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 // Maximum allowed recursion depth.
63 static cl::opt<uint32_t>
64 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
65 cl::desc("Max recurse depth (default = 1000)"));
67 //===----------------------------------------------------------------------===//
69 //===----------------------------------------------------------------------===//
71 /// This class holds the mapping between values and value numbers. It is used
72 /// as an efficient mechanism to determine the expression-wise equivalence of
78 SmallVector<uint32_t, 4> varargs;
80 Expression(uint32_t o = ~2U) : opcode(o) { }
82 bool operator==(const Expression &other) const {
83 if (opcode != other.opcode)
85 if (opcode == ~0U || opcode == ~1U)
87 if (type != other.type)
89 if (varargs != other.varargs)
94 friend hash_code hash_value(const Expression &Value) {
95 return hash_combine(Value.opcode, Value.type,
96 hash_combine_range(Value.varargs.begin(),
97 Value.varargs.end()));
102 DenseMap<Value*, uint32_t> valueNumbering;
103 DenseMap<Expression, uint32_t> expressionNumbering;
105 MemoryDependenceAnalysis *MD;
108 uint32_t nextValueNumber;
110 Expression create_expression(Instruction* I);
111 Expression create_cmp_expression(unsigned Opcode,
112 CmpInst::Predicate Predicate,
113 Value *LHS, Value *RHS);
114 Expression create_extractvalue_expression(ExtractValueInst* EI);
115 uint32_t lookup_or_add_call(CallInst* C);
117 ValueTable() : nextValueNumber(1) { }
118 uint32_t lookup_or_add(Value *V);
119 uint32_t lookup(Value *V) const;
120 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
121 Value *LHS, Value *RHS);
122 void add(Value *V, uint32_t num);
124 void erase(Value *v);
125 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
126 AliasAnalysis *getAliasAnalysis() const { return AA; }
127 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
128 void setDomTree(DominatorTree* D) { DT = D; }
129 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
130 void verifyRemoved(const Value *) const;
135 template <> struct DenseMapInfo<Expression> {
136 static inline Expression getEmptyKey() {
140 static inline Expression getTombstoneKey() {
144 static unsigned getHashValue(const Expression e) {
145 using llvm::hash_value;
146 return static_cast<unsigned>(hash_value(e));
148 static bool isEqual(const Expression &LHS, const Expression &RHS) {
155 //===----------------------------------------------------------------------===//
156 // ValueTable Internal Functions
157 //===----------------------------------------------------------------------===//
159 Expression ValueTable::create_expression(Instruction *I) {
161 e.type = I->getType();
162 e.opcode = I->getOpcode();
163 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
165 e.varargs.push_back(lookup_or_add(*OI));
166 if (I->isCommutative()) {
167 // Ensure that commutative instructions that only differ by a permutation
168 // of their operands get the same value number by sorting the operand value
169 // numbers. Since all commutative instructions have two operands it is more
170 // efficient to sort by hand rather than using, say, std::sort.
171 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
172 if (e.varargs[0] > e.varargs[1])
173 std::swap(e.varargs[0], e.varargs[1]);
176 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
177 // Sort the operand value numbers so x<y and y>x get the same value number.
178 CmpInst::Predicate Predicate = C->getPredicate();
179 if (e.varargs[0] > e.varargs[1]) {
180 std::swap(e.varargs[0], e.varargs[1]);
181 Predicate = CmpInst::getSwappedPredicate(Predicate);
183 e.opcode = (C->getOpcode() << 8) | Predicate;
184 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
185 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
187 e.varargs.push_back(*II);
193 Expression ValueTable::create_cmp_expression(unsigned Opcode,
194 CmpInst::Predicate Predicate,
195 Value *LHS, Value *RHS) {
196 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
197 "Not a comparison!");
199 e.type = CmpInst::makeCmpResultType(LHS->getType());
200 e.varargs.push_back(lookup_or_add(LHS));
201 e.varargs.push_back(lookup_or_add(RHS));
203 // Sort the operand value numbers so x<y and y>x get the same value number.
204 if (e.varargs[0] > e.varargs[1]) {
205 std::swap(e.varargs[0], e.varargs[1]);
206 Predicate = CmpInst::getSwappedPredicate(Predicate);
208 e.opcode = (Opcode << 8) | Predicate;
212 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
213 assert(EI != 0 && "Not an ExtractValueInst?");
215 e.type = EI->getType();
218 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
219 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
220 // EI might be an extract from one of our recognised intrinsics. If it
221 // is we'll synthesize a semantically equivalent expression instead on
222 // an extract value expression.
223 switch (I->getIntrinsicID()) {
224 case Intrinsic::sadd_with_overflow:
225 case Intrinsic::uadd_with_overflow:
226 e.opcode = Instruction::Add;
228 case Intrinsic::ssub_with_overflow:
229 case Intrinsic::usub_with_overflow:
230 e.opcode = Instruction::Sub;
232 case Intrinsic::smul_with_overflow:
233 case Intrinsic::umul_with_overflow:
234 e.opcode = Instruction::Mul;
241 // Intrinsic recognized. Grab its args to finish building the expression.
242 assert(I->getNumArgOperands() == 2 &&
243 "Expect two args for recognised intrinsics.");
244 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
245 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
250 // Not a recognised intrinsic. Fall back to producing an extract value
252 e.opcode = EI->getOpcode();
253 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
255 e.varargs.push_back(lookup_or_add(*OI));
257 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
259 e.varargs.push_back(*II);
264 //===----------------------------------------------------------------------===//
265 // ValueTable External Functions
266 //===----------------------------------------------------------------------===//
268 /// add - Insert a value into the table with a specified value number.
269 void ValueTable::add(Value *V, uint32_t num) {
270 valueNumbering.insert(std::make_pair(V, num));
273 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
274 if (AA->doesNotAccessMemory(C)) {
275 Expression exp = create_expression(C);
276 uint32_t& e = expressionNumbering[exp];
277 if (!e) e = nextValueNumber++;
278 valueNumbering[C] = e;
280 } else if (AA->onlyReadsMemory(C)) {
281 Expression exp = create_expression(C);
282 uint32_t& e = expressionNumbering[exp];
284 e = nextValueNumber++;
285 valueNumbering[C] = e;
289 e = nextValueNumber++;
290 valueNumbering[C] = e;
294 MemDepResult local_dep = MD->getDependency(C);
296 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
297 valueNumbering[C] = nextValueNumber;
298 return nextValueNumber++;
301 if (local_dep.isDef()) {
302 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
304 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
305 valueNumbering[C] = nextValueNumber;
306 return nextValueNumber++;
309 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
310 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
311 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
313 valueNumbering[C] = nextValueNumber;
314 return nextValueNumber++;
318 uint32_t v = lookup_or_add(local_cdep);
319 valueNumbering[C] = v;
324 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
325 MD->getNonLocalCallDependency(CallSite(C));
326 // FIXME: Move the checking logic to MemDep!
329 // Check to see if we have a single dominating call instruction that is
331 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
332 const NonLocalDepEntry *I = &deps[i];
333 if (I->getResult().isNonLocal())
336 // We don't handle non-definitions. If we already have a call, reject
337 // instruction dependencies.
338 if (!I->getResult().isDef() || cdep != 0) {
343 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
344 // FIXME: All duplicated with non-local case.
345 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
346 cdep = NonLocalDepCall;
355 valueNumbering[C] = nextValueNumber;
356 return nextValueNumber++;
359 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
360 valueNumbering[C] = nextValueNumber;
361 return nextValueNumber++;
363 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
364 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
365 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
367 valueNumbering[C] = nextValueNumber;
368 return nextValueNumber++;
372 uint32_t v = lookup_or_add(cdep);
373 valueNumbering[C] = v;
377 valueNumbering[C] = nextValueNumber;
378 return nextValueNumber++;
382 /// lookup_or_add - Returns the value number for the specified value, assigning
383 /// it a new number if it did not have one before.
