst::makeCmpResultType(LHS->getType());
195 e.varargs.push_back(lookup_or_add(LHS));
196 e.varargs.push_back(lookup_or_add(RHS));
198 // Sort the operand value numbers so x<y and y>x get the same value number.
199 if (e.varargs[0] > e.varargs[1]) {
200 std::swap(e.varargs[0], e.varargs[1]);
201 Predicate = CmpInst::getSwappedPredicate(Predicate);
203 e.opcode = (Opcode << 8) | Predicate;
207 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
208 assert(EI != 0 && "Not an ExtractValueInst?");
210 e.type = EI->getType();
213 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
214 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
215 // EI might be an extract from one of our recognised intrinsics. If it
216 // is we'll synthesize a semantically equivalent expression instead on
217 // an extract value expression.
218 switch (I->getIntrinsicID()) {
219 case Intrinsic::sadd_with_overflow:
220 case Intrinsic::uadd_with_overflow:
221 e.opcode = Instruction::Add;
223 case Intrinsic::ssub_with_overflow:
224 case Intrinsic::usub_with_overflow:
225 e.opcode = Instruction::Sub;
227 case Intrinsic::smul_with_overflow:
228 case Intrinsic::umul_with_overflow:
229 e.opcode = Instruction::Mul;
236 // Intrinsic recognized. Grab its args to finish building the expression.
237 assert(I->getNumArgOperands() == 2 &&
238 "Expect two args for recognised intrinsics.");
239 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
240 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
245 // Not a recognised intrinsic. Fall back to producing an extract value
247 e.opcode = EI->getOpcode();
248 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
250 e.varargs.push_back(lookup_or_add(*OI));
252 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
254 e.varargs.push_back(*II);
259 //===----------------------------------------------------------------------===//
260 // ValueTable External Functions
261 //===----------------------------------------------------------------------===//
263 /// add - Insert a value into the table with a specified value number.
264 void ValueTable::add(Value *V, uint32_t num) {
265 valueNumbering.insert(std::make_pair(V, num));
268 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
269 if (AA->doesNotAccessMemory(C)) {
270 Expression exp = create_expression(C);
271 uint32_t& e = expressionNumbering[exp];
272 if (!e) e = nextValueNumber++;
273 valueNumbering[C] = e;
275 } else if (AA->onlyReadsMemory(C)) {
276 Expression exp = create_expression(C);
277 uint32_t& e = expressionNumbering[exp];
279 e = nextValueNumber++;
280 valueNumbering[C] = e;
284 e = nextValueNumber++;
285 valueNumbering[C] = e;
289 MemDepResult local_dep = MD->getDependency(C);
291 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
292 valueNumbering[C] = nextValueNumber;
293 return nextValueNumber++;
296 if (local_dep.isDef()) {
297 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
299 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
300 valueNumbering[C] = nextValueNumber;
301 return nextValueNumber++;
304 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
305 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
306 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
308 valueNumbering[C] = nextValueNumber;
309 return nextValueNumber++;
313 uint32_t v = lookup_or_add(local_cdep);
314 valueNumbering[C] = v;
319 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
320 MD->getNonLocalCallDependency(CallSite(C));
321 // FIXME: Move the checking logic to MemDep!
324 // Check to see if we have a single dominating call instruction that is
326 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
327 const NonLocalDepEntry *I = &deps[i];
328 if (I->getResult().isNonLocal())
331 // We don't handle non-definitions. If we already have a call, reject
332 // instruction dependencies.
333 if (!I->getResult().isDef() || cdep != 0) {
338 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
339 // FIXME: All duplicated with non-local case.
340 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
341 cdep = NonLocalDepCall;
350 valueNumbering[C] = nextValueNumber;
351 return nextValueNumber++;
354 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
355 valueNumbering[C] = nextValueNumber;
356 return nextValueNumber++;
358 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
359 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
360 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
362 valueNumbering[C] = nextValueNumber;
363 return nextValueNumber++;
367 uint32_t v = lookup_or_add(cdep);
368 valueNumbering[C] = v;
372 valueNumbering[C] = nextValueNumber;
373 return nextValueNumber++;
377 /// lookup_or_add - Returns the value number for the specified value, assigning
378 /// it a new number if it did not have one before.
379 uint32_t ValueTable::lookup_or_add(Value *V) {
380 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
381 if (VI != valueNumbering.end())
384 if (!isa<Instruction>(V)) {
385 valueNumbering[V] = nextValueNumber;
386 return nextValueNumber++;
389 Instruction* I = cast<Instruction>(V);
391 switch (I->getOpcode()) {
392 case Instruction::Call:
393 return lookup_or_add_call(cast<CallInst>(I));
394 case Instruction::Add:
395 case Instruction::FAdd:
396 case Instruction::Sub:
397 case Instruction::FSub:
398 case Instruction::Mul:
399 case Instruction::FMul:
400 case Instruction::UDiv:
401 case Instruction::SDiv:
402 case Instruction::FDiv:
403 case Instruction::URem:
404 case Instruction::SRem:
405 case Instruction::FRem:
406 case Instruction::Shl:
407 case Instruction::LShr:
408 case Instruction::AShr:
409 case Instruction::And:
410 case Instruction::Or :
411 case Instruction::Xor:
412 case Instruction::ICmp:
413 case Instruction::FCmp:
414 case Instruction::Trunc:
415 case Instruction::ZExt:
416 case Instruction::SExt:
417 case Instruction::FPToUI:
418 case Instruction::FPToSI:
419 case Instruction::UIToFP:
420 case Instruction::SIToFP:
421 case Instruction::FPTrunc:
422 case Instruction::FPExt:
423 case Instruction::PtrToInt:
424 case Instruction::IntToPtr:
425 case Instruction::BitCast:
426 case Instruction::Select:
427 case Instruction::ExtractElement:
428 case Instruction::InsertElement:
429 case Instruction::ShuffleVector:
430 case Instruction::InsertValue:
431 case Instruction::GetElementPtr:
432 exp = create_expression(I);
434 case Instruction::ExtractValue:
435 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
438 valueNumbering[V] = nextValueNumber;
439 return nextValueNumber++;
442 uint32_t& e = expressionNumbering[exp];
443 if (!e) e = nextValueNumber++;
444 valueNumbering[V] = e;
448 /// lookup - Returns the value number of the specified value. Fails if
449 /// the value has not yet been numbered.
450 uint32_t ValueTable::lookup(Value *V) const {
451 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
452 assert(VI != valueNumbering.end() && "Value not numbered?");
456 /// lookup_or_add_cmp - Returns the value number of the given comparison,
457 /// assigning it a new number if it did not have one before. Useful when
458 /// we deduced the result of a comparison, but don't immediately have an
459 /// instruction realizing that comparison to hand.
460 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
461 CmpInst::Predicate Predicate,
462 Value *LHS, Value *RHS) {
463 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
464 uint32_t& e = expressionNumbering[exp];
465 if (!e) e = nextValueNumber++;
469 /// clear - Remove all entries from the ValueTable.
470 void ValueTable::clear() {
471 valueNumbering.clear();
472 expressionNumbering.clear();
476 /// erase - Remove a value from the value numbering.
477 void ValueTable::erase(Value *V) {
478 valueNumbering.erase(V);
481 /// verifyRemoved - Verify that the value is removed from all internal data
483 void ValueTable::verifyRemoved(const Value *V) const {
484 for (DenseMap<Value*, uint32_t>::const_iterator
485 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
486 assert(I->first != V && "Inst still occurs in value numbering map!");
490 //===----------------------------------------------------------------------===//
492 //===----------------------------------------------------------------------===//
496 class GVN : public FunctionPass {
498 MemoryDependenceAnalysis *MD;
500 const TargetData *TD;
501 const TargetLibraryInfo *TLI;
505 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
506 /// have that value number. Use findLeader to query it.
