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/Transforms/Utils/BasicBlockUtils.h"
35 #include "llvm/Transforms/Utils/SSAUpdater.h"
36 #include "llvm/ADT/DenseMap.h"
37 #include "llvm/ADT/DepthFirstIterator.h"
38 #include "llvm/ADT/SmallPtrSet.h"
39 #include "llvm/ADT/Statistic.h"
40 #include "llvm/Support/Allocator.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/IRBuilder.h"
46 STATISTIC(NumGVNInstr, "Number of instructions deleted");
47 STATISTIC(NumGVNLoad, "Number of loads deleted");
48 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
49 STATISTIC(NumGVNBlocks, "Number of blocks merged");
50 STATISTIC(NumPRELoad, "Number of loads PRE'd");
52 static cl::opt<bool> EnablePRE("enable-pre",
53 cl::init(true), cl::Hidden);
54 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
56 //===----------------------------------------------------------------------===//
58 //===----------------------------------------------------------------------===//
60 /// This class holds the mapping between values and value numbers. It is used
61 /// as an efficient mechanism to determine the expression-wise equivalence of
67 SmallVector<uint32_t, 4> varargs;
69 Expression(uint32_t o = ~2U) : opcode(o) { }
71 bool operator==(const Expression &other) const {
72 if (opcode != other.opcode)
74 if (opcode == ~0U || opcode == ~1U)
76 if (type != other.type)
78 if (varargs != other.varargs)
85 DenseMap<Value*, uint32_t> valueNumbering;
86 DenseMap<Expression, uint32_t> expressionNumbering;
88 MemoryDependenceAnalysis *MD;
91 uint32_t nextValueNumber;
93 Expression create_expression(Instruction* I);
94 uint32_t lookup_or_add_call(CallInst* C);
96 ValueTable() : nextValueNumber(1) { }
97 uint32_t lookup_or_add(Value *V);
98 uint32_t lookup(Value *V) const;
99 void add(Value *V, uint32_t num);
101 void erase(Value *v);
102 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
103 AliasAnalysis *getAliasAnalysis() const { return AA; }
104 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
105 void setDomTree(DominatorTree* D) { DT = D; }
106 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
107 void verifyRemoved(const Value *) const;
112 template <> struct DenseMapInfo<Expression> {
113 static inline Expression getEmptyKey() {
117 static inline Expression getTombstoneKey() {
121 static unsigned getHashValue(const Expression e) {
122 unsigned hash = e.opcode;
124 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
125 (unsigned)((uintptr_t)e.type >> 9));
127 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
128 E = e.varargs.end(); I != E; ++I)
129 hash = *I + hash * 37;
133 static bool isEqual(const Expression &LHS, const Expression &RHS) {
140 //===----------------------------------------------------------------------===//
141 // ValueTable Internal Functions
142 //===----------------------------------------------------------------------===//
145 Expression ValueTable::create_expression(Instruction *I) {
147 e.type = I->getType();
148 e.opcode = I->getOpcode();
149 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
151 e.varargs.push_back(lookup_or_add(*OI));
153 if (CmpInst *C = dyn_cast<CmpInst>(I))
154 e.opcode = (C->getOpcode() << 8) | C->getPredicate();
155 else if (ExtractValueInst *E = dyn_cast<ExtractValueInst>(I)) {
156 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
158 e.varargs.push_back(*II);
159 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
160 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
162 e.varargs.push_back(*II);
168 //===----------------------------------------------------------------------===//
169 // ValueTable External Functions
170 //===----------------------------------------------------------------------===//
172 /// add - Insert a value into the table with a specified value number.
173 void ValueTable::add(Value *V, uint32_t num) {
174 valueNumbering.insert(std::make_pair(V, num));
177 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
178 if (AA->doesNotAccessMemory(C)) {
179 Expression exp = create_expression(C);
180 uint32_t& e = expressionNumbering[exp];
181 if (!e) e = nextValueNumber++;
182 valueNumbering[C] = e;
184 } else if (AA->onlyReadsMemory(C)) {
185 Expression exp = create_expression(C);
186 uint32_t& e = expressionNumbering[exp];
188 e = nextValueNumber++;
189 valueNumbering[C] = e;
193 e = nextValueNumber++;
194 valueNumbering[C] = e;
198 MemDepResult local_dep = MD->getDependency(C);
200 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
201 valueNumbering[C] = nextValueNumber;
202 return nextValueNumber++;
205 if (local_dep.isDef()) {
206 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
208 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
209 valueNumbering[C] = nextValueNumber;
210 return nextValueNumber++;
213 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
214 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
215 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
217 valueNumbering[C] = nextValueNumber;
218 return nextValueNumber++;
222 uint32_t v = lookup_or_add(local_cdep);
223 valueNumbering[C] = v;
228 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
229 MD->getNonLocalCallDependency(CallSite(C));
230 // FIXME: call/call dependencies for readonly calls should return def, not
231 // clobber! Move the checking logic to MemDep!
234 // Check to see if we have a single dominating call instruction that is
236 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
237 const NonLocalDepEntry *I = &deps[i];
238 // Ignore non-local dependencies.
239 if (I->getResult().isNonLocal())
242 // We don't handle non-depedencies. If we already have a call, reject
243 // instruction dependencies.
244 if (I->getResult().isClobber() || cdep != 0) {
249 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
250 // FIXME: All duplicated with non-local case.
251 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
252 cdep = NonLocalDepCall;
261 valueNumbering[C] = nextValueNumber;
262 return nextValueNumber++;
265 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
266 valueNumbering[C] = nextValueNumber;
267 return nextValueNumber++;
269 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
270 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
271 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
273 valueNumbering[C] = nextValueNumber;
274 return nextValueNumber++;
278 uint32_t v = lookup_or_add(cdep);
279 valueNumbering[C] = v;
283 valueNumbering[C] = nextValueNumber;
284 return nextValueNumber++;
288 /// lookup_or_add - Returns the value number for the specified value, assigning
289 /// it a new number if it did not have one before.