384 uint32_t ValueTable::lookup_or_add(Value *V) {
385 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
386 if (VI != valueNumbering.end())
389 if (!isa<Instruction>(V)) {
390 valueNumbering[V] = nextValueNumber;
391 return nextValueNumber++;
394 Instruction* I = cast<Instruction>(V);
396 switch (I->getOpcode()) {
397 case Instruction::Call:
398 return lookup_or_add_call(cast<CallInst>(I));
399 case Instruction::Add:
400 case Instruction::FAdd:
401 case Instruction::Sub:
402 case Instruction::FSub:
403 case Instruction::Mul:
404 case Instruction::FMul:
405 case Instruction::UDiv:
406 case Instruction::SDiv:
407 case Instruction::FDiv:
408 case Instruction::URem:
409 case Instruction::SRem:
410 case Instruction::FRem:
411 case Instruction::Shl:
412 case Instruction::LShr:
413 case Instruction::AShr:
414 case Instruction::And:
415 case Instruction::Or :
416 case Instruction::Xor:
417 case Instruction::ICmp:
418 case Instruction::FCmp:
419 case Instruction::Trunc:
420 case Instruction::ZExt:
421 case Instruction::SExt:
422 case Instruction::FPToUI:
423 case Instruction::FPToSI:
424 case Instruction::UIToFP:
425 case Instruction::SIToFP:
426 case Instruction::FPTrunc:
427 case Instruction::FPExt:
428 case Instruction::PtrToInt:
429 case Instruction::IntToPtr:
430 case Instruction::BitCast:
431 case Instruction::Select:
432 case Instruction::ExtractElement:
433 case Instruction::InsertElement:
434 case Instruction::ShuffleVector:
435 case Instruction::InsertValue:
436 case Instruction::GetElementPtr:
437 exp = create_expression(I);
439 case Instruction::ExtractValue:
440 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
443 valueNumbering[V] = nextValueNumber;
444 return nextValueNumber++;
447 uint32_t& e = expressionNumbering[exp];
448 if (!e) e = nextValueNumber++;
449 valueNumbering[V] = e;
453 /// lookup - Returns the value number of the specified value. Fails if
454 /// the value has not yet been numbered.
455 uint32_t ValueTable::lookup(Value *V) const {
456 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
457 assert(VI != valueNumbering.end() && "Value not numbered?");
461 /// lookup_or_add_cmp - Returns the value number of the given comparison,
462 /// assigning it a new number if it did not have one before. Useful when
463 /// we deduced the result of a comparison, but don't immediately have an
464 /// instruction realizing that comparison to hand.
465 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
466 CmpInst::Predicate Predicate,
467 Value *LHS, Value *RHS) {
468 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
469 uint32_t& e = expressionNumbering[exp];
470 if (!e) e = nextValueNumber++;
474 /// clear - Remove all entries from the ValueTable.
475 void ValueTable::clear() {
476 valueNumbering.clear();
477 expressionNumbering.clear();
481 /// erase - Remove a value from the value numbering.
482 void ValueTable::erase(Value *V) {
483 valueNumbering.erase(V);
486 /// verifyRemoved - Verify that the value is removed from all internal data
488 void ValueTable::verifyRemoved(const Value *V) const {
489 for (DenseMap<Value*, uint32_t>::const_iterator
490 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
491 assert(I->first != V && "Inst still occurs in value numbering map!");
495 //===----------------------------------------------------------------------===//
497 //===----------------------------------------------------------------------===//
501 class GVN : public FunctionPass {
503 MemoryDependenceAnalysis *MD;
505 const TargetData *TD;
506 const TargetLibraryInfo *TLI;
510 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
511 /// have that value number. Use findLeader to query it.
512 struct LeaderTableEntry {
515 LeaderTableEntry *Next;
517 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
518 BumpPtrAllocator TableAllocator;
520 SmallVector<Instruction*, 8> InstrsToErase;
522 static char ID; // Pass identification, replacement for typeid
523 explicit GVN(bool noloads = false)
524 : FunctionPass(ID), NoLoads(noloads), MD(0) {
525 initializeGVNPass(*PassRegistry::getPassRegistry());
528 bool runOnFunction(Function &F);
530 /// markInstructionForDeletion - This removes the specified instruction from
531 /// our various maps and marks it for deletion.
532 void markInstructionForDeletion(Instruction *I) {
534 InstrsToErase.push_back(I);
537 const TargetData *getTargetData() const { return TD; }
538 DominatorTree &getDominatorTree() const { return *DT; }
539 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
540 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
542 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
543 /// its value number.
544 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
545 LeaderTableEntry &Curr = LeaderTable[N];
552 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
555 Node->Next = Curr.Next;
559 /// removeFromLeaderTable - Scan the list of values corresponding to a given
560 /// value number, and remove the given instruction if encountered.
561 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
562 LeaderTableEntry* Prev = 0;
563 LeaderTableEntry* Curr = &LeaderTable[N];
565 while (Curr->Val != I || Curr->BB != BB) {
571 Prev->Next = Curr->Next;
577 LeaderTableEntry* Next = Curr->Next;
578 Curr->Val = Next->Val;
580 Curr->Next = Next->Next;
585 // List of critical edges to be split between iterations.
586 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
588 // This transformation requires dominator postdominator info
589 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
590 AU.addRequired<DominatorTree>();
591 AU.addRequired<TargetLibraryInfo>();
593 AU.addRequired<MemoryDependenceAnalysis>();
594 AU.addRequired<AliasAnalysis>();
596 AU.addPreserved<DominatorTree>();
597 AU.addPreserved<AliasAnalysis>();
602 // FIXME: eliminate or document these better
603 bool processLoad(LoadInst *L);
604 bool processInstruction(Instruction *I);
605 bool processNonLocalLoad(LoadInst *L);
606 bool processBlock(BasicBlock *BB);
607 void dump(DenseMap<uint32_t, Value*> &d);
608 bool iterateOnFunction(Function &F);
609 bool performPRE(Function &F);
610 Value *findLeader(BasicBlock *BB, uint32_t num);
611 void cleanupGlobalSets();
612 void verifyRemoved(const Instruction *I) const;
613 bool splitCriticalEdges();
614 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
616 bool propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root);
622 // createGVNPass - The public interface to this file...
623 FunctionPass *llvm::createGVNPass(bool NoLoads) {
624 return new GVN(NoLoads);
627 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
628 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
629 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
630 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
631 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
632 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
634 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
636 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
637 E = d.end(); I != E; ++I) {
638 errs() << I->first << "\n";
644 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
645 /// we're analyzing is fully available in the specified block. As we go, keep
646 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
647 /// map is actually a tri-state map with the following values:
648 /// 0) we know the block *is not* fully available.
649 /// 1) we know the block *is* fully available.
650 /// 2) we do not know whether the block is fully available or not, but we are
651 /// currently speculating that it will be.
652 /// 3) we are speculating for this block and have used that to speculate for
654 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
655 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
656 uint32_t RecurseDepth) {
657 if (RecurseDepth > MaxRecurseDepth)
660 // Optimistically assume that the block is fully available and check to see
661 // if we already know about this block in one lookup.
662 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
663 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
665 // If the entry already existed for this block, return the precomputed value.
667 // If this is a speculative "available" value, mark it as being used for
668 // speculation of other blocks.
669 if (IV.first->second == 2)
670 IV.first->second = 3;
671 return IV.first->second != 0;
674 // Otherwise, see if it is fully available in all predecessors.
675 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
677 // If this block has no predecessors, it isn't live-in here.
679 goto SpeculationFailure;
681 for (; PI != PE; ++PI)
682 // If the value isn't fully available in one of our predecessors, then it
683 // isn't fully available in this block either. Undo our previous
684 // optimistic assumption and bail out.
685 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
686 goto SpeculationFailure;
690 // SpeculationFailure - If we get here, we found out that this is not, after
691 // all, a fully-available block. We have a problem if we speculated on this and
692 // used the speculation to mark other blocks as available.
694 char &BBVal = FullyAvailableBlocks[BB];
696 // If we didn't speculate on this, just return with it set to false.