507 struct LeaderTableEntry {
510 LeaderTableEntry *Next;
512 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
513 BumpPtrAllocator TableAllocator;
515 SmallVector<Instruction*, 8> InstrsToErase;
517 static char ID; // Pass identification, replacement for typeid
518 explicit GVN(bool noloads = false)
519 : FunctionPass(ID), NoLoads(noloads), MD(0) {
520 initializeGVNPass(*PassRegistry::getPassRegistry());
523 bool runOnFunction(Function &F);
525 /// markInstructionForDeletion - This removes the specified instruction from
526 /// our various maps and marks it for deletion.
527 void markInstructionForDeletion(Instruction *I) {
529 InstrsToErase.push_back(I);
532 const TargetData *getTargetData() const { return TD; }
533 DominatorTree &getDominatorTree() const { return *DT; }
534 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
535 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
537 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
538 /// its value number.
539 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
540 LeaderTableEntry &Curr = LeaderTable[N];
547 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
550 Node->Next = Curr.Next;
554 /// removeFromLeaderTable - Scan the list of values corresponding to a given
555 /// value number, and remove the given value if encountered.
556 void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
557 LeaderTableEntry* Prev = 0;
558 LeaderTableEntry* Curr = &LeaderTable[N];
560 while (Curr->Val != V || Curr->BB != BB) {
566 Prev->Next = Curr->Next;
572 LeaderTableEntry* Next = Curr->Next;
573 Curr->Val = Next->Val;
575 Curr->Next = Next->Next;
580 // List of critical edges to be split between iterations.
581 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
583 // This transformation requires dominator postdominator info
584 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
585 AU.addRequired<DominatorTree>();
586 AU.addRequired<TargetLibraryInfo>();
588 AU.addRequired<MemoryDependenceAnalysis>();
589 AU.addRequired<AliasAnalysis>();
591 AU.addPreserved<DominatorTree>();
592 AU.addPreserved<AliasAnalysis>();
597 // FIXME: eliminate or document these better
598 bool processLoad(LoadInst *L);
599 bool processInstruction(Instruction *I);
600 bool processNonLocalLoad(LoadInst *L);
601 bool processBlock(BasicBlock *BB);
602 void dump(DenseMap<uint32_t, Value*> &d);
603 bool iterateOnFunction(Function &F);
604 bool performPRE(Function &F);
605 Value *findLeader(BasicBlock *BB, uint32_t num);
606 void cleanupGlobalSets();
607 void verifyRemoved(const Instruction *I) const;
608 bool splitCriticalEdges();
609 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
611 bool propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root);
617 // createGVNPass - The public interface to this file...
618 FunctionPass *llvm::createGVNPass(bool NoLoads) {
619 return new GVN(NoLoads);
622 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
623 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
624 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
625 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
626 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
627 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
629 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
631 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
632 E = d.end(); I != E; ++I) {
633 errs() << I->first << "\n";
639 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
640 /// we're analyzing is fully available in the specified block. As we go, keep
641 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
642 /// map is actually a tri-state map with the following values:
643 /// 0) we know the block *is not* fully available.
644 /// 1) we know the block *is* fully available.
645 /// 2) we do not know whether the block is fully available or not, but we are
646 /// currently speculating that it will be.
647 /// 3) we are speculating for this block and have used that to speculate for
649 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
650 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
651 // Optimistically assume that the block is fully available and check to see
652 // if we already know about this block in one lookup.
653 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
654 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
656 // If the entry already existed for this block, return the precomputed value.
658 // If this is a speculative "available" value, mark it as being used for
659 // speculation of other blocks.
660 if (IV.first->second == 2)
661 IV.first->second = 3;
662 return IV.first->second != 0;
665 // Otherwise, see if it is fully available in all predecessors.
666 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
668 // If this block has no predecessors, it isn't live-in here.
670 goto SpeculationFailure;
672 for (; PI != PE; ++PI)
673 // If the value isn't fully available in one of our predecessors, then it
674 // isn't fully available in this block either. Undo our previous
675 // optimistic assumption and bail out.
676 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
677 goto SpeculationFailure;
681 // SpeculationFailure - If we get here, we found out that this is not, after
682 // all, a fully-available block. We have a problem if we speculated on this and
683 // used the speculation to mark other blocks as available.
685 char &BBVal = FullyAvailableBlocks[BB];
687 // If we didn't speculate on this, just return with it set to false.
693 // If we did speculate on this value, we could have blocks set to 1 that are
694 // incorrect. Walk the (transitive) successors of this block and mark them as
696 SmallVector<BasicBlock*, 32> BBWorklist;
697 BBWorklist.push_back(BB);
700 BasicBlock *Entry = BBWorklist.pop_back_val();
701 // Note that this sets blocks to 0 (unavailable) if they happen to not
702 // already be in FullyAvailableBlocks. This is safe.
703 char &EntryVal = FullyAvailableBlocks[Entry];
704 if (EntryVal == 0) continue; // Already unavailable.
706 // Mark as unavailable.
709 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
710 BBWorklist.push_back(*I);
711 } while (!BBWorklist.empty());
717 /// CanCoerceMustAliasedValueToLoad - Return true if
718 /// CoerceAvailableValueToLoadType will succeed.
719 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
721 const TargetData &TD) {
722 // If the loaded or stored value is an first class array or struct, don't try
723 // to transform them. We need to be able to bitcast to integer.
724 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
725 StoredVal->getType()->isStructTy() ||
726 StoredVal->getType()->isArrayTy())
729 // The store has to be at least as big as the load.
730 if (TD.getTypeSizeInBits(StoredVal->getType()) <
731 TD.getTypeSizeInBits(LoadTy))
738 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
739 /// then a load from a must-aliased pointer of a different type, try to coerce
740 /// the stored value. LoadedTy is the type of the load we want to replace and
741 /// InsertPt is the place to insert new instructions.
743 /// If we can't do it, return null.
744 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
746 Instruction *InsertPt,
747 const TargetData &TD) {
748 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
751 // If this is already the right type, just return it.
752 Type *StoredValTy = StoredVal->getType();
754 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
755 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
757 // If the store and reload are the same size, we can always reuse it.
758 if (StoreSize == LoadSize) {
759 // Pointer to Pointer -> use bitcast.
760 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy())
761 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
763 // Convert source pointers to integers, which can be bitcast.
764 if (StoredValTy->isPointerTy()) {
765 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
766 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
769 Type *TypeToCastTo = LoadedTy;
770 if (TypeToCastTo->isPointerTy())
771 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
773 if (StoredValTy != TypeToCastTo)
774 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
776 // Cast to pointer if the load needs a pointer type.
777 if (LoadedTy->isPointerTy())
778 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
783 // If the loaded value is smaller than the available value, then we can
784 // extract out a piece from it. If the available value is too small, then we
785 // can't do anything.
786 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
788 // Convert source pointers to integers, which can be manipulated.
789 if (StoredValTy->isPointerTy()) {
790 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
791 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
794 // Convert vectors and fp to integer, which can be manipulated.
795 if (!StoredValTy->isIntegerTy()) {
796 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
797 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
800 // If this is a big-endian system, we need to shift the value down to the low
801 // bits so that a truncate will work.
802 if (TD.isBigEndian()) {
803 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
804 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
807 // Truncate the integer to the right size now.
808 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
809 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
811 if (LoadedTy == NewIntTy)
814 // If the result is a pointer, inttoptr.