290 uint32_t ValueTable::lookup_or_add(Value *V) {
291 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
292 if (VI != valueNumbering.end())
295 if (!isa<Instruction>(V)) {
296 valueNumbering[V] = nextValueNumber;
297 return nextValueNumber++;
300 Instruction* I = cast<Instruction>(V);
302 switch (I->getOpcode()) {
303 case Instruction::Call:
304 return lookup_or_add_call(cast<CallInst>(I));
305 case Instruction::Add:
306 case Instruction::FAdd:
307 case Instruction::Sub:
308 case Instruction::FSub:
309 case Instruction::Mul:
310 case Instruction::FMul:
311 case Instruction::UDiv:
312 case Instruction::SDiv:
313 case Instruction::FDiv:
314 case Instruction::URem:
315 case Instruction::SRem:
316 case Instruction::FRem:
317 case Instruction::Shl:
318 case Instruction::LShr:
319 case Instruction::AShr:
320 case Instruction::And:
321 case Instruction::Or :
322 case Instruction::Xor:
323 case Instruction::ICmp:
324 case Instruction::FCmp:
325 case Instruction::Trunc:
326 case Instruction::ZExt:
327 case Instruction::SExt:
328 case Instruction::FPToUI:
329 case Instruction::FPToSI:
330 case Instruction::UIToFP:
331 case Instruction::SIToFP:
332 case Instruction::FPTrunc:
333 case Instruction::FPExt:
334 case Instruction::PtrToInt:
335 case Instruction::IntToPtr:
336 case Instruction::BitCast:
337 case Instruction::Select:
338 case Instruction::ExtractElement:
339 case Instruction::InsertElement:
340 case Instruction::ShuffleVector:
341 case Instruction::ExtractValue:
342 case Instruction::InsertValue:
343 case Instruction::GetElementPtr:
344 exp = create_expression(I);
347 valueNumbering[V] = nextValueNumber;
348 return nextValueNumber++;
351 uint32_t& e = expressionNumbering[exp];
352 if (!e) e = nextValueNumber++;
353 valueNumbering[V] = e;
357 /// lookup - Returns the value number of the specified value. Fails if
358 /// the value has not yet been numbered.
359 uint32_t ValueTable::lookup(Value *V) const {
360 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
361 assert(VI != valueNumbering.end() && "Value not numbered?");
365 /// clear - Remove all entries from the ValueTable.
366 void ValueTable::clear() {
367 valueNumbering.clear();
368 expressionNumbering.clear();
372 /// erase - Remove a value from the value numbering.
373 void ValueTable::erase(Value *V) {
374 valueNumbering.erase(V);
377 /// verifyRemoved - Verify that the value is removed from all internal data
379 void ValueTable::verifyRemoved(const Value *V) const {
380 for (DenseMap<Value*, uint32_t>::const_iterator
381 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
382 assert(I->first != V && "Inst still occurs in value numbering map!");
386 //===----------------------------------------------------------------------===//
388 //===----------------------------------------------------------------------===//
392 class GVN : public FunctionPass {
394 MemoryDependenceAnalysis *MD;
396 const TargetData *TD;
400 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
401 /// have that value number. Use findLeader to query it.
402 struct LeaderTableEntry {
405 LeaderTableEntry *Next;
407 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
408 BumpPtrAllocator TableAllocator;
410 SmallVector<Instruction*, 8> InstrsToErase;
412 static char ID; // Pass identification, replacement for typeid
413 explicit GVN(bool noloads = false)
414 : FunctionPass(ID), NoLoads(noloads), MD(0) {
415 initializeGVNPass(*PassRegistry::getPassRegistry());
418 bool runOnFunction(Function &F);
420 /// markInstructionForDeletion - This removes the specified instruction from
421 /// our various maps and marks it for deletion.
422 void markInstructionForDeletion(Instruction *I) {
424 InstrsToErase.push_back(I);
427 const TargetData *getTargetData() const { return TD; }
428 DominatorTree &getDominatorTree() const { return *DT; }
429 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
430 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
432 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
433 /// its value number.
434 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
435 LeaderTableEntry &Curr = LeaderTable[N];
442 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
445 Node->Next = Curr.Next;
449 /// removeFromLeaderTable - Scan the list of values corresponding to a given
450 /// value number, and remove the given value if encountered.
451 void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
452 LeaderTableEntry* Prev = 0;
453 LeaderTableEntry* Curr = &LeaderTable[N];
455 while (Curr->Val != V || Curr->BB != BB) {
461 Prev->Next = Curr->Next;
467 LeaderTableEntry* Next = Curr->Next;
468 Curr->Val = Next->Val;
470 Curr->Next = Next->Next;
475 // List of critical edges to be split between iterations.
476 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
478 // This transformation requires dominator postdominator info
479 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
480 AU.addRequired<DominatorTree>();
482 AU.addRequired<MemoryDependenceAnalysis>();
483 AU.addRequired<AliasAnalysis>();
485 AU.addPreserved<DominatorTree>();
486 AU.addPreserved<AliasAnalysis>();
491 // FIXME: eliminate or document these better
492 bool processLoad(LoadInst *L);
493 bool processInstruction(Instruction *I);
494 bool processNonLocalLoad(LoadInst *L);
495 bool processBlock(BasicBlock *BB);
496 void dump(DenseMap<uint32_t, Value*> &d);
497 bool iterateOnFunction(Function &F);
498 bool performPRE(Function &F);
499 Value *findLeader(BasicBlock *BB, uint32_t num);
500 void cleanupGlobalSets();
501 void verifyRemoved(const Instruction *I) const;
502 bool splitCriticalEdges();
508 // createGVNPass - The public interface to this file...
509 FunctionPass *llvm::createGVNPass(bool NoLoads) {
510 return new GVN(NoLoads);
513 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
514 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
515 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
516 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
517 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
519 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
521 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
522 E = d.end(); I != E; ++I) {
523 errs() << I->first << "\n";
529 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
530 /// we're analyzing is fully available in the specified block. As we go, keep
531 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
532 /// map is actually a tri-state map with the following values:
533 /// 0) we know the block *is not* fully available.
534 /// 1) we know the block *is* fully available.
535 /// 2) we do not know whether the block is fully available or not, but we are
536 /// currently speculating that it will be.
537 /// 3) we are speculating for this block and have used that to speculate for
539 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
540 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
541 // Optimistically assume that the block is fully available and check to see
542 // if we already know about this block in one lookup.
543 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
544 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
546 // If the entry already existed for this block, return the precomputed value.
548 // If this is a speculative "available" value, mark it as being used for
549 // speculation of other blocks.
550 if (IV.first->second == 2)
551 IV.first->second = 3;
552 return IV.first->second != 0;
555 // Otherwise, see if it is fully available in all predecessors.
556 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
558 // If this block has no predecessors, it isn't live-in here.
560 goto SpeculationFailure;
562 for (; PI != PE; ++PI)
563 // If the value isn't fully available in one of our predecessors, then it
564 // isn't fully available in this block either. Undo our previous
565 // optimistic assumption and bail out.
566 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
567 goto SpeculationFailure;
571 // SpeculationFailure - If we get here, we found out that this is not, after
572 // all, a fully-available block. We have a problem if we speculated on this and
573 // used the speculation to mark other blocks as available.
575 char &BBVal = FullyAvailableBlocks[BB];
577 // If we didn't speculate on this, just return with it set to false.
583 // If we did speculate on this value, we could have blocks set to 1 that are
584 // incorrect. Walk the (transitive) successors of this block and mark them as
586 SmallVector<BasicBlock*, 32> BBWorklist;
587 BBWorklist.push_back(BB);
590 BasicBlock *Entry = BBWorklist.pop_back_val();
591 // Note that this sets blocks to 0 (unavailable) if they happen to not
592 // already be in FullyAvailableBlocks. This is safe.
593 char &EntryVal = FullyAvailableBlocks[Entry];
594 if (EntryVal == 0) continue; // Already unavailable.
596 // Mark as unavailable.