702 // If we did speculate on this value, we could have blocks set to 1 that are
703 // incorrect. Walk the (transitive) successors of this block and mark them as
705 SmallVector<BasicBlock*, 32> BBWorklist;
706 BBWorklist.push_back(BB);
709 BasicBlock *Entry = BBWorklist.pop_back_val();
710 // Note that this sets blocks to 0 (unavailable) if they happen to not
711 // already be in FullyAvailableBlocks. This is safe.
712 char &EntryVal = FullyAvailableBlocks[Entry];
713 if (EntryVal == 0) continue; // Already unavailable.
715 // Mark as unavailable.
718 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
719 BBWorklist.push_back(*I);
720 } while (!BBWorklist.empty());
726 /// CanCoerceMustAliasedValueToLoad - Return true if
727 /// CoerceAvailableValueToLoadType will succeed.
728 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
730 const TargetData &TD) {
731 // If the loaded or stored value is an first class array or struct, don't try
732 // to transform them. We need to be able to bitcast to integer.
733 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
734 StoredVal->getType()->isStructTy() ||
735 StoredVal->getType()->isArrayTy())
738 // The store has to be at least as big as the load.
739 if (TD.getTypeSizeInBits(StoredVal->getType()) <
740 TD.getTypeSizeInBits(LoadTy))
747 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
748 /// then a load from a must-aliased pointer of a different type, try to coerce
749 /// the stored value. LoadedTy is the type of the load we want to replace and
750 /// InsertPt is the place to insert new instructions.
752 /// If we can't do it, return null.
753 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
755 Instruction *InsertPt,
756 const TargetData &TD) {
757 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
760 // If this is already the right type, just return it.
761 Type *StoredValTy = StoredVal->getType();
763 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
764 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
766 // If the store and reload are the same size, we can always reuse it.
767 if (StoreSize == LoadSize) {
768 // Pointer to Pointer -> use bitcast.
769 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy())
770 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
772 // Convert source pointers to integers, which can be bitcast.
773 if (StoredValTy->isPointerTy()) {
774 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
775 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
778 Type *TypeToCastTo = LoadedTy;
779 if (TypeToCastTo->isPointerTy())
780 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
782 if (StoredValTy != TypeToCastTo)
783 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
785 // Cast to pointer if the load needs a pointer type.
786 if (LoadedTy->isPointerTy())
787 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
792 // If the loaded value is smaller than the available value, then we can
793 // extract out a piece from it. If the available value is too small, then we
794 // can't do anything.
795 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
797 // Convert source pointers to integers, which can be manipulated.
798 if (StoredValTy->isPointerTy()) {
799 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
800 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
803 // Convert vectors and fp to integer, which can be manipulated.
804 if (!StoredValTy->isIntegerTy()) {
805 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
806 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
809 // If this is a big-endian system, we need to shift the value down to the low
810 // bits so that a truncate will work.
811 if (TD.isBigEndian()) {
812 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
813 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
816 // Truncate the integer to the right size now.
817 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
818 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
820 if (LoadedTy == NewIntTy)
823 // If the result is a pointer, inttoptr.
824 if (LoadedTy->isPointerTy())
825 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
827 // Otherwise, bitcast.
828 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
831 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
832 /// memdep query of a load that ends up being a clobbering memory write (store,
833 /// memset, memcpy, memmove). This means that the write *may* provide bits used
834 /// by the load but we can't be sure because the pointers don't mustalias.
836 /// Check this case to see if there is anything more we can do before we give
837 /// up. This returns -1 if we have to give up, or a byte number in the stored
838 /// value of the piece that feeds the load.
839 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
841 uint64_t WriteSizeInBits,
842 const TargetData &TD) {
843 // If the loaded or stored value is a first class array or struct, don't try
844 // to transform them. We need to be able to bitcast to integer.
845 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
848 int64_t StoreOffset = 0, LoadOffset = 0;
849 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD);
850 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
851 if (StoreBase != LoadBase)
854 // If the load and store are to the exact same address, they should have been
855 // a must alias. AA must have gotten confused.
856 // FIXME: Study to see if/when this happens. One case is forwarding a memset
857 // to a load from the base of the memset.
859 if (LoadOffset == StoreOffset) {
860 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
861 << "Base = " << *StoreBase << "\n"
862 << "Store Ptr = " << *WritePtr << "\n"
863 << "Store Offs = " << StoreOffset << "\n"
864 << "Load Ptr = " << *LoadPtr << "\n";
869 // If the load and store don't overlap at all, the store doesn't provide
870 // anything to the load. In this case, they really don't alias at all, AA
871 // must have gotten confused.
872 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
874 if ((WriteSizeInBits & 7) | (LoadSize & 7))
876 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
880 bool isAAFailure = false;
881 if (StoreOffset < LoadOffset)
882 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
884 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
888 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
889 << "Base = " << *StoreBase << "\n"
890 << "Store Ptr = " << *WritePtr << "\n"
891 << "Store Offs = " << StoreOffset << "\n"
892 << "Load Ptr = " << *LoadPtr << "\n";
898 // If the Load isn't completely contained within the stored bits, we don't
899 // have all the bits to feed it. We could do something crazy in the future
900 // (issue a smaller load then merge the bits in) but this seems unlikely to be
902 if (StoreOffset > LoadOffset ||
903 StoreOffset+StoreSize < LoadOffset+LoadSize)
906 // Okay, we can do this transformation. Return the number of bytes into the
907 // store that the load is.
908 return LoadOffset-StoreOffset;
911 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
912 /// memdep query of a load that ends up being a clobbering store.
913 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
915 const TargetData &TD) {
916 // Cannot handle reading from store of first-class aggregate yet.
917 if (DepSI->getValueOperand()->getType()->isStructTy() ||
918 DepSI->getValueOperand()->getType()->isArrayTy())
921 Value *StorePtr = DepSI->getPointerOperand();
922 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
923 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
924 StorePtr, StoreSize, TD);
927 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
928 /// memdep query of a load that ends up being clobbered by another load. See if
929 /// the other load can feed into the second load.
930 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
931 LoadInst *DepLI, const TargetData &TD){
932 // Cannot handle reading from store of first-class aggregate yet.
933 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
936 Value *DepPtr = DepLI->getPointerOperand();
937 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType());
938 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD);
939 if (R != -1) return R;
941 // If we have a load/load clobber an DepLI can be widened to cover this load,
942 // then we should widen it!
943 int64_t LoadOffs = 0;
944 const Value *LoadBase =
945 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD);
946 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
948 unsigned Size = MemoryDependenceAnalysis::
949 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD);
950 if (Size == 0) return -1;
952 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD);
957 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
959 const TargetData &TD) {
960 // If the mem operation is a non-constant size, we can't handle it.
961 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
962 if (SizeCst == 0) return -1;
963 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
965 // If this is memset, we just need to see if the offset is valid in the size
967 if (MI->getIntrinsicID() == Intrinsic::memset)
968 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
971 // If we have a memcpy/memmove, the only case we can handle is if this is a
972 // copy from constant memory. In that case, we can read directly from the
974 MemTransferInst *MTI = cast<MemTransferInst>(MI);
976 Constant *Src = dyn_cast<Constant>(MTI->getSource());
977 if (Src == 0) return -1;
979 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
980 if (GV == 0 || !GV->isConstant()) return -1;
982 // See if the access is within the bounds of the transfer.
983 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
984 MI->getDest(), MemSizeInBits, TD);
988 // Otherwise, see if we can constant fold a load from the constant with the
989 // offset applied as appropriate.
990 Src = ConstantExpr::getBitCast(Src,
991 llvm::Type::getInt8PtrTy(Src->getContext()));
992 Constant *OffsetCst =
993 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
994 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
995 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
996 if (ConstantFoldLoadFromConstPtr(Src, &TD))
1002 /// GetStoreValueForLoad - This function is called when we have a
1003 /// memdep query of a load that ends up being a clobbering store. This means
1004 /// that the store provides bits used by the load but we the pointers don't
1005 /// mustalias. Check this case to see if there is anything more we can do
1006 /// before we give up.