815 if (LoadedTy->isPointerTy())
816 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
818 // Otherwise, bitcast.
819 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
822 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
823 /// memdep query of a load that ends up being a clobbering memory write (store,
824 /// memset, memcpy, memmove). This means that the write *may* provide bits used
825 /// by the load but we can't be sure because the pointers don't mustalias.
827 /// Check this case to see if there is anything more we can do before we give
828 /// up. This returns -1 if we have to give up, or a byte number in the stored
829 /// value of the piece that feeds the load.
830 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
832 uint64_t WriteSizeInBits,
833 const TargetData &TD) {
834 // If the loaded or stored value is a 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())
839 int64_t StoreOffset = 0, LoadOffset = 0;
840 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD);
841 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
842 if (StoreBase != LoadBase)
845 // If the load and store are to the exact same address, they should have been
846 // a must alias. AA must have gotten confused.
847 // FIXME: Study to see if/when this happens. One case is forwarding a memset
848 // to a load from the base of the memset.
850 if (LoadOffset == StoreOffset) {
851 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
852 << "Base = " << *StoreBase << "\n"
853 << "Store Ptr = " << *WritePtr << "\n"
854 << "Store Offs = " << StoreOffset << "\n"
855 << "Load Ptr = " << *LoadPtr << "\n";
860 // If the load and store don't overlap at all, the store doesn't provide
861 // anything to the load. In this case, they really don't alias at all, AA
862 // must have gotten confused.
863 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
865 if ((WriteSizeInBits & 7) | (LoadSize & 7))
867 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
871 bool isAAFailure = false;
872 if (StoreOffset < LoadOffset)
873 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
875 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
879 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
880 << "Base = " << *StoreBase << "\n"
881 << "Store Ptr = " << *WritePtr << "\n"
882 << "Store Offs = " << StoreOffset << "\n"
883 << "Load Ptr = " << *LoadPtr << "\n";
889 // If the Load isn't completely contained within the stored bits, we don't
890 // have all the bits to feed it. We could do something crazy in the future
891 // (issue a smaller load then merge the bits in) but this seems unlikely to be
893 if (StoreOffset > LoadOffset ||
894 StoreOffset+StoreSize < LoadOffset+LoadSize)
897 // Okay, we can do this transformation. Return the number of bytes into the
898 // store that the load is.
899 return LoadOffset-StoreOffset;
902 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
903 /// memdep query of a load that ends up being a clobbering store.
904 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
906 const TargetData &TD) {
907 // Cannot handle reading from store of first-class aggregate yet.
908 if (DepSI->getValueOperand()->getType()->isStructTy() ||
909 DepSI->getValueOperand()->getType()->isArrayTy())
912 Value *StorePtr = DepSI->getPointerOperand();
913 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
914 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
915 StorePtr, StoreSize, TD);
918 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
919 /// memdep query of a load that ends up being clobbered by another load. See if
920 /// the other load can feed into the second load.
921 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
922 LoadInst *DepLI, const TargetData &TD){
923 // Cannot handle reading from store of first-class aggregate yet.
924 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
927 Value *DepPtr = DepLI->getPointerOperand();
928 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType());
929 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD);
930 if (R != -1) return R;
932 // If we have a load/load clobber an DepLI can be widened to cover this load,
933 // then we should widen it!
934 int64_t LoadOffs = 0;
935 const Value *LoadBase =
936 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD);
937 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
939 unsigned Size = MemoryDependenceAnalysis::
940 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD);
941 if (Size == 0) return -1;
943 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD);
948 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
950 const TargetData &TD) {
951 // If the mem operation is a non-constant size, we can't handle it.
952 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
953 if (SizeCst == 0) return -1;
954 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
956 // If this is memset, we just need to see if the offset is valid in the size
958 if (MI->getIntrinsicID() == Intrinsic::memset)
959 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
962 // If we have a memcpy/memmove, the only case we can handle is if this is a
963 // copy from constant memory. In that case, we can read directly from the
965 MemTransferInst *MTI = cast<MemTransferInst>(MI);
967 Constant *Src = dyn_cast<Constant>(MTI->getSource());
968 if (Src == 0) return -1;
970 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
971 if (GV == 0 || !GV->isConstant()) return -1;
973 // See if the access is within the bounds of the transfer.
974 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
975 MI->getDest(), MemSizeInBits, TD);
979 // Otherwise, see if we can constant fold a load from the constant with the
980 // offset applied as appropriate.
981 Src = ConstantExpr::getBitCast(Src,
982 llvm::Type::getInt8PtrTy(Src->getContext()));
983 Constant *OffsetCst =
984 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
985 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
986 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
987 if (ConstantFoldLoadFromConstPtr(Src, &TD))
993 /// GetStoreValueForLoad - This function is called when we have a
994 /// memdep query of a load that ends up being a clobbering store. This means
995 /// that the store provides bits used by the load but we the pointers don't
996 /// mustalias. Check this case to see if there is anything more we can do
997 /// before we give up.
998 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1000 Instruction *InsertPt, const TargetData &TD){
1001 LLVMContext &Ctx = SrcVal->getType()->getContext();
1003 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1004 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
1006 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1008 // Compute which bits of the stored value are being used by the load. Convert
1009 // to an integer type to start with.
1010 if (SrcVal->getType()->isPointerTy())
1011 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx));
1012 if (!SrcVal->getType()->isIntegerTy())
1013 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1015 // Shift the bits to the least significant depending on endianness.
1017 if (TD.isLittleEndian())
1018 ShiftAmt = Offset*8;
1020 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1023 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1025 if (LoadSize != StoreSize)
1026 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1028 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1031 /// GetLoadValueForLoad - This function is called when we have a
1032 /// memdep query of a load that ends up being a clobbering load. This means
1033 /// that the load *may* provide bits used by the load but we can't be sure
1034 /// because the pointers don't mustalias. Check this case to see if there is
1035 /// anything more we can do before we give up.
1036 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1037 Type *LoadTy, Instruction *InsertPt,
1039 const TargetData &TD = *gvn.getTargetData();
1040 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1041 // widen SrcVal out to a larger load.
1042 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType());
1043 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
1044 if (Offset+LoadSize > SrcValSize) {
1045 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1046 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1047 // If we have a load/load clobber an DepLI can be widened to cover this
1048 // load, then we should widen it to the next power of 2 size big enough!
1049 unsigned NewLoadSize = Offset+LoadSize;
1050 if (!isPowerOf2_32(NewLoadSize))
1051 NewLoadSize = NextPowerOf2(NewLoadSize);
1053 Value *PtrVal = SrcVal->getPointerOperand();
1055 // Insert the new load after the old load. This ensures that subsequent
1056 // memdep queries will find the new load. We can't easily remove the old
1057 // load completely because it is already in the value numbering table.
1058 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1060 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1061 DestPTy = PointerType::get(DestPTy,
1062 cast<PointerType>(PtrVal->getType())->getAddressSpace());
1063 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1064 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1065 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1066 NewLoad->takeName(SrcVal);
1067 NewLoad->setAlignment(SrcVal->getAlignment());
1069 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1070 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1072 // Replace uses of the original load with the wider load. On a big endian
1073 // system, we need to shift down to get the relevant bits.
1074 Value *RV = NewLoad;
1075 if (TD.isBigEndian())
1076 RV = Builder.CreateLShr(RV,
1077 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1078 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1079 SrcVal->replaceAllUsesWith(RV);
1081 // We would like to use gvn.markInstructionForDeletion here, but we can't
1082 // because the load is already memoized into the leader map table that GVN
1083 // tracks. It is potentially possible to remove the load from the table,
1084 // but then there all of the operations based on it would need to be
1085 // rehashed. Just leave the dead load around.