599 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
600 BBWorklist.push_back(*I);
601 } while (!BBWorklist.empty());
607 /// CanCoerceMustAliasedValueToLoad - Return true if
608 /// CoerceAvailableValueToLoadType will succeed.
609 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
611 const TargetData &TD) {
612 // If the loaded or stored value is an first class array or struct, don't try
613 // to transform them. We need to be able to bitcast to integer.
614 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
615 StoredVal->getType()->isStructTy() ||
616 StoredVal->getType()->isArrayTy())
619 // The store has to be at least as big as the load.
620 if (TD.getTypeSizeInBits(StoredVal->getType()) <
621 TD.getTypeSizeInBits(LoadTy))
628 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
629 /// then a load from a must-aliased pointer of a different type, try to coerce
630 /// the stored value. LoadedTy is the type of the load we want to replace and
631 /// InsertPt is the place to insert new instructions.
633 /// If we can't do it, return null.
634 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
635 const Type *LoadedTy,
636 Instruction *InsertPt,
637 const TargetData &TD) {
638 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
641 // If this is already the right type, just return it.
642 const Type *StoredValTy = StoredVal->getType();
644 uint64_t StoreSize = TD.getTypeStoreSizeInBits(StoredValTy);
645 uint64_t LoadSize = TD.getTypeStoreSizeInBits(LoadedTy);
647 // If the store and reload are the same size, we can always reuse it.
648 if (StoreSize == LoadSize) {
649 // Pointer to Pointer -> use bitcast.
650 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy())
651 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
653 // Convert source pointers to integers, which can be bitcast.
654 if (StoredValTy->isPointerTy()) {
655 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
656 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
659 const Type *TypeToCastTo = LoadedTy;
660 if (TypeToCastTo->isPointerTy())
661 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
663 if (StoredValTy != TypeToCastTo)
664 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
666 // Cast to pointer if the load needs a pointer type.
667 if (LoadedTy->isPointerTy())
668 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
673 // If the loaded value is smaller than the available value, then we can
674 // extract out a piece from it. If the available value is too small, then we
675 // can't do anything.
676 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
678 // Convert source pointers to integers, which can be manipulated.
679 if (StoredValTy->isPointerTy()) {
680 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
681 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
684 // Convert vectors and fp to integer, which can be manipulated.
685 if (!StoredValTy->isIntegerTy()) {
686 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
687 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
690 // If this is a big-endian system, we need to shift the value down to the low
691 // bits so that a truncate will work.
692 if (TD.isBigEndian()) {
693 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
694 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
697 // Truncate the integer to the right size now.
698 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
699 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
701 if (LoadedTy == NewIntTy)
704 // If the result is a pointer, inttoptr.
705 if (LoadedTy->isPointerTy())
706 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
708 // Otherwise, bitcast.
709 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
712 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
713 /// memdep query of a load that ends up being a clobbering memory write (store,
714 /// memset, memcpy, memmove). This means that the write *may* provide bits used
715 /// by the load but we can't be sure because the pointers don't mustalias.
717 /// Check this case to see if there is anything more we can do before we give
718 /// up. This returns -1 if we have to give up, or a byte number in the stored
719 /// value of the piece that feeds the load.
720 static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr,
722 uint64_t WriteSizeInBits,
723 const TargetData &TD) {
724 // If the loaded or stored value is an first class array or struct, don't try
725 // to transform them. We need to be able to bitcast to integer.
726 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
729 int64_t StoreOffset = 0, LoadOffset = 0;
730 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD);
731 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
732 if (StoreBase != LoadBase)
735 // If the load and store are to the exact same address, they should have been
736 // a must alias. AA must have gotten confused.
737 // FIXME: Study to see if/when this happens. One case is forwarding a memset
738 // to a load from the base of the memset.
740 if (LoadOffset == StoreOffset) {
741 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
742 << "Base = " << *StoreBase << "\n"
743 << "Store Ptr = " << *WritePtr << "\n"
744 << "Store Offs = " << StoreOffset << "\n"
745 << "Load Ptr = " << *LoadPtr << "\n";
750 // If the load and store don't overlap at all, the store doesn't provide
751 // anything to the load. In this case, they really don't alias at all, AA
752 // must have gotten confused.
753 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
755 if ((WriteSizeInBits & 7) | (LoadSize & 7))
757 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
761 bool isAAFailure = false;
762 if (StoreOffset < LoadOffset)
763 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
765 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
769 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
770 << "Base = " << *StoreBase << "\n"
771 << "Store Ptr = " << *WritePtr << "\n"
772 << "Store Offs = " << StoreOffset << "\n"
773 << "Load Ptr = " << *LoadPtr << "\n";
779 // If the Load isn't completely contained within the stored bits, we don't
780 // have all the bits to feed it. We could do something crazy in the future
781 // (issue a smaller load then merge the bits in) but this seems unlikely to be
783 if (StoreOffset > LoadOffset ||
784 StoreOffset+StoreSize < LoadOffset+LoadSize)
787 // Okay, we can do this transformation. Return the number of bytes into the
788 // store that the load is.
789 return LoadOffset-StoreOffset;
792 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
793 /// memdep query of a load that ends up being a clobbering store.
794 static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr,
796 const TargetData &TD) {
797 // Cannot handle reading from store of first-class aggregate yet.
798 if (DepSI->getValueOperand()->getType()->isStructTy() ||
799 DepSI->getValueOperand()->getType()->isArrayTy())
802 Value *StorePtr = DepSI->getPointerOperand();
803 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
804 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
805 StorePtr, StoreSize, TD);
808 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
809 /// memdep query of a load that ends up being clobbered by another load. See if
810 /// the other load can feed into the second load.
811 static int AnalyzeLoadFromClobberingLoad(const Type *LoadTy, Value *LoadPtr,
812 LoadInst *DepLI, const TargetData &TD){
813 // Cannot handle reading from store of first-class aggregate yet.
814 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
817 Value *DepPtr = DepLI->getPointerOperand();
818 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType());
819 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD);
820 if (R != -1) return R;
822 // If we have a load/load clobber an DepLI can be widened to cover this load,
823 // then we should widen it!
824 int64_t LoadOffs = 0;
825 const Value *LoadBase =
826 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD);
827 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
829 unsigned Size = MemoryDependenceAnalysis::
830 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD);
831 if (Size == 0) return -1;
833 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD);
838 static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr,
840 const TargetData &TD) {
841 // If the mem operation is a non-constant size, we can't handle it.
842 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
843 if (SizeCst == 0) return -1;
844 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
846 // If this is memset, we just need to see if the offset is valid in the size
848 if (MI->getIntrinsicID() == Intrinsic::memset)
849 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
852 // If we have a memcpy/memmove, the only case we can handle is if this is a
853 // copy from constant memory. In that case, we can read directly from the
855 MemTransferInst *MTI = cast<MemTransferInst>(MI);
857 Constant *Src = dyn_cast<Constant>(MTI->getSource());
858 if (Src == 0) return -1;
860 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
861 if (GV == 0 || !GV->isConstant()) return -1;
863 // See if the access is within the bounds of the transfer.