1007 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1009 Instruction *InsertPt, const TargetData &TD){
1010 LLVMContext &Ctx = SrcVal->getType()->getContext();
1012 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1013 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
1015 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1017 // Compute which bits of the stored value are being used by the load. Convert
1018 // to an integer type to start with.
1019 if (SrcVal->getType()->isPointerTy())
1020 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx));
1021 if (!SrcVal->getType()->isIntegerTy())
1022 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1024 // Shift the bits to the least significant depending on endianness.
1026 if (TD.isLittleEndian())
1027 ShiftAmt = Offset*8;
1029 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1032 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1034 if (LoadSize != StoreSize)
1035 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1037 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1040 /// GetLoadValueForLoad - This function is called when we have a
1041 /// memdep query of a load that ends up being a clobbering load. This means
1042 /// that the load *may* provide bits used by the load but we can't be sure
1043 /// because the pointers don't mustalias. Check this case to see if there is
1044 /// anything more we can do before we give up.
1045 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1046 Type *LoadTy, Instruction *InsertPt,
1048 const TargetData &TD = *gvn.getTargetData();
1049 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1050 // widen SrcVal out to a larger load.
1051 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType());
1052 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
1053 if (Offset+LoadSize > SrcValSize) {
1054 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1055 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1056 // If we have a load/load clobber an DepLI can be widened to cover this
1057 // load, then we should widen it to the next power of 2 size big enough!
1058 unsigned NewLoadSize = Offset+LoadSize;
1059 if (!isPowerOf2_32(NewLoadSize))
1060 NewLoadSize = NextPowerOf2(NewLoadSize);
1062 Value *PtrVal = SrcVal->getPointerOperand();
1064 // Insert the new load after the old load. This ensures that subsequent
1065 // memdep queries will find the new load. We can't easily remove the old
1066 // load completely because it is already in the value numbering table.
1067 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1069 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1070 DestPTy = PointerType::get(DestPTy,
1071 cast<PointerType>(PtrVal->getType())->getAddressSpace());
1072 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1073 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1074 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1075 NewLoad->takeName(SrcVal);
1076 NewLoad->setAlignment(SrcVal->getAlignment());
1078 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1079 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1081 // Replace uses of the original load with the wider load. On a big endian
1082 // system, we need to shift down to get the relevant bits.
1083 Value *RV = NewLoad;
1084 if (TD.isBigEndian())
1085 RV = Builder.CreateLShr(RV,
1086 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1087 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1088 SrcVal->replaceAllUsesWith(RV);
1090 // We would like to use gvn.markInstructionForDeletion here, but we can't
1091 // because the load is already memoized into the leader map table that GVN
1092 // tracks. It is potentially possible to remove the load from the table,
1093 // but then there all of the operations based on it would need to be
1094 // rehashed. Just leave the dead load around.
1095 gvn.getMemDep().removeInstruction(SrcVal);
1099 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD);
1103 /// GetMemInstValueForLoad - This function is called when we have a
1104 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1105 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1106 Type *LoadTy, Instruction *InsertPt,
1107 const TargetData &TD){
1108 LLVMContext &Ctx = LoadTy->getContext();
1109 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1111 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1113 // We know that this method is only called when the mem transfer fully
1114 // provides the bits for the load.
1115 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1116 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1117 // independently of what the offset is.
1118 Value *Val = MSI->getValue();
1120 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1122 Value *OneElt = Val;
1124 // Splat the value out to the right number of bits.
1125 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1126 // If we can double the number of bytes set, do it.
1127 if (NumBytesSet*2 <= LoadSize) {
1128 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1129 Val = Builder.CreateOr(Val, ShVal);
1134 // Otherwise insert one byte at a time.
1135 Value *ShVal = Builder.CreateShl(Val, 1*8);
1136 Val = Builder.CreateOr(OneElt, ShVal);
1140 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1143 // Otherwise, this is a memcpy/memmove from a constant global.
1144 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1145 Constant *Src = cast<Constant>(MTI->getSource());
1147 // Otherwise, see if we can constant fold a load from the constant with the
1148 // offset applied as appropriate.
1149 Src = ConstantExpr::getBitCast(Src,
1150 llvm::Type::getInt8PtrTy(Src->getContext()));
1151 Constant *OffsetCst =
1152 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1153 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1154 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1155 return ConstantFoldLoadFromConstPtr(Src, &TD);
1160 struct AvailableValueInBlock {
1161 /// BB - The basic block in question.
1164 SimpleVal, // A simple offsetted value that is accessed.
1165 LoadVal, // A value produced by a load.
1166 MemIntrin // A memory intrinsic which is loaded from.
1169 /// V - The value that is live out of the block.
1170 PointerIntPair<Value *, 2, ValType> Val;
1172 /// Offset - The byte offset in Val that is interesting for the load query.
1175 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1176 unsigned Offset = 0) {
1177 AvailableValueInBlock Res;
1179 Res.Val.setPointer(V);
1180 Res.Val.setInt(SimpleVal);
1181 Res.Offset = Offset;
1185 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1186 unsigned Offset = 0) {
1187 AvailableValueInBlock Res;
1189 Res.Val.setPointer(MI);
1190 Res.Val.setInt(MemIntrin);
1191 Res.Offset = Offset;
1195 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
1196 unsigned Offset = 0) {
1197 AvailableValueInBlock Res;
1199 Res.Val.setPointer(LI);
1200 Res.Val.setInt(LoadVal);
1201 Res.Offset = Offset;
1205 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
1206 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
1207 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
1209 Value *getSimpleValue() const {
1210 assert(isSimpleValue() && "Wrong accessor");
1211 return Val.getPointer();
1214 LoadInst *getCoercedLoadValue() const {
1215 assert(isCoercedLoadValue() && "Wrong accessor");
1216 return cast<LoadInst>(Val.getPointer());
1219 MemIntrinsic *getMemIntrinValue() const {
1220 assert(isMemIntrinValue() && "Wrong accessor");
1221 return cast<MemIntrinsic>(Val.getPointer());
1224 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
1225 /// defined here to the specified type. This handles various coercion cases.
1226 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1228 if (isSimpleValue()) {
1229 Res = getSimpleValue();
1230 if (Res->getType() != LoadTy) {
1231 const TargetData *TD = gvn.getTargetData();
1232 assert(TD && "Need target data to handle type mismatch case");
1233 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1236 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1237 << *getSimpleValue() << '\n'
1238 << *Res << '\n' << "\n\n\n");
1240 } else if (isCoercedLoadValue()) {
1241 LoadInst *Load = getCoercedLoadValue();
1242 if (Load->getType() == LoadTy && Offset == 0) {
1245 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1248 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1249 << *getCoercedLoadValue() << '\n'
1250 << *Res << '\n' << "\n\n\n");
1253 const TargetData *TD = gvn.getTargetData();
1254 assert(TD && "Need target data to handle type mismatch case");
1255 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1256 LoadTy, BB->getTerminator(), *TD);
1257 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1258 << " " << *getMemIntrinValue() << '\n'
1259 << *Res << '\n' << "\n\n\n");
1265 } // end anonymous namespace
1267 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1268 /// construct SSA form, allowing us to eliminate LI. This returns the value
1269 /// that should be used at LI's definition site.
1270 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1271 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1273 // Check for the fully redundant, dominating load case. In this case, we can
1274 // just use the dominating value directly.
1275 if (ValuesPerBlock.size() == 1 &&
1276 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1278 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()->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 static bool isLifetimeStart(const Instruction *Inst) {
1323 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1324 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1328 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1329 /// non-local by performing PHI construction.
1330 bool GVN::processNonLocalLoad(LoadInst *LI) {
1331 // Find the non-local dependencies of the load.
1332 SmallVector<NonLocalDepResult, 64> Deps;
1333 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1334 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1335 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
1336 // << Deps.size() << *LI << '\n');
1338 // If we had to process more than one hundred blocks to find the
1339 // dependencies, this load isn't worth worrying about. Optimizing
1340 // it will be too expensive.