1086 gvn.getMemDep().removeInstruction(SrcVal);
1090 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD);
1094 /// GetMemInstValueForLoad - This function is called when we have a
1095 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1096 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1097 Type *LoadTy, Instruction *InsertPt,
1098 const TargetData &TD){
1099 LLVMContext &Ctx = LoadTy->getContext();
1100 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1102 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1104 // We know that this method is only called when the mem transfer fully
1105 // provides the bits for the load.
1106 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1107 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1108 // independently of what the offset is.
1109 Value *Val = MSI->getValue();
1111 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1113 Value *OneElt = Val;
1115 // Splat the value out to the right number of bits.
1116 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1117 // If we can double the number of bytes set, do it.
1118 if (NumBytesSet*2 <= LoadSize) {
1119 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1120 Val = Builder.CreateOr(Val, ShVal);
1125 // Otherwise insert one byte at a time.
1126 Value *ShVal = Builder.CreateShl(Val, 1*8);
1127 Val = Builder.CreateOr(OneElt, ShVal);
1131 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1134 // Otherwise, this is a memcpy/memmove from a constant global.
1135 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1136 Constant *Src = cast<Constant>(MTI->getSource());
1138 // Otherwise, see if we can constant fold a load from the constant with the
1139 // offset applied as appropriate.
1140 Src = ConstantExpr::getBitCast(Src,
1141 llvm::Type::getInt8PtrTy(Src->getContext()));
1142 Constant *OffsetCst =
1143 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1144 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1145 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1146 return ConstantFoldLoadFromConstPtr(Src, &TD);
1151 struct AvailableValueInBlock {
1152 /// BB - The basic block in question.
1155 SimpleVal, // A simple offsetted value that is accessed.
1156 LoadVal, // A value produced by a load.
1157 MemIntrin // A memory intrinsic which is loaded from.
1160 /// V - The value that is live out of the block.
1161 PointerIntPair<Value *, 2, ValType> Val;
1163 /// Offset - The byte offset in Val that is interesting for the load query.
1166 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1167 unsigned Offset = 0) {
1168 AvailableValueInBlock Res;
1170 Res.Val.setPointer(V);
1171 Res.Val.setInt(SimpleVal);
1172 Res.Offset = Offset;
1176 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1177 unsigned Offset = 0) {
1178 AvailableValueInBlock Res;
1180 Res.Val.setPointer(MI);
1181 Res.Val.setInt(MemIntrin);
1182 Res.Offset = Offset;
1186 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
1187 unsigned Offset = 0) {
1188 AvailableValueInBlock Res;
1190 Res.Val.setPointer(LI);
1191 Res.Val.setInt(LoadVal);
1192 Res.Offset = Offset;
1196 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
1197 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
1198 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
1200 Value *getSimpleValue() const {
1201 assert(isSimpleValue() && "Wrong accessor");
1202 return Val.getPointer();
1205 LoadInst *getCoercedLoadValue() const {
1206 assert(isCoercedLoadValue() && "Wrong accessor");
1207 return cast<LoadInst>(Val.getPointer());
1210 MemIntrinsic *getMemIntrinValue() const {
1211 assert(isMemIntrinValue() && "Wrong accessor");
1212 return cast<MemIntrinsic>(Val.getPointer());
1215 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
1216 /// defined here to the specified type. This handles various coercion cases.
1217 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1219 if (isSimpleValue()) {
1220 Res = getSimpleValue();
1221 if (Res->getType() != LoadTy) {
1222 const TargetData *TD = gvn.getTargetData();
1223 assert(TD && "Need target data to handle type mismatch case");
1224 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1227 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1228 << *getSimpleValue() << '\n'
1229 << *Res << '\n' << "\n\n\n");
1231 } else if (isCoercedLoadValue()) {
1232 LoadInst *Load = getCoercedLoadValue();
1233 if (Load->getType() == LoadTy && Offset == 0) {
1236 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1239 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1240 << *getCoercedLoadValue() << '\n'
1241 << *Res << '\n' << "\n\n\n");
1244 const TargetData *TD = gvn.getTargetData();
1245 assert(TD && "Need target data to handle type mismatch case");
1246 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1247 LoadTy, BB->getTerminator(), *TD);
1248 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1249 << " " << *getMemIntrinValue() << '\n'
1250 << *Res << '\n' << "\n\n\n");
1256 } // end anonymous namespace
1258 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1259 /// construct SSA form, allowing us to eliminate LI. This returns the value
1260 /// that should be used at LI's definition site.
1261 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1262 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1264 // Check for the fully redundant, dominating load case. In this case, we can
1265 // just use the dominating value directly.
1266 if (ValuesPerBlock.size() == 1 &&
1267 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1269 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1271 // Otherwise, we have to construct SSA form.
1272 SmallVector<PHINode*, 8> NewPHIs;
1273 SSAUpdater SSAUpdate(&NewPHIs);
1274 SSAUpdate.Initialize(LI->getType(), LI->getName());
1276 Type *LoadTy = LI->getType();
1278 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1279 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1280 BasicBlock *BB = AV.BB;
1282 if (SSAUpdate.HasValueForBlock(BB))
1285 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1288 // Perform PHI construction.
1289 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1291 // If new PHI nodes were created, notify alias analysis.
1292 if (V->getType()->isPointerTy()) {
1293 AliasAnalysis *AA = gvn.getAliasAnalysis();
1295 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1296 AA->copyValue(LI, NewPHIs[i]);
1298 // Now that we've copied information to the new PHIs, scan through
1299 // them again and inform alias analysis that we've added potentially
1300 // escaping uses to any values that are operands to these PHIs.
1301 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1302 PHINode *P = NewPHIs[i];
1303 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1304 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1305 AA->addEscapingUse(P->getOperandUse(jj));
1313 static bool isLifetimeStart(const Instruction *Inst) {
1314 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1315 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1319 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1320 /// non-local by performing PHI construction.
1321 bool GVN::processNonLocalLoad(LoadInst *LI) {
1322 // Find the non-local dependencies of the load.
1323 SmallVector<NonLocalDepResult, 64> Deps;
1324 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1325 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1326 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
1327 // << Deps.size() << *LI << '\n');
1329 // If we had to process more than one hundred blocks to find the
1330 // dependencies, this load isn't worth worrying about. Optimizing
1331 // it will be too expensive.
1332 unsigned NumDeps = Deps.size();
1336 // If we had a phi translation failure, we'll have a single entry which is a
1337 // clobber in the current block. Reject this early.
1339 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1341 dbgs() << "GVN: non-local load ";
1342 WriteAsOperand(dbgs(), LI);
1343 dbgs() << " has unknown dependencies\n";
1348 // Filter out useless results (non-locals, etc). Keep track of the blocks
1349 // where we have a value available in repl, also keep track of whether we see
1350 // dependencies that produce an unknown value for the load (such as a call
1351 // that could potentially clobber the load).
1352 SmallVector<AvailableValueInBlock, 64> ValuesPerBlock;
1353 SmallVector<BasicBlock*, 64> UnavailableBlocks;
1355 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1356 BasicBlock *DepBB = Deps[i].getBB();
1357 MemDepResult DepInfo = Deps[i].getResult();
1359 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1360 UnavailableBlocks.push_back(DepBB);
1364 if (DepInfo.isClobber()) {
1365 // The address being loaded in this non-local block may not be the same as
1366 // the pointer operand of the load if PHI translation occurs. Make sure
1367 // to consider the right address.