864 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
865 MI->getDest(), MemSizeInBits, TD);
869 // Otherwise, see if we can constant fold a load from the constant with the
870 // offset applied as appropriate.
871 Src = ConstantExpr::getBitCast(Src,
872 llvm::Type::getInt8PtrTy(Src->getContext()));
873 Constant *OffsetCst =
874 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
875 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
876 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
877 if (ConstantFoldLoadFromConstPtr(Src, &TD))
883 /// GetStoreValueForLoad - This function is called when we have a
884 /// memdep query of a load that ends up being a clobbering store. This means
885 /// that the store provides bits used by the load but we the pointers don't
886 /// mustalias. Check this case to see if there is anything more we can do
887 /// before we give up.
888 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
890 Instruction *InsertPt, const TargetData &TD){
891 LLVMContext &Ctx = SrcVal->getType()->getContext();
893 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
894 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
896 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
898 // Compute which bits of the stored value are being used by the load. Convert
899 // to an integer type to start with.
900 if (SrcVal->getType()->isPointerTy())
901 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp");
902 if (!SrcVal->getType()->isIntegerTy())
903 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8),
906 // Shift the bits to the least significant depending on endianness.
908 if (TD.isLittleEndian())
911 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
914 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp");
916 if (LoadSize != StoreSize)
917 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8),
920 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
923 /// GetStoreValueForLoad - This function is called when we have a
924 /// memdep query of a load that ends up being a clobbering load. This means
925 /// that the load *may* provide bits used by the load but we can't be sure
926 /// because the pointers don't mustalias. Check this case to see if there is
927 /// anything more we can do before we give up.
928 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
929 const Type *LoadTy, Instruction *InsertPt,
931 const TargetData &TD = *gvn.getTargetData();
932 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
933 // widen SrcVal out to a larger load.
934 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType());
935 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
936 if (Offset+LoadSize > SrcValSize) {
937 assert(!SrcVal->isVolatile() && "Cannot widen volatile load!");
938 assert(isa<IntegerType>(SrcVal->getType())&&"Can't widen non-integer load");
939 // If we have a load/load clobber an DepLI can be widened to cover this
940 // load, then we should widen it to the next power of 2 size big enough!
941 unsigned NewLoadSize = Offset+LoadSize;
942 if (!isPowerOf2_32(NewLoadSize))
943 NewLoadSize = NextPowerOf2(NewLoadSize);
945 Value *PtrVal = SrcVal->getPointerOperand();
947 // Insert the new load after the old load. This ensures that subsequent
948 // memdep queries will find the new load. We can't easily remove the old
949 // load completely because it is already in the value numbering table.
950 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
951 const Type *DestPTy =
952 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
953 DestPTy = PointerType::get(DestPTy,
954 cast<PointerType>(PtrVal->getType())->getAddressSpace());
956 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
957 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
958 NewLoad->takeName(SrcVal);
959 NewLoad->setAlignment(SrcVal->getAlignment());
961 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
962 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
964 // Replace uses of the original load with the wider load. On a big endian
965 // system, we need to shift down to get the relevant bits.
967 if (TD.isBigEndian())
968 RV = Builder.CreateLShr(RV,
969 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
970 RV = Builder.CreateTrunc(RV, SrcVal->getType());
971 SrcVal->replaceAllUsesWith(RV);
973 // We would like to use gvn.markInstructionForDeletion here, but we can't
974 // because the load is already memoized into the leader map table that GVN
975 // tracks. It is potentially possible to remove the load from the table,
976 // but then there all of the operations based on it would need to be
977 // rehashed. Just leave the dead load around.
978 gvn.getMemDep().removeInstruction(SrcVal);
982 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD);
986 /// GetMemInstValueForLoad - This function is called when we have a
987 /// memdep query of a load that ends up being a clobbering mem intrinsic.
988 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
989 const Type *LoadTy, Instruction *InsertPt,
990 const TargetData &TD){
991 LLVMContext &Ctx = LoadTy->getContext();
992 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
994 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
996 // We know that this method is only called when the mem transfer fully
997 // provides the bits for the load.
998 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
999 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1000 // independently of what the offset is.
1001 Value *Val = MSI->getValue();
1003 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1005 Value *OneElt = Val;
1007 // Splat the value out to the right number of bits.
1008 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1009 // If we can double the number of bytes set, do it.
1010 if (NumBytesSet*2 <= LoadSize) {
1011 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1012 Val = Builder.CreateOr(Val, ShVal);
1017 // Otherwise insert one byte at a time.
1018 Value *ShVal = Builder.CreateShl(Val, 1*8);
1019 Val = Builder.CreateOr(OneElt, ShVal);
1023 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1026 // Otherwise, this is a memcpy/memmove from a constant global.
1027 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1028 Constant *Src = cast<Constant>(MTI->getSource());
1030 // Otherwise, see if we can constant fold a load from the constant with the
1031 // offset applied as appropriate.
1032 Src = ConstantExpr::getBitCast(Src,
1033 llvm::Type::getInt8PtrTy(Src->getContext()));
1034 Constant *OffsetCst =
1035 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1036 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
1037 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1038 return ConstantFoldLoadFromConstPtr(Src, &TD);
1043 struct AvailableValueInBlock {
1044 /// BB - The basic block in question.
1047 SimpleVal, // A simple offsetted value that is accessed.
1048 LoadVal, // A value produced by a load.
1049 MemIntrin // A memory intrinsic which is loaded from.
1052 /// V - The value that is live out of the block.
1053 PointerIntPair<Value *, 2, ValType> Val;
1055 /// Offset - The byte offset in Val that is interesting for the load query.
1058 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1059 unsigned Offset = 0) {
1060 AvailableValueInBlock Res;
1062 Res.Val.setPointer(V);
1063 Res.Val.setInt(SimpleVal);
1064 Res.Offset = Offset;
1068 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1069 unsigned Offset = 0) {
1070 AvailableValueInBlock Res;
1072 Res.Val.setPointer(MI);
1073 Res.Val.setInt(MemIntrin);
1074 Res.Offset = Offset;
1078 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
1079 unsigned Offset = 0) {
1080 AvailableValueInBlock Res;
1082 Res.Val.setPointer(LI);
1083 Res.Val.setInt(LoadVal);
1084 Res.Offset = Offset;
1088 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
1089 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
1090 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
1092 Value *getSimpleValue() const {
1093 assert(isSimpleValue() && "Wrong accessor");
1094 return Val.getPointer();
1097 LoadInst *getCoercedLoadValue() const {
1098 assert(isCoercedLoadValue() && "Wrong accessor");
1099 return cast<LoadInst>(Val.getPointer());
1102 MemIntrinsic *getMemIntrinValue() const {
1103 assert(isMemIntrinValue() && "Wrong accessor");
1104 return cast<MemIntrinsic>(Val.getPointer());
1107 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
1108 /// defined here to the specified type. This handles various coercion cases.