1341 unsigned NumDeps = Deps.size();
1345 // If we had a phi translation failure, we'll have a single entry which is a
1346 // clobber in the current block. Reject this early.
1348 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1350 dbgs() << "GVN: non-local load ";
1351 WriteAsOperand(dbgs(), LI);
1352 dbgs() << " has unknown dependencies\n";
1357 // Filter out useless results (non-locals, etc). Keep track of the blocks
1358 // where we have a value available in repl, also keep track of whether we see
1359 // dependencies that produce an unknown value for the load (such as a call
1360 // that could potentially clobber the load).
1361 SmallVector<AvailableValueInBlock, 64> ValuesPerBlock;
1362 SmallVector<BasicBlock*, 64> UnavailableBlocks;
1364 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1365 BasicBlock *DepBB = Deps[i].getBB();
1366 MemDepResult DepInfo = Deps[i].getResult();
1368 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1369 UnavailableBlocks.push_back(DepBB);
1373 if (DepInfo.isClobber()) {
1374 // The address being loaded in this non-local block may not be the same as
1375 // the pointer operand of the load if PHI translation occurs. Make sure
1376 // to consider the right address.
1377 Value *Address = Deps[i].getAddress();
1379 // If the dependence is to a store that writes to a superset of the bits
1380 // read by the load, we can extract the bits we need for the load from the
1382 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1383 if (TD && Address) {
1384 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1387 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1388 DepSI->getValueOperand(),
1395 // Check to see if we have something like this:
1398 // if we have this, replace the later with an extraction from the former.
1399 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1400 // If this is a clobber and L is the first instruction in its block, then
1401 // we have the first instruction in the entry block.
1402 if (DepLI != LI && Address && TD) {
1403 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
1404 LI->getPointerOperand(),
1408 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1415 // If the clobbering value is a memset/memcpy/memmove, see if we can
1416 // forward a value on from it.
1417 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1418 if (TD && Address) {
1419 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1422 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1429 UnavailableBlocks.push_back(DepBB);
1433 // DepInfo.isDef() here
1435 Instruction *DepInst = DepInfo.getInst();
1437 // Loading the allocation -> undef.
1438 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1439 // Loading immediately after lifetime begin -> undef.
1440 isLifetimeStart(DepInst)) {
1441 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1442 UndefValue::get(LI->getType())));
1446 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1447 // Reject loads and stores that are to the same address but are of
1448 // different types if we have to.
1449 if (S->getValueOperand()->getType() != LI->getType()) {
1450 // If the stored value is larger or equal to the loaded value, we can
1452 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1453 LI->getType(), *TD)) {
1454 UnavailableBlocks.push_back(DepBB);
1459 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1460 S->getValueOperand()));
1464 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1465 // If the types mismatch and we can't handle it, reject reuse of the load.
1466 if (LD->getType() != LI->getType()) {
1467 // If the stored value is larger or equal to the loaded value, we can
1469 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1470 UnavailableBlocks.push_back(DepBB);
1474 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1478 UnavailableBlocks.push_back(DepBB);
1482 // If we have no predecessors that produce a known value for this load, exit
1484 if (ValuesPerBlock.empty()) return false;
1486 // If all of the instructions we depend on produce a known value for this
1487 // load, then it is fully redundant and we can use PHI insertion to compute
1488 // its value. Insert PHIs and remove the fully redundant value now.
1489 if (UnavailableBlocks.empty()) {
1490 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1492 // Perform PHI construction.
1493 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1494 LI->replaceAllUsesWith(V);
1496 if (isa<PHINode>(V))
1498 if (V->getType()->isPointerTy())
1499 MD->invalidateCachedPointerInfo(V);
1500 markInstructionForDeletion(LI);
1505 if (!EnablePRE || !EnableLoadPRE)
1508 // Okay, we have *some* definitions of the value. This means that the value
1509 // is available in some of our (transitive) predecessors. Lets think about
1510 // doing PRE of this load. This will involve inserting a new load into the
1511 // predecessor when it's not available. We could do this in general, but
1512 // prefer to not increase code size. As such, we only do this when we know
1513 // that we only have to insert *one* load (which means we're basically moving
1514 // the load, not inserting a new one).
1516 SmallPtrSet<BasicBlock *, 4> Blockers;
1517 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1518 Blockers.insert(UnavailableBlocks[i]);
1520 // Let's find the first basic block with more than one predecessor. Walk
1521 // backwards through predecessors if needed.
1522 BasicBlock *LoadBB = LI->getParent();
1523 BasicBlock *TmpBB = LoadBB;
1525 bool isSinglePred = false;
1526 bool allSingleSucc = true;
1527 while (TmpBB->getSinglePredecessor()) {
1528 isSinglePred = true;
1529 TmpBB = TmpBB->getSinglePredecessor();
1530 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1532 if (Blockers.count(TmpBB))
1535 // If any of these blocks has more than one successor (i.e. if the edge we
1536 // just traversed was critical), then there are other paths through this
1537 // block along which the load may not be anticipated. Hoisting the load
1538 // above this block would be adding the load to execution paths along
1539 // which it was not previously executed.
1540 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1547 // FIXME: It is extremely unclear what this loop is doing, other than
1548 // artificially restricting loadpre.
1551 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1552 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1553 if (AV.isSimpleValue())
1554 // "Hot" Instruction is in some loop (because it dominates its dep.
1556 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1557 if (DT->dominates(LI, I)) {
1563 // We are interested only in "hot" instructions. We don't want to do any
1564 // mis-optimizations here.
1569 // Check to see how many predecessors have the loaded value fully
1571 DenseMap<BasicBlock*, Value*> PredLoads;
1572 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1573 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1574 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1575 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1576 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1578 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
1579 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1581 BasicBlock *Pred = *PI;
1582 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1585 PredLoads[Pred] = 0;
1587 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1588 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1589 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1590 << Pred->getName() << "': " << *LI << '\n');
1594 if (LoadBB->isLandingPad()) {
1596 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1597 << Pred->getName() << "': " << *LI << '\n');
1601 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
1602 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
1606 if (!NeedToSplit.empty()) {
1607 toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
1611 // Decide whether PRE is profitable for this load.
1612 unsigned NumUnavailablePreds = PredLoads.size();
1613 assert(NumUnavailablePreds != 0 &&
1614 "Fully available value should be eliminated above!");
1616 // If this load is unavailable in multiple predecessors, reject it.
1617 // FIXME: If we could restructure the CFG, we could make a common pred with
1618 // all the preds that don't have an available LI and insert a new load into
1620 if (NumUnavailablePreds != 1)
1623 // Check if the load can safely be moved to all the unavailable predecessors.
1624 bool CanDoPRE = true;
1625 SmallVector<Instruction*, 8> NewInsts;
1626 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1627 E = PredLoads.end(); I != E; ++I) {
1628 BasicBlock *UnavailablePred = I->first;
1630 // Do PHI translation to get its value in the predecessor if necessary. The
1631 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1633 // If all preds have a single successor, then we know it is safe to insert
1634 // the load on the pred (?!?), so we can insert code to materialize the
1635 // pointer if it is not available.
1636 PHITransAddr Address(LI->getPointerOperand(), TD);
1638 if (allSingleSucc) {
1639 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1642 Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
1643 LoadPtr = Address.getAddr();
1646 // If we couldn't find or insert a computation of this phi translated value,
1649 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1650 << *LI->getPointerOperand() << "\n");
1655 // Make sure it is valid to move this load here. We have to watch out for:
1656 // @1 = getelementptr (i8* p, ...
1657 // test p and branch if == 0
1659 // It is valid to have the getelementptr before the test, even if p can
1660 // be 0, as getelementptr only does address arithmetic.
1661 // If we are not pushing the value through any multiple-successor blocks
1662 // we do not have this case. Otherwise, check that the load is safe to
1663 // put anywhere; this can be improved, but should be conservatively safe.
1664 if (!allSingleSucc &&
1665 // FIXME: REEVALUTE THIS.