1368 Value *Address = Deps[i].getAddress();
1370 // If the dependence is to a store that writes to a superset of the bits
1371 // read by the load, we can extract the bits we need for the load from the
1373 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1374 if (TD && Address) {
1375 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1378 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1379 DepSI->getValueOperand(),
1386 // Check to see if we have something like this:
1389 // if we have this, replace the later with an extraction from the former.
1390 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1391 // If this is a clobber and L is the first instruction in its block, then
1392 // we have the first instruction in the entry block.
1393 if (DepLI != LI && Address && TD) {
1394 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
1395 LI->getPointerOperand(),
1399 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1406 // If the clobbering value is a memset/memcpy/memmove, see if we can
1407 // forward a value on from it.
1408 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1409 if (TD && Address) {
1410 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1413 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1420 UnavailableBlocks.push_back(DepBB);
1424 // DepInfo.isDef() here
1426 Instruction *DepInst = DepInfo.getInst();
1428 // Loading the allocation -> undef.
1429 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1430 // Loading immediately after lifetime begin -> undef.
1431 isLifetimeStart(DepInst)) {
1432 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1433 UndefValue::get(LI->getType())));
1437 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1438 // Reject loads and stores that are to the same address but are of
1439 // different types if we have to.
1440 if (S->getValueOperand()->getType() != LI->getType()) {
1441 // If the stored value is larger or equal to the loaded value, we can
1443 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1444 LI->getType(), *TD)) {
1445 UnavailableBlocks.push_back(DepBB);
1450 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1451 S->getValueOperand()));
1455 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1456 // If the types mismatch and we can't handle it, reject reuse of the load.
1457 if (LD->getType() != LI->getType()) {
1458 // If the stored value is larger or equal to the loaded value, we can
1460 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1461 UnavailableBlocks.push_back(DepBB);
1465 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1469 UnavailableBlocks.push_back(DepBB);
1473 // If we have no predecessors that produce a known value for this load, exit
1475 if (ValuesPerBlock.empty()) return false;
1477 // If all of the instructions we depend on produce a known value for this
1478 // load, then it is fully redundant and we can use PHI insertion to compute
1479 // its value. Insert PHIs and remove the fully redundant value now.
1480 if (UnavailableBlocks.empty()) {
1481 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1483 // Perform PHI construction.
1484 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1485 LI->replaceAllUsesWith(V);
1487 if (isa<PHINode>(V))
1489 if (V->getType()->isPointerTy())
1490 MD->invalidateCachedPointerInfo(V);
1491 markInstructionForDeletion(LI);
1496 if (!EnablePRE || !EnableLoadPRE)
1499 // Okay, we have *some* definitions of the value. This means that the value
1500 // is available in some of our (transitive) predecessors. Lets think about
1501 // doing PRE of this load. This will involve inserting a new load into the
1502 // predecessor when it's not available. We could do this in general, but
1503 // prefer to not increase code size. As such, we only do this when we know
1504 // that we only have to insert *one* load (which means we're basically moving
1505 // the load, not inserting a new one).
1507 SmallPtrSet<BasicBlock *, 4> Blockers;
1508 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1509 Blockers.insert(UnavailableBlocks[i]);
1511 // Let's find the first basic block with more than one predecessor. Walk
1512 // backwards through predecessors if needed.
1513 BasicBlock *LoadBB = LI->getParent();
1514 BasicBlock *TmpBB = LoadBB;
1516 bool isSinglePred = false;
1517 bool allSingleSucc = true;
1518 while (TmpBB->getSinglePredecessor()) {
1519 isSinglePred = true;
1520 TmpBB = TmpBB->getSinglePredecessor();
1521 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1523 if (Blockers.count(TmpBB))
1526 // If any of these blocks has more than one successor (i.e. if the edge we
1527 // just traversed was critical), then there are other paths through this
1528 // block along which the load may not be anticipated. Hoisting the load
1529 // above this block would be adding the load to execution paths along
1530 // which it was not previously executed.
1531 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1538 // FIXME: It is extremely unclear what this loop is doing, other than
1539 // artificially restricting loadpre.
1542 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1543 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1544 if (AV.isSimpleValue())
1545 // "Hot" Instruction is in some loop (because it dominates its dep.
1547 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1548 if (DT->dominates(LI, I)) {
1554 // We are interested only in "hot" instructions. We don't want to do any
1555 // mis-optimizations here.
1560 // Check to see how many predecessors have the loaded value fully
1562 DenseMap<BasicBlock*, Value*> PredLoads;
1563 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1564 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1565 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1566 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1567 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1569 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
1570 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1572 BasicBlock *Pred = *PI;
1573 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) {
1576 PredLoads[Pred] = 0;
1578 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1579 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1580 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1581 << Pred->getName() << "': " << *LI << '\n');
1585 if (LoadBB->isLandingPad()) {
1587 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1588 << Pred->getName() << "': " << *LI << '\n');
1592 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
1593 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
1597 if (!NeedToSplit.empty()) {
1598 toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
1602 // Decide whether PRE is profitable for this load.
1603 unsigned NumUnavailablePreds = PredLoads.size();
1604 assert(NumUnavailablePreds != 0 &&
1605 "Fully available value should be eliminated above!");
1607 // If this load is unavailable in multiple predecessors, reject it.
1608 // FIXME: If we could restructure the CFG, we could make a common pred with
1609 // all the preds that don't have an available LI and insert a new load into
1611 if (NumUnavailablePreds != 1)
1614 // Check if the load can safely be moved to all the unavailable predecessors.
1615 bool CanDoPRE = true;
1616 SmallVector<Instruction*, 8> NewInsts;
1617 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1618 E = PredLoads.end(); I != E; ++I) {
1619 BasicBlock *UnavailablePred = I->first;
1621 // Do PHI translation to get its value in the predecessor if necessary. The
1622 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1624 // If all preds have a single successor, then we know it is safe to insert
1625 // the load on the pred (?!?), so we can insert code to materialize the
1626 // pointer if it is not available.
1627 PHITransAddr Address(LI->getPointerOperand(), TD);
1629 if (allSingleSucc) {
1630 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1633 Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
1634 LoadPtr = Address.getAddr();
1637 // If we couldn't find or insert a computation of this phi translated value,
1640 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1641 << *LI->getPointerOperand() << "\n");
1646 // Make sure it is valid to move this load here. We have to watch out for:
1647 // @1 = getelementptr (i8* p, ...
1648 // test p and branch if == 0
1650 // It is valid to have the getelementptr before the test, even if p can
1651 // be 0, as getelementptr only does address arithmetic.
1652 // If we are not pushing the value through any multiple-successor blocks
1653 // we do not have this case. Otherwise, check that the load is safe to
1654 // put anywhere; this can be improved, but should be conservatively safe.
1655 if (!allSingleSucc &&
1656 // FIXME: REEVALUTE THIS.
1657 !isSafeToLoadUnconditionally(LoadPtr,
1658 UnavailablePred->getTerminator(),
1659 LI->getAlignment(), TD)) {
1664 I->second = LoadPtr;
1668 while (!NewInsts.empty()) {
1669 Instruction *I = NewInsts.pop_back_val();
1670 if (MD) MD->removeInstruction(I);
1671 I->eraseFromParent();
1676 // Okay, we can eliminate this load by inserting a reload in the predecessor
1677 // and using PHI construction to get the value in the other predecessors, do
1679 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1680 DEBUG(if (!NewInsts.empty())
1681 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1682 << *NewInsts.back() << '\n');
1684 // Assign value numbers to the new instructions.
1685 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1686 // FIXME: We really _ought_ to insert these value numbers into their
1687 // parent's availability map. However, in doing so, we risk getting into
1688 // ordering issues. If a block hasn't been processed yet, we would be
1689 // marking a value as AVAIL-IN, which isn't what we intend.