1109 Value *MaterializeAdjustedValue(const Type *LoadTy, GVN &gvn) const {
1111 if (isSimpleValue()) {
1112 Res = getSimpleValue();
1113 if (Res->getType() != LoadTy) {
1114 const TargetData *TD = gvn.getTargetData();
1115 assert(TD && "Need target data to handle type mismatch case");
1116 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1119 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1120 << *getSimpleValue() << '\n'
1121 << *Res << '\n' << "\n\n\n");
1123 } else if (isCoercedLoadValue()) {
1124 LoadInst *Load = getCoercedLoadValue();
1125 if (Load->getType() == LoadTy && Offset == 0) {
1128 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1131 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1132 << *getCoercedLoadValue() << '\n'
1133 << *Res << '\n' << "\n\n\n");
1136 const TargetData *TD = gvn.getTargetData();
1137 assert(TD && "Need target data to handle type mismatch case");
1138 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1139 LoadTy, BB->getTerminator(), *TD);
1140 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1141 << " " << *getMemIntrinValue() << '\n'
1142 << *Res << '\n' << "\n\n\n");
1148 } // end anonymous namespace
1150 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1151 /// construct SSA form, allowing us to eliminate LI. This returns the value
1152 /// that should be used at LI's definition site.
1153 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1154 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1156 // Check for the fully redundant, dominating load case. In this case, we can
1157 // just use the dominating value directly.
1158 if (ValuesPerBlock.size() == 1 &&
1159 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1161 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1163 // Otherwise, we have to construct SSA form.
1164 SmallVector<PHINode*, 8> NewPHIs;
1165 SSAUpdater SSAUpdate(&NewPHIs);
1166 SSAUpdate.Initialize(LI->getType(), LI->getName());
1168 const Type *LoadTy = LI->getType();
1170 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1171 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1172 BasicBlock *BB = AV.BB;
1174 if (SSAUpdate.HasValueForBlock(BB))
1177 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1180 // Perform PHI construction.
1181 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1183 // If new PHI nodes were created, notify alias analysis.
1184 if (V->getType()->isPointerTy()) {
1185 AliasAnalysis *AA = gvn.getAliasAnalysis();
1187 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1188 AA->copyValue(LI, NewPHIs[i]);
1190 // Now that we've copied information to the new PHIs, scan through
1191 // them again and inform alias analysis that we've added potentially
1192 // escaping uses to any values that are operands to these PHIs.
1193 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1194 PHINode *P = NewPHIs[i];
1195 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii)
1196 AA->addEscapingUse(P->getOperandUse(2*ii));
1203 static bool isLifetimeStart(const Instruction *Inst) {
1204 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1205 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1209 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1210 /// non-local by performing PHI construction.
1211 bool GVN::processNonLocalLoad(LoadInst *LI) {
1212 // Find the non-local dependencies of the load.
1213 SmallVector<NonLocalDepResult, 64> Deps;
1214 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1215 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1216 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
1217 // << Deps.size() << *LI << '\n');
1219 // If we had to process more than one hundred blocks to find the
1220 // dependencies, this load isn't worth worrying about. Optimizing
1221 // it will be too expensive.
1222 if (Deps.size() > 100)
1225 // If we had a phi translation failure, we'll have a single entry which is a
1226 // clobber in the current block. Reject this early.
1227 if (Deps.size() == 1 && Deps[0].getResult().isClobber() &&
1228 Deps[0].getResult().getInst()->getParent() == LI->getParent()) {
1230 dbgs() << "GVN: non-local load ";
1231 WriteAsOperand(dbgs(), LI);
1232 dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n';
1237 // Filter out useless results (non-locals, etc). Keep track of the blocks
1238 // where we have a value available in repl, also keep track of whether we see
1239 // dependencies that produce an unknown value for the load (such as a call
1240 // that could potentially clobber the load).
1241 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
1242 SmallVector<BasicBlock*, 16> UnavailableBlocks;
1244 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
1245 BasicBlock *DepBB = Deps[i].getBB();
1246 MemDepResult DepInfo = Deps[i].getResult();
1248 if (DepInfo.isClobber()) {
1249 // The address being loaded in this non-local block may not be the same as
1250 // the pointer operand of the load if PHI translation occurs. Make sure
1251 // to consider the right address.
1252 Value *Address = Deps[i].getAddress();
1254 // If the dependence is to a store that writes to a superset of the bits
1255 // read by the load, we can extract the bits we need for the load from the
1257 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1258 if (TD && Address) {
1259 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1262 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1263 DepSI->getValueOperand(),
1270 // Check to see if we have something like this:
1273 // if we have this, replace the later with an extraction from the former.
1274 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1275 // If this is a clobber and L is the first instruction in its block, then
1276 // we have the first instruction in the entry block.
1277 if (DepLI != LI && Address && TD) {
1278 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
1279 LI->getPointerOperand(),
1283 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1290 // If the clobbering value is a memset/memcpy/memmove, see if we can
1291 // forward a value on from it.
1292 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1293 if (TD && Address) {
1294 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1297 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1304 UnavailableBlocks.push_back(DepBB);
1308 Instruction *DepInst = DepInfo.getInst();
1310 // Loading the allocation -> undef.
1311 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1312 // Loading immediately after lifetime begin -> undef.
1313 isLifetimeStart(DepInst)) {
1314 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1315 UndefValue::get(LI->getType())));
1319 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1320 // Reject loads and stores that are to the same address but are of
1321 // different types if we have to.
1322 if (S->getValueOperand()->getType() != LI->getType()) {
1323 // If the stored value is larger or equal to the loaded value, we can
1325 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1326 LI->getType(), *TD)) {
1327 UnavailableBlocks.push_back(DepBB);
1332 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1333 S->getValueOperand()));
1337 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1338 // If the types mismatch and we can't handle it, reject reuse of the load.
1339 if (LD->getType() != LI->getType()) {
1340 // If the stored value is larger or equal to the loaded value, we can
1342 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1343 UnavailableBlocks.push_back(DepBB);
1347 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1351 UnavailableBlocks.push_back(DepBB);
1355 // If we have no predecessors that produce a known value for this load, exit
1357 if (ValuesPerBlock.empty()) return false;
1359 // If all of the instructions we depend on produce a known value for this
1360 // load, then it is fully redundant and we can use PHI insertion to compute
1361 // its value. Insert PHIs and remove the fully redundant value now.
1362 if (UnavailableBlocks.empty()) {
1363 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1365 // Perform PHI construction.
1366 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1367 LI->replaceAllUsesWith(V);
1369 if (isa<PHINode>(V))
1371 if (V->getType()->isPointerTy())
1372 MD->invalidateCachedPointerInfo(V);
1373 markInstructionForDeletion(LI);
1378 if (!EnablePRE || !EnableLoadPRE)
1381 // Okay, we have *some* definitions of the value. This means that the value
1382 // is available in some of our (transitive) predecessors. Lets think about
1383 // doing PRE of this load. This will involve inserting a new load into the
1384 // predecessor when it's not available. We could do this in general, but
1385 // prefer to not increase code size. As such, we only do this when we know
1386 // that we only have to insert *one* load (which means we're basically moving
1387 // the load, not inserting a new one).