1666 !isSafeToLoadUnconditionally(LoadPtr,
1667 UnavailablePred->getTerminator(),
1668 LI->getAlignment(), TD)) {
1673 I->second = LoadPtr;
1677 while (!NewInsts.empty()) {
1678 Instruction *I = NewInsts.pop_back_val();
1679 if (MD) MD->removeInstruction(I);
1680 I->eraseFromParent();
1685 // Okay, we can eliminate this load by inserting a reload in the predecessor
1686 // and using PHI construction to get the value in the other predecessors, do
1688 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1689 DEBUG(if (!NewInsts.empty())
1690 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1691 << *NewInsts.back() << '\n');
1693 // Assign value numbers to the new instructions.
1694 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1695 // FIXME: We really _ought_ to insert these value numbers into their
1696 // parent's availability map. However, in doing so, we risk getting into
1697 // ordering issues. If a block hasn't been processed yet, we would be
1698 // marking a value as AVAIL-IN, which isn't what we intend.
1699 VN.lookup_or_add(NewInsts[i]);
1702 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1703 E = PredLoads.end(); I != E; ++I) {
1704 BasicBlock *UnavailablePred = I->first;
1705 Value *LoadPtr = I->second;
1707 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1709 UnavailablePred->getTerminator());
1711 // Transfer the old load's TBAA tag to the new load.
1712 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1713 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1715 // Transfer DebugLoc.
1716 NewLoad->setDebugLoc(LI->getDebugLoc());
1718 // Add the newly created load.
1719 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1721 MD->invalidateCachedPointerInfo(LoadPtr);
1722 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1725 // Perform PHI construction.
1726 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1727 LI->replaceAllUsesWith(V);
1728 if (isa<PHINode>(V))
1730 if (V->getType()->isPointerTy())
1731 MD->invalidateCachedPointerInfo(V);
1732 markInstructionForDeletion(LI);
1737 /// processLoad - Attempt to eliminate a load, first by eliminating it
1738 /// locally, and then attempting non-local elimination if that fails.
1739 bool GVN::processLoad(LoadInst *L) {
1746 if (L->use_empty()) {
1747 markInstructionForDeletion(L);
1751 // ... to a pointer that has been loaded from before...
1752 MemDepResult Dep = MD->getDependency(L);
1754 // If we have a clobber and target data is around, see if this is a clobber
1755 // that we can fix up through code synthesis.
1756 if (Dep.isClobber() && TD) {
1757 // Check to see if we have something like this:
1758 // store i32 123, i32* %P
1759 // %A = bitcast i32* %P to i8*
1760 // %B = gep i8* %A, i32 1
1763 // We could do that by recognizing if the clobber instructions are obviously
1764 // a common base + constant offset, and if the previous store (or memset)
1765 // completely covers this load. This sort of thing can happen in bitfield
1767 Value *AvailVal = 0;
1768 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1769 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1770 L->getPointerOperand(),
1773 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1774 L->getType(), L, *TD);
1777 // Check to see if we have something like this:
1780 // if we have this, replace the later with an extraction from the former.
1781 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1782 // If this is a clobber and L is the first instruction in its block, then
1783 // we have the first instruction in the entry block.
1787 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1788 L->getPointerOperand(),
1791 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1794 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1795 // a value on from it.
1796 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1797 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1798 L->getPointerOperand(),
1801 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
1805 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1806 << *AvailVal << '\n' << *L << "\n\n\n");
1808 // Replace the load!
1809 L->replaceAllUsesWith(AvailVal);
1810 if (AvailVal->getType()->isPointerTy())
1811 MD->invalidateCachedPointerInfo(AvailVal);
1812 markInstructionForDeletion(L);
1818 // If the value isn't available, don't do anything!
1819 if (Dep.isClobber()) {
1821 // fast print dep, using operator<< on instruction is too slow.
1822 dbgs() << "GVN: load ";
1823 WriteAsOperand(dbgs(), L);
1824 Instruction *I = Dep.getInst();
1825 dbgs() << " is clobbered by " << *I << '\n';
1830 // If it is defined in another block, try harder.
1831 if (Dep.isNonLocal())
1832 return processNonLocalLoad(L);
1836 // fast print dep, using operator<< on instruction is too slow.
1837 dbgs() << "GVN: load ";
1838 WriteAsOperand(dbgs(), L);
1839 dbgs() << " has unknown dependence\n";
1844 Instruction *DepInst = Dep.getInst();
1845 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1846 Value *StoredVal = DepSI->getValueOperand();
1848 // The store and load are to a must-aliased pointer, but they may not
1849 // actually have the same type. See if we know how to reuse the stored
1850 // value (depending on its type).
1851 if (StoredVal->getType() != L->getType()) {
1853 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1858 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1859 << '\n' << *L << "\n\n\n");
1866 L->replaceAllUsesWith(StoredVal);
1867 if (StoredVal->getType()->isPointerTy())
1868 MD->invalidateCachedPointerInfo(StoredVal);
1869 markInstructionForDeletion(L);
1874 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1875 Value *AvailableVal = DepLI;
1877 // The loads are of a must-aliased pointer, but they may not actually have
1878 // the same type. See if we know how to reuse the previously loaded value
1879 // (depending on its type).
1880 if (DepLI->getType() != L->getType()) {
1882 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1884 if (AvailableVal == 0)
1887 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1888 << "\n" << *L << "\n\n\n");
1895 L->replaceAllUsesWith(AvailableVal);
1896 if (DepLI->getType()->isPointerTy())
1897 MD->invalidateCachedPointerInfo(DepLI);
1898 markInstructionForDeletion(L);
1903 // If this load really doesn't depend on anything, then we must be loading an
1904 // undef value. This can happen when loading for a fresh allocation with no
1905 // intervening stores, for example.
1906 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1907 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1908 markInstructionForDeletion(L);
1913 // If this load occurs either right after a lifetime begin,
1914 // then the loaded value is undefined.
1915 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1916 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1917 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1918 markInstructionForDeletion(L);
1927 // findLeader - In order to find a leader for a given value number at a
1928 // specific basic block, we first obtain the list of all Values for that number,
1929 // and then scan the list to find one whose block dominates the block in
1930 // question. This is fast because dominator tree queries consist of only
1931 // a few comparisons of DFS numbers.
1932 Value *GVN::findLeader(BasicBlock *BB, uint32_t num) {
1933 LeaderTableEntry Vals = LeaderTable[num];
1934 if (!Vals.Val) return 0;
1937 if (DT->dominates(Vals.BB, BB)) {
1939 if (isa<Constant>(Val)) return Val;
1942 LeaderTableEntry* Next = Vals.Next;
1944 if (DT->dominates(Next->BB, BB)) {
1945 if (isa<Constant>(Next->Val)) return Next->Val;
1946 if (!Val) Val = Next->Val;
1955 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
1956 /// use is dominated by the given basic block. Returns the number of uses that
1958 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
1961 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1963 Use &U = (UI++).getUse();
1965 // If From occurs as a phi node operand then the use implicitly lives in the
1966 // corresponding incoming block. Otherwise it is the block containing the
1967 // user that must be dominated by Root.
1968 BasicBlock *UsingBlock;
1969 if (PHINode *PN = dyn_cast<PHINode>(U.getUser()))
1970 UsingBlock = PN->getIncomingBlock(U);
1972 UsingBlock = cast<Instruction>(U.getUser())->getParent();
1974 if (DT->dominates(Root, UsingBlock)) {
1982 /// propagateEquality - The given values are known to be equal in every block
1983 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1984 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1985 bool GVN::propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root) {
1986 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
1987 Worklist.push_back(std::make_pair(LHS, RHS));
1988 bool Changed = false;
1990 while (!Worklist.empty()) {
1991 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1992 LHS = Item.first; RHS = Item.second;
1994 if (LHS == RHS) continue;
1995 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1997 // Don't try to propagate equalities between constants.
1998 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2000 // Prefer a constant on the right-hand side, or an Argument if no constants.