1690 VN.lookup_or_add(NewInsts[i]);
1693 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1694 E = PredLoads.end(); I != E; ++I) {
1695 BasicBlock *UnavailablePred = I->first;
1696 Value *LoadPtr = I->second;
1698 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1700 UnavailablePred->getTerminator());
1702 // Transfer the old load's TBAA tag to the new load.
1703 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1704 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1706 // Transfer DebugLoc.
1707 NewLoad->setDebugLoc(LI->getDebugLoc());
1709 // Add the newly created load.
1710 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1712 MD->invalidateCachedPointerInfo(LoadPtr);
1713 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1716 // Perform PHI construction.
1717 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1718 LI->replaceAllUsesWith(V);
1719 if (isa<PHINode>(V))
1721 if (V->getType()->isPointerTy())
1722 MD->invalidateCachedPointerInfo(V);
1723 markInstructionForDeletion(LI);
1728 /// processLoad - Attempt to eliminate a load, first by eliminating it
1729 /// locally, and then attempting non-local elimination if that fails.
1730 bool GVN::processLoad(LoadInst *L) {
1737 if (L->use_empty()) {
1738 markInstructionForDeletion(L);
1742 // ... to a pointer that has been loaded from before...
1743 MemDepResult Dep = MD->getDependency(L);
1745 // If we have a clobber and target data is around, see if this is a clobber
1746 // that we can fix up through code synthesis.
1747 if (Dep.isClobber() && TD) {
1748 // Check to see if we have something like this:
1749 // store i32 123, i32* %P
1750 // %A = bitcast i32* %P to i8*
1751 // %B = gep i8* %A, i32 1
1754 // We could do that by recognizing if the clobber instructions are obviously
1755 // a common base + constant offset, and if the previous store (or memset)
1756 // completely covers this load. This sort of thing can happen in bitfield
1758 Value *AvailVal = 0;
1759 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1760 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1761 L->getPointerOperand(),
1764 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1765 L->getType(), L, *TD);
1768 // Check to see if we have something like this:
1771 // if we have this, replace the later with an extraction from the former.
1772 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1773 // If this is a clobber and L is the first instruction in its block, then
1774 // we have the first instruction in the entry block.
1778 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1779 L->getPointerOperand(),
1782 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1785 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1786 // a value on from it.
1787 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1788 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1789 L->getPointerOperand(),
1792 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
1796 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1797 << *AvailVal << '\n' << *L << "\n\n\n");
1799 // Replace the load!
1800 L->replaceAllUsesWith(AvailVal);
1801 if (AvailVal->getType()->isPointerTy())
1802 MD->invalidateCachedPointerInfo(AvailVal);
1803 markInstructionForDeletion(L);
1809 // If the value isn't available, don't do anything!
1810 if (Dep.isClobber()) {
1812 // fast print dep, using operator<< on instruction is too slow.
1813 dbgs() << "GVN: load ";
1814 WriteAsOperand(dbgs(), L);
1815 Instruction *I = Dep.getInst();
1816 dbgs() << " is clobbered by " << *I << '\n';
1821 // If it is defined in another block, try harder.
1822 if (Dep.isNonLocal())
1823 return processNonLocalLoad(L);
1827 // fast print dep, using operator<< on instruction is too slow.
1828 dbgs() << "GVN: load ";
1829 WriteAsOperand(dbgs(), L);
1830 dbgs() << " has unknown dependence\n";
1835 Instruction *DepInst = Dep.getInst();
1836 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1837 Value *StoredVal = DepSI->getValueOperand();
1839 // The store and load are to a must-aliased pointer, but they may not
1840 // actually have the same type. See if we know how to reuse the stored
1841 // value (depending on its type).
1842 if (StoredVal->getType() != L->getType()) {
1844 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1849 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1850 << '\n' << *L << "\n\n\n");
1857 L->replaceAllUsesWith(StoredVal);
1858 if (StoredVal->getType()->isPointerTy())
1859 MD->invalidateCachedPointerInfo(StoredVal);
1860 markInstructionForDeletion(L);
1865 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1866 Value *AvailableVal = DepLI;
1868 // The loads are of a must-aliased pointer, but they may not actually have
1869 // the same type. See if we know how to reuse the previously loaded value
1870 // (depending on its type).
1871 if (DepLI->getType() != L->getType()) {
1873 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1875 if (AvailableVal == 0)
1878 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1879 << "\n" << *L << "\n\n\n");
1886 L->replaceAllUsesWith(AvailableVal);
1887 if (DepLI->getType()->isPointerTy())
1888 MD->invalidateCachedPointerInfo(DepLI);
1889 markInstructionForDeletion(L);
1894 // If this load really doesn't depend on anything, then we must be loading an
1895 // undef value. This can happen when loading for a fresh allocation with no
1896 // intervening stores, for example.
1897 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1898 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1899 markInstructionForDeletion(L);
1904 // If this load occurs either right after a lifetime begin,
1905 // then the loaded value is undefined.
1906 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1907 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1908 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1909 markInstructionForDeletion(L);
1918 // findLeader - In order to find a leader for a given value number at a
1919 // specific basic block, we first obtain the list of all Values for that number,
1920 // and then scan the list to find one whose block dominates the block in
1921 // question. This is fast because dominator tree queries consist of only
1922 // a few comparisons of DFS numbers.
1923 Value *GVN::findLeader(BasicBlock *BB, uint32_t num) {
1924 LeaderTableEntry Vals = LeaderTable[num];
1925 if (!Vals.Val) return 0;
1928 if (DT->dominates(Vals.BB, BB)) {
1930 if (isa<Constant>(Val)) return Val;
1933 LeaderTableEntry* Next = Vals.Next;
1935 if (DT->dominates(Next->BB, BB)) {
1936 if (isa<Constant>(Next->Val)) return Next->Val;
1937 if (!Val) Val = Next->Val;
1946 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
1947 /// use is dominated by the given basic block. Returns the number of uses that
1949 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
1952 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1954 Use &U = (UI++).getUse();
1956 // If From occurs as a phi node operand then the use implicitly lives in the
1957 // corresponding incoming block. Otherwise it is the block containing the
1958 // user that must be dominated by Root.
1959 BasicBlock *UsingBlock;
1960 if (PHINode *PN = dyn_cast<PHINode>(U.getUser()))
1961 UsingBlock = PN->getIncomingBlock(U);
1963 UsingBlock = cast<Instruction>(U.getUser())->getParent();
1965 if (DT->dominates(Root, UsingBlock)) {
1973 /// propagateEquality - The given values are known to be equal in every block
1974 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1975 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1976 bool GVN::propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root) {
1977 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
1978 Worklist.push_back(std::make_pair(LHS, RHS));
1979 bool Changed = false;
1981 while (!Worklist.empty()) {
1982 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1983 LHS = Item.first; RHS = Item.second;
1985 if (LHS == RHS) continue;
1986 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1988 // Don't try to propagate equalities between constants.
1989 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
1991 // Prefer a constant on the right-hand side, or an Argument if no constants.
1992 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1993 std::swap(LHS, RHS);
1994 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1996 // If there is no obvious reason to prefer the left-hand side over the right-
1997 // hand side, ensure the longest lived term is on the right-hand side, so the
1998 // shortest lived term will be replaced by the longest lived. This tends to
1999 // expose more simplifications.