1389 SmallPtrSet<BasicBlock *, 4> Blockers;
1390 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1391 Blockers.insert(UnavailableBlocks[i]);
1393 // Lets find first basic block with more than one predecessor. Walk backwards
1394 // through predecessors if needed.
1395 BasicBlock *LoadBB = LI->getParent();
1396 BasicBlock *TmpBB = LoadBB;
1398 bool isSinglePred = false;
1399 bool allSingleSucc = true;
1400 while (TmpBB->getSinglePredecessor()) {
1401 isSinglePred = true;
1402 TmpBB = TmpBB->getSinglePredecessor();
1403 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1405 if (Blockers.count(TmpBB))
1408 // If any of these blocks has more than one successor (i.e. if the edge we
1409 // just traversed was critical), then there are other paths through this
1410 // block along which the load may not be anticipated. Hoisting the load
1411 // above this block would be adding the load to execution paths along
1412 // which it was not previously executed.
1413 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1420 // FIXME: It is extremely unclear what this loop is doing, other than
1421 // artificially restricting loadpre.
1424 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1425 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1426 if (AV.isSimpleValue())
1427 // "Hot" Instruction is in some loop (because it dominates its dep.
1429 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1430 if (DT->dominates(LI, I)) {
1436 // We are interested only in "hot" instructions. We don't want to do any
1437 // mis-optimizations here.
1442 // Check to see how many predecessors have the loaded value fully
1444 DenseMap<BasicBlock*, Value*> PredLoads;
1445 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1446 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1447 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1448 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1449 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1451 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
1452 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1454 BasicBlock *Pred = *PI;
1455 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) {
1458 PredLoads[Pred] = 0;
1460 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1461 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1462 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1463 << Pred->getName() << "': " << *LI << '\n');
1466 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
1467 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
1470 if (!NeedToSplit.empty()) {
1471 toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
1475 // Decide whether PRE is profitable for this load.
1476 unsigned NumUnavailablePreds = PredLoads.size();
1477 assert(NumUnavailablePreds != 0 &&
1478 "Fully available value should be eliminated above!");
1480 // If this load is unavailable in multiple predecessors, reject it.
1481 // FIXME: If we could restructure the CFG, we could make a common pred with
1482 // all the preds that don't have an available LI and insert a new load into
1484 if (NumUnavailablePreds != 1)
1487 // Check if the load can safely be moved to all the unavailable predecessors.
1488 bool CanDoPRE = true;
1489 SmallVector<Instruction*, 8> NewInsts;
1490 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1491 E = PredLoads.end(); I != E; ++I) {
1492 BasicBlock *UnavailablePred = I->first;
1494 // Do PHI translation to get its value in the predecessor if necessary. The
1495 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1497 // If all preds have a single successor, then we know it is safe to insert
1498 // the load on the pred (?!?), so we can insert code to materialize the
1499 // pointer if it is not available.
1500 PHITransAddr Address(LI->getPointerOperand(), TD);
1502 if (allSingleSucc) {
1503 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1506 Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
1507 LoadPtr = Address.getAddr();
1510 // If we couldn't find or insert a computation of this phi translated value,
1513 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1514 << *LI->getPointerOperand() << "\n");
1519 // Make sure it is valid to move this load here. We have to watch out for:
1520 // @1 = getelementptr (i8* p, ...
1521 // test p and branch if == 0
1523 // It is valid to have the getelementptr before the test, even if p can
1524 // be 0, as getelementptr only does address arithmetic.
1525 // If we are not pushing the value through any multiple-successor blocks
1526 // we do not have this case. Otherwise, check that the load is safe to
1527 // put anywhere; this can be improved, but should be conservatively safe.
1528 if (!allSingleSucc &&
1529 // FIXME: REEVALUTE THIS.
1530 !isSafeToLoadUnconditionally(LoadPtr,
1531 UnavailablePred->getTerminator(),
1532 LI->getAlignment(), TD)) {
1537 I->second = LoadPtr;
1541 while (!NewInsts.empty()) {
1542 Instruction *I = NewInsts.pop_back_val();
1543 if (MD) MD->removeInstruction(I);
1544 I->eraseFromParent();
1549 // Okay, we can eliminate this load by inserting a reload in the predecessor
1550 // and using PHI construction to get the value in the other predecessors, do
1552 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1553 DEBUG(if (!NewInsts.empty())
1554 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1555 << *NewInsts.back() << '\n');
1557 // Assign value numbers to the new instructions.
1558 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1559 // FIXME: We really _ought_ to insert these value numbers into their
1560 // parent's availability map. However, in doing so, we risk getting into
1561 // ordering issues. If a block hasn't been processed yet, we would be
1562 // marking a value as AVAIL-IN, which isn't what we intend.
1563 VN.lookup_or_add(NewInsts[i]);
1566 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1567 E = PredLoads.end(); I != E; ++I) {
1568 BasicBlock *UnavailablePred = I->first;
1569 Value *LoadPtr = I->second;
1571 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1573 UnavailablePred->getTerminator());
1575 // Transfer the old load's TBAA tag to the new load.
1576 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1577 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1579 // Add the newly created load.
1580 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1582 MD->invalidateCachedPointerInfo(LoadPtr);
1583 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1586 // Perform PHI construction.
1587 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1588 LI->replaceAllUsesWith(V);
1589 if (isa<PHINode>(V))
1591 if (V->getType()->isPointerTy())
1592 MD->invalidateCachedPointerInfo(V);
1593 markInstructionForDeletion(LI);
1598 /// processLoad - Attempt to eliminate a load, first by eliminating it
1599 /// locally, and then attempting non-local elimination if that fails.
1600 bool GVN::processLoad(LoadInst *L) {
1604 if (L->isVolatile())
1607 // ... to a pointer that has been loaded from before...
1608 MemDepResult Dep = MD->getDependency(L);
1610 // If we have a clobber and target data is around, see if this is a clobber
1611 // that we can fix up through code synthesis.
1612 if (Dep.isClobber() && TD) {
1613 // Check to see if we have something like this:
1614 // store i32 123, i32* %P
1615 // %A = bitcast i32* %P to i8*
1616 // %B = gep i8* %A, i32 1
1619 // We could do that by recognizing if the clobber instructions are obviously
1620 // a common base + constant offset, and if the previous store (or memset)
1621 // completely covers this load. This sort of thing can happen in bitfield
1623 Value *AvailVal = 0;
1624 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1625 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1626 L->getPointerOperand(),
1629 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1630 L->getType(), L, *TD);
1633 // Check to see if we have something like this:
1636 // if we have this, replace the later with an extraction from the former.
1637 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1638 // If this is a clobber and L is the first instruction in its block, then
1639 // we have the first instruction in the entry block.
1643 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1644 L->getPointerOperand(),
1647 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1650 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1651 // a value on from it.
1652 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1653 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1654 L->getPointerOperand(),
1657 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
1661 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1662 << *AvailVal << '\n' << *L << "\n\n\n");
1664 // Replace the load!