2001 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2002 std::swap(LHS, RHS);
2003 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2005 // If there is no obvious reason to prefer the left-hand side over the right-
2006 // hand side, ensure the longest lived term is on the right-hand side, so the
2007 // shortest lived term will be replaced by the longest lived. This tends to
2008 // expose more simplifications.
2009 uint32_t LVN = VN.lookup_or_add(LHS);
2010 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2011 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2012 // Move the 'oldest' value to the right-hand side, using the value number as
2014 uint32_t RVN = VN.lookup_or_add(RHS);
2016 std::swap(LHS, RHS);
2020 assert((!isa<Instruction>(RHS) ||
2021 DT->properlyDominates(cast<Instruction>(RHS)->getParent(), Root)) &&
2022 "Instruction doesn't dominate scope!");
2024 // If value numbering later sees that an instruction in the scope is equal
2025 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2026 // the invariant that instructions only occur in the leader table for their
2027 // own value number (this is used by removeFromLeaderTable), do not do this
2028 // if RHS is an instruction (if an instruction in the scope is morphed into
2029 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2030 // using the leader table is about compiling faster, not optimizing better).
2031 if (!isa<Instruction>(RHS))
2032 addToLeaderTable(LVN, RHS, Root);
2034 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2035 // LHS always has at least one use that is not dominated by Root, this will
2036 // never do anything if LHS has only one use.
2037 if (!LHS->hasOneUse()) {
2038 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2039 Changed |= NumReplacements > 0;
2040 NumGVNEqProp += NumReplacements;
2043 // Now try to deduce additional equalities from this one. For example, if the
2044 // known equality was "(A != B)" == "false" then it follows that A and B are
2045 // equal in the scope. Only boolean equalities with an explicit true or false
2046 // RHS are currently supported.
2047 if (!RHS->getType()->isIntegerTy(1))
2048 // Not a boolean equality - bail out.
2050 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2052 // RHS neither 'true' nor 'false' - bail out.
2054 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2055 bool isKnownTrue = CI->isAllOnesValue();
2056 bool isKnownFalse = !isKnownTrue;
2058 // If "A && B" is known true then both A and B are known true. If "A || B"
2059 // is known false then both A and B are known false.
2061 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2062 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2063 Worklist.push_back(std::make_pair(A, RHS));
2064 Worklist.push_back(std::make_pair(B, RHS));
2068 // If we are propagating an equality like "(A == B)" == "true" then also
2069 // propagate the equality A == B. When propagating a comparison such as
2070 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2071 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
2072 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2074 // If "A == B" is known true, or "A != B" is known false, then replace
2075 // A with B everywhere in the scope.
2076 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2077 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2078 Worklist.push_back(std::make_pair(Op0, Op1));
2080 // If "A >= B" is known true, replace "A < B" with false everywhere.
2081 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2082 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2083 // Since we don't have the instruction "A < B" immediately to hand, work out
2084 // the value number that it would have and use that to find an appropriate
2085 // instruction (if any).
2086 uint32_t NextNum = VN.getNextUnusedValueNumber();
2087 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2088 // If the number we were assigned was brand new then there is no point in
2089 // looking for an instruction realizing it: there cannot be one!
2090 if (Num < NextNum) {
2091 Value *NotCmp = findLeader(Root, Num);
2092 if (NotCmp && isa<Instruction>(NotCmp)) {
2093 unsigned NumReplacements =
2094 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2095 Changed |= NumReplacements > 0;
2096 NumGVNEqProp += NumReplacements;
2099 // Ensure that any instruction in scope that gets the "A < B" value number
2100 // is replaced with false.
2101 addToLeaderTable(Num, NotVal, Root);
2110 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
2111 /// true if every path from the entry block to 'Dst' passes via this edge. In
2112 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2113 static bool isOnlyReachableViaThisEdge(BasicBlock *Src, BasicBlock *Dst,
2114 DominatorTree *DT) {
2115 // While in theory it is interesting to consider the case in which Dst has
2116 // more than one predecessor, because Dst might be part of a loop which is
2117 // only reachable from Src, in practice it is pointless since at the time
2118 // GVN runs all such loops have preheaders, which means that Dst will have
2119 // been changed to have only one predecessor, namely Src.
2120 BasicBlock *Pred = Dst->getSinglePredecessor();
2121 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2126 /// processInstruction - When calculating availability, handle an instruction
2127 /// by inserting it into the appropriate sets
2128 bool GVN::processInstruction(Instruction *I) {
2129 // Ignore dbg info intrinsics.
2130 if (isa<DbgInfoIntrinsic>(I))
2133 // If the instruction can be easily simplified then do so now in preference
2134 // to value numbering it. Value numbering often exposes redundancies, for
2135 // example if it determines that %y is equal to %x then the instruction
2136 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2137 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) {
2138 I->replaceAllUsesWith(V);
2139 if (MD && V->getType()->isPointerTy())
2140 MD->invalidateCachedPointerInfo(V);
2141 markInstructionForDeletion(I);
2146 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2147 if (processLoad(LI))
2150 unsigned Num = VN.lookup_or_add(LI);
2151 addToLeaderTable(Num, LI, LI->getParent());
2155 // For conditional branches, we can perform simple conditional propagation on
2156 // the condition value itself.
2157 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2158 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
2161 Value *BranchCond = BI->getCondition();
2163 BasicBlock *TrueSucc = BI->getSuccessor(0);
2164 BasicBlock *FalseSucc = BI->getSuccessor(1);
2165 BasicBlock *Parent = BI->getParent();
2166 bool Changed = false;
2168 if (isOnlyReachableViaThisEdge(Parent, TrueSucc, DT))
2169 Changed |= propagateEquality(BranchCond,
2170 ConstantInt::getTrue(TrueSucc->getContext()),
2173 if (isOnlyReachableViaThisEdge(Parent, FalseSucc, DT))
2174 Changed |= propagateEquality(BranchCond,
2175 ConstantInt::getFalse(FalseSucc->getContext()),
2181 // For switches, propagate the case values into the case destinations.
2182 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2183 Value *SwitchCond = SI->getCondition();
2184 BasicBlock *Parent = SI->getParent();
2185 bool Changed = false;
2186 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2188 BasicBlock *Dst = i.getCaseSuccessor();
2189 if (isOnlyReachableViaThisEdge(Parent, Dst, DT))
2190 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), Dst);
2195 // Instructions with void type don't return a value, so there's
2196 // no point in trying to find redundancies in them.
2197 if (I->getType()->isVoidTy()) return false;
2199 uint32_t NextNum = VN.getNextUnusedValueNumber();
2200 unsigned Num = VN.lookup_or_add(I);
2202 // Allocations are always uniquely numbered, so we can save time and memory
2203 // by fast failing them.
2204 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2205 addToLeaderTable(Num, I, I->getParent());
2209 // If the number we were assigned was a brand new VN, then we don't
2210 // need to do a lookup to see if the number already exists
2211 // somewhere in the domtree: it can't!
2212 if (Num >= NextNum) {
2213 addToLeaderTable(Num, I, I->getParent());
2217 // Perform fast-path value-number based elimination of values inherited from
2219 Value *repl = findLeader(I->getParent(), Num);
2221 // Failure, just remember this instance for future use.
2222 addToLeaderTable(Num, I, I->getParent());
2227 I->replaceAllUsesWith(repl);
2228 if (MD && repl->getType()->isPointerTy())
2229 MD->invalidateCachedPointerInfo(repl);
2230 markInstructionForDeletion(I);
2234 /// runOnFunction - This is the main transformation entry point for a function.
2235 bool GVN::runOnFunction(Function& F) {
2237 MD = &getAnalysis<MemoryDependenceAnalysis>();
2238 DT = &getAnalysis<DominatorTree>();
2239 TD = getAnalysisIfAvailable<TargetData>();
2240 TLI = &getAnalysis<TargetLibraryInfo>();
2241 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2245 bool Changed = false;
2246 bool ShouldContinue = true;
2248 // Merge unconditional branches, allowing PRE to catch more
2249 // optimization opportunities.