2000 uint32_t LVN = VN.lookup_or_add(LHS);
2001 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2002 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2003 // Move the 'oldest' value to the right-hand side, using the value number as
2005 uint32_t RVN = VN.lookup_or_add(RHS);
2007 std::swap(LHS, RHS);
2011 assert((!isa<Instruction>(RHS) ||
2012 DT->properlyDominates(cast<Instruction>(RHS)->getParent(), Root)) &&
2013 "Instruction doesn't dominate scope!");
2015 // If value numbering later deduces that an instruction in the scope is equal
2016 // to 'LHS' then ensure it will be turned into 'RHS'.
2017 addToLeaderTable(LVN, RHS, Root);
2019 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2020 // LHS always has at least one use that is not dominated by Root, this will
2021 // never do anything if LHS has only one use.
2022 if (!LHS->hasOneUse()) {
2023 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2024 Changed |= NumReplacements > 0;
2025 NumGVNEqProp += NumReplacements;
2028 // Now try to deduce additional equalities from this one. For example, if the
2029 // known equality was "(A != B)" == "false" then it follows that A and B are
2030 // equal in the scope. Only boolean equalities with an explicit true or false
2031 // RHS are currently supported.
2032 if (!RHS->getType()->isIntegerTy(1))
2033 // Not a boolean equality - bail out.
2035 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2037 // RHS neither 'true' nor 'false' - bail out.
2039 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2040 bool isKnownTrue = CI->isAllOnesValue();
2041 bool isKnownFalse = !isKnownTrue;
2043 // If "A && B" is known true then both A and B are known true. If "A || B"
2044 // is known false then both A and B are known false.
2046 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2047 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2048 Worklist.push_back(std::make_pair(A, RHS));
2049 Worklist.push_back(std::make_pair(B, RHS));
2053 // If we are propagating an equality like "(A == B)" == "true" then also
2054 // propagate the equality A == B. When propagating a comparison such as
2055 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2056 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
2057 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2059 // If "A == B" is known true, or "A != B" is known false, then replace
2060 // A with B everywhere in the scope.
2061 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2062 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2063 Worklist.push_back(std::make_pair(Op0, Op1));
2065 // If "A >= B" is known true, replace "A < B" with false everywhere.
2066 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2067 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2068 // Since we don't have the instruction "A < B" immediately to hand, work out
2069 // the value number that it would have and use that to find an appropriate
2070 // instruction (if any).
2071 uint32_t NextNum = VN.getNextUnusedValueNumber();
2072 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2073 // If the number we were assigned was brand new then there is no point in
2074 // looking for an instruction realizing it: there cannot be one!
2075 if (Num < NextNum) {
2076 Value *NotCmp = findLeader(Root, Num);
2077 if (NotCmp && isa<Instruction>(NotCmp)) {
2078 unsigned NumReplacements =
2079 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2080 Changed |= NumReplacements > 0;
2081 NumGVNEqProp += NumReplacements;
2084 // Ensure that any instruction in scope that gets the "A < B" value number
2085 // is replaced with false.
2086 addToLeaderTable(Num, NotVal, Root);
2095 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
2096 /// true if every path from the entry block to 'Dst' passes via this edge. In
2097 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2098 static bool isOnlyReachableViaThisEdge(BasicBlock *Src, BasicBlock *Dst,
2099 DominatorTree *DT) {
2100 // While in theory it is interesting to consider the case in which Dst has
2101 // more than one predecessor, because Dst might be part of a loop which is
2102 // only reachable from Src, in practice it is pointless since at the time
2103 // GVN runs all such loops have preheaders, which means that Dst will have
2104 // been changed to have only one predecessor, namely Src.
2105 BasicBlock *Pred = Dst->getSinglePredecessor();
2106 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2111 /// processInstruction - When calculating availability, handle an instruction
2112 /// by inserting it into the appropriate sets
2113 bool GVN::processInstruction(Instruction *I) {
2114 // Ignore dbg info intrinsics.
2115 if (isa<DbgInfoIntrinsic>(I))
2118 // If the instruction can be easily simplified then do so now in preference
2119 // to value numbering it. Value numbering often exposes redundancies, for
2120 // example if it determines that %y is equal to %x then the instruction
2121 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2122 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) {
2123 I->replaceAllUsesWith(V);
2124 if (MD && V->getType()->isPointerTy())
2125 MD->invalidateCachedPointerInfo(V);
2126 markInstructionForDeletion(I);
2131 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2132 if (processLoad(LI))
2135 unsigned Num = VN.lookup_or_add(LI);
2136 addToLeaderTable(Num, LI, LI->getParent());
2140 // For conditional branches, we can perform simple conditional propagation on
2141 // the condition value itself.
2142 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2143 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
2146 Value *BranchCond = BI->getCondition();
2148 BasicBlock *TrueSucc = BI->getSuccessor(0);
2149 BasicBlock *FalseSucc = BI->getSuccessor(1);
2150 BasicBlock *Parent = BI->getParent();
2151 bool Changed = false;
2153 if (isOnlyReachableViaThisEdge(Parent, TrueSucc, DT))
2154 Changed |= propagateEquality(BranchCond,
2155 ConstantInt::getTrue(TrueSucc->getContext()),
2158 if (isOnlyReachableViaThisEdge(Parent, FalseSucc, DT))
2159 Changed |= propagateEquality(BranchCond,
2160 ConstantInt::getFalse(FalseSucc->getContext()),
2166 // For switches, propagate the case values into the case destinations.
2167 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2168 Value *SwitchCond = SI->getCondition();
2169 BasicBlock *Parent = SI->getParent();
2170 bool Changed = false;
2171 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2173 BasicBlock *Dst = i.getCaseSuccessor();
2174 if (isOnlyReachableViaThisEdge(Parent, Dst, DT))
2175 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), Dst);
2180 // Instructions with void type don't return a value, so there's
2181 // no point in trying to find redundancies in them.
2182 if (I->getType()->isVoidTy()) return false;
2184 uint32_t NextNum = VN.getNextUnusedValueNumber();
2185 unsigned Num = VN.lookup_or_add(I);
2187 // Allocations are always uniquely numbered, so we can save time and memory
2188 // by fast failing them.
2189 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2190 addToLeaderTable(Num, I, I->getParent());
2194 // If the number we were assigned was a brand new VN, then we don't
2195 // need to do a lookup to see if the number already exists
2196 // somewhere in the domtree: it can't!
2197 if (Num >= NextNum) {
2198 addToLeaderTable(Num, I, I->getParent());
2202 // Perform fast-path value-number based elimination of values inherited from
2204 Value *repl = findLeader(I->getParent(), Num);
2206 // Failure, just remember this instance for future use.
2207 addToLeaderTable(Num, I, I->getParent());
2212 I->replaceAllUsesWith(repl);
2213 if (MD && repl->getType()->isPointerTy())
2214 MD->invalidateCachedPointerInfo(repl);
2215 markInstructionForDeletion(I);
2219 /// runOnFunction - This is the main transformation entry point for a function.
2220 bool GVN::runOnFunction(Function& F) {
2222 MD = &getAnalysis<MemoryDependenceAnalysis>();
2223 DT = &getAnalysis<DominatorTree>();
2224 TD = getAnalysisIfAvailable<TargetData>();
2225 TLI = &getAnalysis<TargetLibraryInfo>();
2226 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2230 bool Changed = false;
2231 bool ShouldContinue = true;
2233 // Merge unconditional branches, allowing PRE to catch more
2234 // optimization opportunities.
2235 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2236 BasicBlock *BB = FI++;
2238 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2239 if (removedBlock) ++NumGVNBlocks;
2241 Changed |= removedBlock;
2244 unsigned Iteration = 0;
2245 while (ShouldContinue) {
2246 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2247 ShouldContinue = iterateOnFunction(F);
2248 if (splitCriticalEdges())
2249 ShouldContinue = true;
2250 Changed |= ShouldContinue;
2255 bool PREChanged = true;
2256 while (PREChanged) {
2257 PREChanged = performPRE(F);
2258 Changed |= PREChanged;
2261 // FIXME: Should perform GVN again after PRE does something. PRE can move
2262 // computations into blocks where they become fully redundant. Note that
2263 // we can't do this until PRE's critical edge splitting updates memdep.