1665 L->replaceAllUsesWith(AvailVal);
1666 if (AvailVal->getType()->isPointerTy())
1667 MD->invalidateCachedPointerInfo(AvailVal);
1668 markInstructionForDeletion(L);
1674 // If the value isn't available, don't do anything!
1675 if (Dep.isClobber()) {
1677 // fast print dep, using operator<< on instruction is too slow.
1678 dbgs() << "GVN: load ";
1679 WriteAsOperand(dbgs(), L);
1680 Instruction *I = Dep.getInst();
1681 dbgs() << " is clobbered by " << *I << '\n';
1686 // If it is defined in another block, try harder.
1687 if (Dep.isNonLocal())
1688 return processNonLocalLoad(L);
1690 Instruction *DepInst = Dep.getInst();
1691 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1692 Value *StoredVal = DepSI->getValueOperand();
1694 // The store and load are to a must-aliased pointer, but they may not
1695 // actually have the same type. See if we know how to reuse the stored
1696 // value (depending on its type).
1697 if (StoredVal->getType() != L->getType()) {
1699 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1704 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1705 << '\n' << *L << "\n\n\n");
1712 L->replaceAllUsesWith(StoredVal);
1713 if (StoredVal->getType()->isPointerTy())
1714 MD->invalidateCachedPointerInfo(StoredVal);
1715 markInstructionForDeletion(L);
1720 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1721 Value *AvailableVal = DepLI;
1723 // The loads are of a must-aliased pointer, but they may not actually have
1724 // the same type. See if we know how to reuse the previously loaded value
1725 // (depending on its type).
1726 if (DepLI->getType() != L->getType()) {
1728 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1730 if (AvailableVal == 0)
1733 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1734 << "\n" << *L << "\n\n\n");
1741 L->replaceAllUsesWith(AvailableVal);
1742 if (DepLI->getType()->isPointerTy())
1743 MD->invalidateCachedPointerInfo(DepLI);
1744 markInstructionForDeletion(L);
1749 // If this load really doesn't depend on anything, then we must be loading an
1750 // undef value. This can happen when loading for a fresh allocation with no
1751 // intervening stores, for example.
1752 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1753 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1754 markInstructionForDeletion(L);
1759 // If this load occurs either right after a lifetime begin,
1760 // then the loaded value is undefined.
1761 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1762 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1763 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1764 markInstructionForDeletion(L);
1773 // findLeader - In order to find a leader for a given value number at a
1774 // specific basic block, we first obtain the list of all Values for that number,
1775 // and then scan the list to find one whose block dominates the block in
1776 // question. This is fast because dominator tree queries consist of only
1777 // a few comparisons of DFS numbers.
1778 Value *GVN::findLeader(BasicBlock *BB, uint32_t num) {
1779 LeaderTableEntry Vals = LeaderTable[num];
1780 if (!Vals.Val) return 0;
1783 if (DT->dominates(Vals.BB, BB)) {
1785 if (isa<Constant>(Val)) return Val;
1788 LeaderTableEntry* Next = Vals.Next;
1790 if (DT->dominates(Next->BB, BB)) {
1791 if (isa<Constant>(Next->Val)) return Next->Val;
1792 if (!Val) Val = Next->Val;
1802 /// processInstruction - When calculating availability, handle an instruction
1803 /// by inserting it into the appropriate sets
1804 bool GVN::processInstruction(Instruction *I) {
1805 // Ignore dbg info intrinsics.
1806 if (isa<DbgInfoIntrinsic>(I))
1809 // If the instruction can be easily simplified then do so now in preference
1810 // to value numbering it. Value numbering often exposes redundancies, for
1811 // example if it determines that %y is equal to %x then the instruction
1812 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1813 if (Value *V = SimplifyInstruction(I, TD, DT)) {
1814 I->replaceAllUsesWith(V);
1815 if (MD && V->getType()->isPointerTy())
1816 MD->invalidateCachedPointerInfo(V);
1817 markInstructionForDeletion(I);
1821 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1822 if (processLoad(LI))
1825 unsigned Num = VN.lookup_or_add(LI);
1826 addToLeaderTable(Num, LI, LI->getParent());
1830 // For conditions branches, we can perform simple conditional propagation on
1831 // the condition value itself.
1832 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1833 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1836 Value *BranchCond = BI->getCondition();
1837 uint32_t CondVN = VN.lookup_or_add(BranchCond);
1839 BasicBlock *TrueSucc = BI->getSuccessor(0);
1840 BasicBlock *FalseSucc = BI->getSuccessor(1);
1842 if (TrueSucc->getSinglePredecessor())
1843 addToLeaderTable(CondVN,
1844 ConstantInt::getTrue(TrueSucc->getContext()),
1846 if (FalseSucc->getSinglePredecessor())
1847 addToLeaderTable(CondVN,
1848 ConstantInt::getFalse(TrueSucc->getContext()),
1854 // Instructions with void type don't return a value, so there's
1855 // no point in trying to find redudancies in them.
1856 if (I->getType()->isVoidTy()) return false;
1858 uint32_t NextNum = VN.getNextUnusedValueNumber();
1859 unsigned Num = VN.lookup_or_add(I);
1861 // Allocations are always uniquely numbered, so we can save time and memory
1862 // by fast failing them.
1863 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
1864 addToLeaderTable(Num, I, I->getParent());
1868 // If the number we were assigned was a brand new VN, then we don't
1869 // need to do a lookup to see if the number already exists
1870 // somewhere in the domtree: it can't!
1871 if (Num == NextNum) {
1872 addToLeaderTable(Num, I, I->getParent());
1876 // Perform fast-path value-number based elimination of values inherited from
1878 Value *repl = findLeader(I->getParent(), Num);
1880 // Failure, just remember this instance for future use.
1881 addToLeaderTable(Num, I, I->getParent());
1886 I->replaceAllUsesWith(repl);
1887 if (MD && repl->getType()->isPointerTy())
1888 MD->invalidateCachedPointerInfo(repl);
1889 markInstructionForDeletion(I);
1893 /// runOnFunction - This is the main transformation entry point for a function.
1894 bool GVN::runOnFunction(Function& F) {
1896 MD = &getAnalysis<MemoryDependenceAnalysis>();
1897 DT = &getAnalysis<DominatorTree>();
1898 TD = getAnalysisIfAvailable<TargetData>();
1899 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1903 bool Changed = false;
1904 bool ShouldContinue = true;
1906 // Merge unconditional branches, allowing PRE to catch more
1907 // optimization opportunities.
1908 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1909 BasicBlock *BB = FI++;
1911 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1912 if (removedBlock) ++NumGVNBlocks;
1914 Changed |= removedBlock;
1917 unsigned Iteration = 0;
1918 while (ShouldContinue) {
1919 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
1920 ShouldContinue = iterateOnFunction(F);
1921 if (splitCriticalEdges())
1922 ShouldContinue = true;
1923 Changed |= ShouldContinue;
1928 bool PREChanged = true;
1929 while (PREChanged) {
1930 PREChanged = performPRE(F);
1931 Changed |= PREChanged;
1934 // FIXME: Should perform GVN again after PRE does something. PRE can move
1935 // computations into blocks where they become fully redundant. Note that
1936 // we can't do this until PRE's critical edge splitting updates memdep.