2250 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2251 BasicBlock *BB = FI++;
2253 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2254 if (removedBlock) ++NumGVNBlocks;
2256 Changed |= removedBlock;
2259 unsigned Iteration = 0;
2260 while (ShouldContinue) {
2261 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2262 ShouldContinue = iterateOnFunction(F);
2263 if (splitCriticalEdges())
2264 ShouldContinue = true;
2265 Changed |= ShouldContinue;
2270 bool PREChanged = true;
2271 while (PREChanged) {
2272 PREChanged = performPRE(F);
2273 Changed |= PREChanged;
2276 // FIXME: Should perform GVN again after PRE does something. PRE can move
2277 // computations into blocks where they become fully redundant. Note that
2278 // we can't do this until PRE's critical edge splitting updates memdep.
2279 // Actually, when this happens, we should just fully integrate PRE into GVN.
2281 cleanupGlobalSets();
2287 bool GVN::processBlock(BasicBlock *BB) {
2288 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2289 // (and incrementing BI before processing an instruction).
2290 assert(InstrsToErase.empty() &&
2291 "We expect InstrsToErase to be empty across iterations");
2292 bool ChangedFunction = false;
2294 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2296 ChangedFunction |= processInstruction(BI);
2297 if (InstrsToErase.empty()) {
2302 // If we need some instructions deleted, do it now.
2303 NumGVNInstr += InstrsToErase.size();
2305 // Avoid iterator invalidation.
2306 bool AtStart = BI == BB->begin();
2310 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(),
2311 E = InstrsToErase.end(); I != E; ++I) {
2312 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2313 if (MD) MD->removeInstruction(*I);
2314 (*I)->eraseFromParent();
2315 DEBUG(verifyRemoved(*I));
2317 InstrsToErase.clear();
2325 return ChangedFunction;
2328 /// performPRE - Perform a purely local form of PRE that looks for diamond
2329 /// control flow patterns and attempts to perform simple PRE at the join point.
2330 bool GVN::performPRE(Function &F) {
2331 bool Changed = false;
2332 DenseMap<BasicBlock*, Value*> predMap;
2333 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2334 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2335 BasicBlock *CurrentBlock = *DI;
2337 // Nothing to PRE in the entry block.
2338 if (CurrentBlock == &F.getEntryBlock()) continue;
2340 // Don't perform PRE on a landing pad.
2341 if (CurrentBlock->isLandingPad()) continue;
2343 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2344 BE = CurrentBlock->end(); BI != BE; ) {
2345 Instruction *CurInst = BI++;
2347 if (isa<AllocaInst>(CurInst) ||
2348 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2349 CurInst->getType()->isVoidTy() ||
2350 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2351 isa<DbgInfoIntrinsic>(CurInst))
2354 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2355 // sinking the compare again, and it would force the code generator to
2356 // move the i1 from processor flags or predicate registers into a general
2357 // purpose register.
2358 if (isa<CmpInst>(CurInst))
2361 // We don't currently value number ANY inline asm calls.
2362 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2363 if (CallI->isInlineAsm())
2366 uint32_t ValNo = VN.lookup(CurInst);
2368 // Look for the predecessors for PRE opportunities. We're
2369 // only trying to solve the basic diamond case, where
2370 // a value is computed in the successor and one predecessor,
2371 // but not the other. We also explicitly disallow cases
2372 // where the successor is its own predecessor, because they're
2373 // more complicated to get right.
2374 unsigned NumWith = 0;
2375 unsigned NumWithout = 0;
2376 BasicBlock *PREPred = 0;
2379 for (pred_iterator PI = pred_begin(CurrentBlock),
2380 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2381 BasicBlock *P = *PI;
2382 // We're not interested in PRE where the block is its
2383 // own predecessor, or in blocks with predecessors
2384 // that are not reachable.
2385 if (P == CurrentBlock) {
2388 } else if (!DT->dominates(&F.getEntryBlock(), P)) {
2393 Value* predV = findLeader(P, ValNo);
2397 } else if (predV == CurInst) {
2405 // Don't do PRE when it might increase code size, i.e. when
2406 // we would need to insert instructions in more than one pred.
2407 if (NumWithout != 1 || NumWith == 0)
2410 // Don't do PRE across indirect branch.
2411 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2414 // We can't do PRE safely on a critical edge, so instead we schedule
2415 // the edge to be split and perform the PRE the next time we iterate
2417 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2418 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2419 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2423 // Instantiate the expression in the predecessor that lacked it.
2424 // Because we are going top-down through the block, all value numbers
2425 // will be available in the predecessor by the time we need them. Any
2426 // that weren't originally present will have been instantiated earlier
2428 Instruction *PREInstr = CurInst->clone();
2429 bool success = true;
2430 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2431 Value *Op = PREInstr->getOperand(i);
2432 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2435 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2436 PREInstr->setOperand(i, V);
2443 // Fail out if we encounter an operand that is not available in
2444 // the PRE predecessor. This is typically because of loads which
2445 // are not value numbered precisely.
2448 DEBUG(verifyRemoved(PREInstr));
2452 PREInstr->insertBefore(PREPred->getTerminator());
2453 PREInstr->setName(CurInst->getName() + ".pre");
2454 PREInstr->setDebugLoc(CurInst->getDebugLoc());
2455 predMap[PREPred] = PREInstr;
2456 VN.add(PREInstr, ValNo);
2459 // Update the availability map to include the new instruction.
2460 addToLeaderTable(ValNo, PREInstr, PREPred);
2462 // Create a PHI to make the value available in this block.
2463 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2464 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE),
2465 CurInst->getName() + ".pre-phi",
2466 CurrentBlock->begin());
2467 for (pred_iterator PI = PB; PI != PE; ++PI) {
2468 BasicBlock *P = *PI;
2469 Phi->addIncoming(predMap[P], P);
2473 addToLeaderTable(ValNo, Phi, CurrentBlock);
2474 Phi->setDebugLoc(CurInst->getDebugLoc());
2475 CurInst->replaceAllUsesWith(Phi);
2476 if (Phi->getType()->isPointerTy()) {
2477 // Because we have added a PHI-use of the pointer value, it has now
2478 // "escaped" from alias analysis' perspective. We need to inform
2480 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
2482 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2483 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2487 MD->invalidateCachedPointerInfo(Phi);
2490 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2492 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2493 if (MD) MD->removeInstruction(CurInst);
2494 CurInst->eraseFromParent();
2495 DEBUG(verifyRemoved(CurInst));
2500 if (splitCriticalEdges())
2506 /// splitCriticalEdges - Split critical edges found during the previous
2507 /// iteration that may enable further optimization.
2508 bool GVN::splitCriticalEdges() {
2509 if (toSplit.empty())
2512 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2513 SplitCriticalEdge(Edge.first, Edge.second, this);
2514 } while (!toSplit.empty());
2515 if (MD) MD->invalidateCachedPredecessors();
2519 /// iterateOnFunction - Executes one iteration of GVN
2520 bool GVN::iterateOnFunction(Function &F) {
2521 cleanupGlobalSets();
2523 // Top-down walk of the dominator tree
2524 bool Changed = false;
2526 // Needed for value numbering with phi construction to work.
2527 ReversePostOrderTraversal<Function*> RPOT(&F);
2528 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2529 RE = RPOT.end(); RI != RE; ++RI)
2530 Changed |= processBlock(*RI);
2532 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2533 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2534 Changed |= processBlock(DI->getBlock());
2540 void GVN::cleanupGlobalSets() {
2542 LeaderTable.clear();
2543 TableAllocator.Reset();
2546 /// verifyRemoved - Verify that the specified instruction does not occur in our
2547 /// internal data structures.
2548 void GVN::verifyRemoved(const Instruction *Inst) const {
2549 VN.verifyRemoved(Inst);
2551 // Walk through the value number scope to make sure the instruction isn't
2552 // ferreted away in it.
2553 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2554 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2555 const LeaderTableEntry *Node = &I->second;
2556 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2558 while (Node->Next) {
2560 assert(Node->Val != Inst && "Inst still in value numbering scope!");