2264 // Actually, when this happens, we should just fully integrate PRE into GVN.
2266 cleanupGlobalSets();
2272 bool GVN::processBlock(BasicBlock *BB) {
2273 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2274 // (and incrementing BI before processing an instruction).
2275 assert(InstrsToErase.empty() &&
2276 "We expect InstrsToErase to be empty across iterations");
2277 bool ChangedFunction = false;
2279 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2281 ChangedFunction |= processInstruction(BI);
2282 if (InstrsToErase.empty()) {
2287 // If we need some instructions deleted, do it now.
2288 NumGVNInstr += InstrsToErase.size();
2290 // Avoid iterator invalidation.
2291 bool AtStart = BI == BB->begin();
2295 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(),
2296 E = InstrsToErase.end(); I != E; ++I) {
2297 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2298 if (MD) MD->removeInstruction(*I);
2299 (*I)->eraseFromParent();
2300 DEBUG(verifyRemoved(*I));
2302 InstrsToErase.clear();
2310 return ChangedFunction;
2313 /// performPRE - Perform a purely local form of PRE that looks for diamond
2314 /// control flow patterns and attempts to perform simple PRE at the join point.
2315 bool GVN::performPRE(Function &F) {
2316 bool Changed = false;
2317 DenseMap<BasicBlock*, Value*> predMap;
2318 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2319 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2320 BasicBlock *CurrentBlock = *DI;
2322 // Nothing to PRE in the entry block.
2323 if (CurrentBlock == &F.getEntryBlock()) continue;
2325 // Don't perform PRE on a landing pad.
2326 if (CurrentBlock->isLandingPad()) continue;
2328 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2329 BE = CurrentBlock->end(); BI != BE; ) {
2330 Instruction *CurInst = BI++;
2332 if (isa<AllocaInst>(CurInst) ||
2333 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2334 CurInst->getType()->isVoidTy() ||
2335 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2336 isa<DbgInfoIntrinsic>(CurInst))
2339 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2340 // sinking the compare again, and it would force the code generator to
2341 // move the i1 from processor flags or predicate registers into a general
2342 // purpose register.
2343 if (isa<CmpInst>(CurInst))
2346 // We don't currently value number ANY inline asm calls.
2347 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2348 if (CallI->isInlineAsm())
2351 uint32_t ValNo = VN.lookup(CurInst);
2353 // Look for the predecessors for PRE opportunities. We're
2354 // only trying to solve the basic diamond case, where
2355 // a value is computed in the successor and one predecessor,
2356 // but not the other. We also explicitly disallow cases
2357 // where the successor is its own predecessor, because they're
2358 // more complicated to get right.
2359 unsigned NumWith = 0;
2360 unsigned NumWithout = 0;
2361 BasicBlock *PREPred = 0;
2364 for (pred_iterator PI = pred_begin(CurrentBlock),
2365 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2366 BasicBlock *P = *PI;
2367 // We're not interested in PRE where the block is its
2368 // own predecessor, or in blocks with predecessors
2369 // that are not reachable.
2370 if (P == CurrentBlock) {
2373 } else if (!DT->dominates(&F.getEntryBlock(), P)) {
2378 Value* predV = findLeader(P, ValNo);
2382 } else if (predV == CurInst) {
2390 // Don't do PRE when it might increase code size, i.e. when
2391 // we would need to insert instructions in more than one pred.
2392 if (NumWithout != 1 || NumWith == 0)
2395 // Don't do PRE across indirect branch.
2396 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2399 // We can't do PRE safely on a critical edge, so instead we schedule
2400 // the edge to be split and perform the PRE the next time we iterate
2402 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2403 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2404 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2408 // Instantiate the expression in the predecessor that lacked it.
2409 // Because we are going top-down through the block, all value numbers
2410 // will be available in the predecessor by the time we need them. Any
2411 // that weren't originally present will have been instantiated earlier
2413 Instruction *PREInstr = CurInst->clone();
2414 bool success = true;
2415 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2416 Value *Op = PREInstr->getOperand(i);
2417 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2420 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2421 PREInstr->setOperand(i, V);
2428 // Fail out if we encounter an operand that is not available in
2429 // the PRE predecessor. This is typically because of loads which
2430 // are not value numbered precisely.
2433 DEBUG(verifyRemoved(PREInstr));
2437 PREInstr->insertBefore(PREPred->getTerminator());
2438 PREInstr->setName(CurInst->getName() + ".pre");
2439 PREInstr->setDebugLoc(CurInst->getDebugLoc());
2440 predMap[PREPred] = PREInstr;
2441 VN.add(PREInstr, ValNo);
2444 // Update the availability map to include the new instruction.
2445 addToLeaderTable(ValNo, PREInstr, PREPred);
2447 // Create a PHI to make the value available in this block.
2448 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2449 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE),
2450 CurInst->getName() + ".pre-phi",
2451 CurrentBlock->begin());
2452 for (pred_iterator PI = PB; PI != PE; ++PI) {
2453 BasicBlock *P = *PI;
2454 Phi->addIncoming(predMap[P], P);
2458 addToLeaderTable(ValNo, Phi, CurrentBlock);
2459 Phi->setDebugLoc(CurInst->getDebugLoc());
2460 CurInst->replaceAllUsesWith(Phi);
2461 if (Phi->getType()->isPointerTy()) {
2462 // Because we have added a PHI-use of the pointer value, it has now
2463 // "escaped" from alias analysis' perspective. We need to inform
2465 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
2467 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2468 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2472 MD->invalidateCachedPointerInfo(Phi);
2475 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2477 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2478 if (MD) MD->removeInstruction(CurInst);
2479 CurInst->eraseFromParent();
2480 DEBUG(verifyRemoved(CurInst));
2485 if (splitCriticalEdges())
2491 /// splitCriticalEdges - Split critical edges found during the previous
2492 /// iteration that may enable further optimization.
2493 bool GVN::splitCriticalEdges() {
2494 if (toSplit.empty())
2497 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2498 SplitCriticalEdge(Edge.first, Edge.second, this);
2499 } while (!toSplit.empty());
2500 if (MD) MD->invalidateCachedPredecessors();
2504 /// iterateOnFunction - Executes one iteration of GVN
2505 bool GVN::iterateOnFunction(Function &F) {
2506 cleanupGlobalSets();
2508 // Top-down walk of the dominator tree
2509 bool Changed = false;
2511 // Needed for value numbering with phi construction to work.
2512 ReversePostOrderTraversal<Function*> RPOT(&F);
2513 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2514 RE = RPOT.end(); RI != RE; ++RI)
2515 Changed |= processBlock(*RI);
2517 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2518 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2519 Changed |= processBlock(DI->getBlock());
2525 void GVN::cleanupGlobalSets() {
2527 LeaderTable.clear();
2528 TableAllocator.Reset();
2531 /// verifyRemoved - Verify that the specified instruction does not occur in our
2532 /// internal data structures.
2533 void GVN::verifyRemoved(const Instruction *Inst) const {
2534 VN.verifyRemoved(Inst);
2536 // Walk through the value number scope to make sure the instruction isn't
2537 // ferreted away in it.
2538 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2539 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2540 const LeaderTableEntry *Node = &I->second;
2541 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2543 while (Node->Next) {
2545 assert(Node->Val != Inst && "Inst still in value numbering scope!");