1937 // Actually, when this happens, we should just fully integrate PRE into GVN.
1939 cleanupGlobalSets();
1945 bool GVN::processBlock(BasicBlock *BB) {
1946 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
1947 // (and incrementing BI before processing an instruction).
1948 assert(InstrsToErase.empty() &&
1949 "We expect InstrsToErase to be empty across iterations");
1950 bool ChangedFunction = false;
1952 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1954 ChangedFunction |= processInstruction(BI);
1955 if (InstrsToErase.empty()) {
1960 // If we need some instructions deleted, do it now.
1961 NumGVNInstr += InstrsToErase.size();
1963 // Avoid iterator invalidation.
1964 bool AtStart = BI == BB->begin();
1968 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(),
1969 E = InstrsToErase.end(); I != E; ++I) {
1970 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
1971 if (MD) MD->removeInstruction(*I);
1972 (*I)->eraseFromParent();
1973 DEBUG(verifyRemoved(*I));
1975 InstrsToErase.clear();
1983 return ChangedFunction;
1986 /// performPRE - Perform a purely local form of PRE that looks for diamond
1987 /// control flow patterns and attempts to perform simple PRE at the join point.
1988 bool GVN::performPRE(Function &F) {
1989 bool Changed = false;
1990 DenseMap<BasicBlock*, Value*> predMap;
1991 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
1992 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
1993 BasicBlock *CurrentBlock = *DI;
1995 // Nothing to PRE in the entry block.
1996 if (CurrentBlock == &F.getEntryBlock()) continue;
1998 for (BasicBlock::iterator BI = CurrentBlock->begin(),
1999 BE = CurrentBlock->end(); BI != BE; ) {
2000 Instruction *CurInst = BI++;
2002 if (isa<AllocaInst>(CurInst) ||
2003 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2004 CurInst->getType()->isVoidTy() ||
2005 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2006 isa<DbgInfoIntrinsic>(CurInst))
2009 // We don't currently value number ANY inline asm calls.
2010 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2011 if (CallI->isInlineAsm())
2014 uint32_t ValNo = VN.lookup(CurInst);
2016 // Look for the predecessors for PRE opportunities. We're
2017 // only trying to solve the basic diamond case, where
2018 // a value is computed in the successor and one predecessor,
2019 // but not the other. We also explicitly disallow cases
2020 // where the successor is its own predecessor, because they're
2021 // more complicated to get right.
2022 unsigned NumWith = 0;
2023 unsigned NumWithout = 0;
2024 BasicBlock *PREPred = 0;
2027 for (pred_iterator PI = pred_begin(CurrentBlock),
2028 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2029 BasicBlock *P = *PI;
2030 // We're not interested in PRE where the block is its
2031 // own predecessor, or in blocks with predecessors
2032 // that are not reachable.
2033 if (P == CurrentBlock) {
2036 } else if (!DT->dominates(&F.getEntryBlock(), P)) {
2041 Value* predV = findLeader(P, ValNo);
2045 } else if (predV == CurInst) {
2053 // Don't do PRE when it might increase code size, i.e. when
2054 // we would need to insert instructions in more than one pred.
2055 if (NumWithout != 1 || NumWith == 0)
2058 // Don't do PRE across indirect branch.
2059 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2062 // We can't do PRE safely on a critical edge, so instead we schedule
2063 // the edge to be split and perform the PRE the next time we iterate
2065 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2066 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2067 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2071 // Instantiate the expression in the predecessor that lacked it.
2072 // Because we are going top-down through the block, all value numbers
2073 // will be available in the predecessor by the time we need them. Any
2074 // that weren't originally present will have been instantiated earlier
2076 Instruction *PREInstr = CurInst->clone();
2077 bool success = true;
2078 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2079 Value *Op = PREInstr->getOperand(i);
2080 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2083 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2084 PREInstr->setOperand(i, V);
2091 // Fail out if we encounter an operand that is not available in
2092 // the PRE predecessor. This is typically because of loads which
2093 // are not value numbered precisely.
2096 DEBUG(verifyRemoved(PREInstr));
2100 PREInstr->insertBefore(PREPred->getTerminator());
2101 PREInstr->setName(CurInst->getName() + ".pre");
2102 predMap[PREPred] = PREInstr;
2103 VN.add(PREInstr, ValNo);
2106 // Update the availability map to include the new instruction.
2107 addToLeaderTable(ValNo, PREInstr, PREPred);
2109 // Create a PHI to make the value available in this block.
2110 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2111 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE),
2112 CurInst->getName() + ".pre-phi",
2113 CurrentBlock->begin());
2114 for (pred_iterator PI = PB; PI != PE; ++PI) {
2115 BasicBlock *P = *PI;
2116 Phi->addIncoming(predMap[P], P);
2120 addToLeaderTable(ValNo, Phi, CurrentBlock);
2122 CurInst->replaceAllUsesWith(Phi);
2123 if (Phi->getType()->isPointerTy()) {
2124 // Because we have added a PHI-use of the pointer value, it has now
2125 // "escaped" from alias analysis' perspective. We need to inform
2127 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii)
2128 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(2*ii));
2131 MD->invalidateCachedPointerInfo(Phi);
2134 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2136 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2137 if (MD) MD->removeInstruction(CurInst);
2138 CurInst->eraseFromParent();
2139 DEBUG(verifyRemoved(CurInst));
2144 if (splitCriticalEdges())
2150 /// splitCriticalEdges - Split critical edges found during the previous
2151 /// iteration that may enable further optimization.
2152 bool GVN::splitCriticalEdges() {
2153 if (toSplit.empty())
2156 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2157 SplitCriticalEdge(Edge.first, Edge.second, this);
2158 } while (!toSplit.empty());
2159 if (MD) MD->invalidateCachedPredecessors();
2163 /// iterateOnFunction - Executes one iteration of GVN
2164 bool GVN::iterateOnFunction(Function &F) {
2165 cleanupGlobalSets();
2167 // Top-down walk of the dominator tree
2168 bool Changed = false;
2170 // Needed for value numbering with phi construction to work.
2171 ReversePostOrderTraversal<Function*> RPOT(&F);
2172 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2173 RE = RPOT.end(); RI != RE; ++RI)
2174 Changed |= processBlock(*RI);
2176 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2177 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2178 Changed |= processBlock(DI->getBlock());
2184 void GVN::cleanupGlobalSets() {
2186 LeaderTable.clear();
2187 TableAllocator.Reset();
2190 /// verifyRemoved - Verify that the specified instruction does not occur in our
2191 /// internal data structures.
2192 void GVN::verifyRemoved(const Instruction *Inst) const {
2193 VN.verifyRemoved(Inst);
2195 // Walk through the value number scope to make sure the instruction isn't
2196 // ferreted away in it.
2197 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2198 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2199 const LeaderTableEntry *Node = &I->second;
2200 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2202 while (Node->Next) {
2204 assert(Node->Val != Inst && "Inst still in value numbering scope!");