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;
70 Expression(uint32_t o) : opcode(o) { }
72 bool operator==(const Expression &other) const {
73 if (opcode != other.opcode)
75 else if (opcode == ~0U || opcode == ~1U)
77 else if (type != other.type)
79 else if (varargs != other.varargs)
87 DenseMap<Value*, uint32_t> valueNumbering;
88 DenseMap<Expression, uint32_t> expressionNumbering;
90 MemoryDependenceAnalysis* MD;
93 uint32_t nextValueNumber;
95 Expression create_expression(Instruction* I);
96 uint32_t lookup_or_add_call(CallInst* C);
98 ValueTable() : nextValueNumber(1) { }
99 uint32_t lookup_or_add(Value *V);
100 uint32_t lookup(Value *V) const;
101 void add(Value *V, uint32_t num);
103 void erase(Value *v);
104 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
105 AliasAnalysis *getAliasAnalysis() const { return AA; }
106 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
107 void setDomTree(DominatorTree* D) { DT = D; }
108 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
109 void verifyRemoved(const Value *) const;
114 template <> struct DenseMapInfo<Expression> {
115 static inline Expression getEmptyKey() {
119 static inline Expression getTombstoneKey() {
123 static unsigned getHashValue(const Expression e) {
124 unsigned hash = e.opcode;
126 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
127 (unsigned)((uintptr_t)e.type >> 9));
129 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
130 E = e.varargs.end(); I != E; ++I)
131 hash = *I + hash * 37;
135 static bool isEqual(const Expression &LHS, const Expression &RHS) {
142 //===----------------------------------------------------------------------===//
143 // ValueTable Internal Functions
144 //===----------------------------------------------------------------------===//
147 Expression ValueTable::create_expression(Instruction *I) {
149 e.type = I->getType();
150 e.opcode = I->getOpcode();
151 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
153 e.varargs.push_back(lookup_or_add(*OI));
155 if (CmpInst *C = dyn_cast<CmpInst>(I))
156 e.opcode = (C->getOpcode() << 8) | C->getPredicate();
157 else if (ExtractValueInst *E = dyn_cast<ExtractValueInst>(I)) {
158 for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
160 e.varargs.push_back(*II);
161 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
162 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
164 e.varargs.push_back(*II);
170 //===----------------------------------------------------------------------===//
171 // ValueTable External Functions
172 //===----------------------------------------------------------------------===//
174 /// add - Insert a value into the table with a specified value number.
175 void ValueTable::add(Value *V, uint32_t num) {
176 valueNumbering.insert(std::make_pair(V, num));
179 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
180 if (AA->doesNotAccessMemory(C)) {
181 Expression exp = create_expression(C);
182 uint32_t& e = expressionNumbering[exp];
183 if (!e) e = nextValueNumber++;
184 valueNumbering[C] = e;
186 } else if (AA->onlyReadsMemory(C)) {
187 Expression exp = create_expression(C);
188 uint32_t& e = expressionNumbering[exp];
190 e = nextValueNumber++;
191 valueNumbering[C] = e;
195 e = nextValueNumber++;
196 valueNumbering[C] = e;
200 MemDepResult local_dep = MD->getDependency(C);
202 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
203 valueNumbering[C] = nextValueNumber;
204 return nextValueNumber++;
207 if (local_dep.isDef()) {
208 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
210 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
211 valueNumbering[C] = nextValueNumber;
212 return nextValueNumber++;
215 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
216 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
217 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
219 valueNumbering[C] = nextValueNumber;
220 return nextValueNumber++;
224 uint32_t v = lookup_or_add(local_cdep);
225 valueNumbering[C] = v;
230 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
231 MD->getNonLocalCallDependency(CallSite(C));
232 // FIXME: call/call dependencies for readonly calls should return def, not
233 // clobber! Move the checking logic to MemDep!
236 // Check to see if we have a single dominating call instruction that is
238 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
239 const NonLocalDepEntry *I = &deps[i];
240 // Ignore non-local dependencies.
241 if (I->getResult().isNonLocal())
244 // We don't handle non-depedencies. If we already have a call, reject
245 // instruction dependencies.
246 if (I->getResult().isClobber() || cdep != 0) {
251 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
252 // FIXME: All duplicated with non-local case.
253 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
254 cdep = NonLocalDepCall;
263 valueNumbering[C] = nextValueNumber;
264 return nextValueNumber++;
267 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
268 valueNumbering[C] = nextValueNumber;
269 return nextValueNumber++;
271 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
272 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
273 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
275 valueNumbering[C] = nextValueNumber;
276 return nextValueNumber++;
280 uint32_t v = lookup_or_add(cdep);
281 valueNumbering[C] = v;
285 valueNumbering[C] = nextValueNumber;
286 return nextValueNumber++;
290 /// lookup_or_add - Returns the value number for the specified value, assigning
291 /// it a new number if it did not have one before.
292 uint32_t ValueTable::lookup_or_add(Value *V) {
293 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
294 if (VI != valueNumbering.end())
297 if (!isa<Instruction>(V)) {
298 valueNumbering[V] = nextValueNumber;
299 return nextValueNumber++;
302 Instruction* I = cast<Instruction>(V);
304 switch (I->getOpcode()) {
305 case Instruction::Call:
306 return lookup_or_add_call(cast<CallInst>(I));
307 case Instruction::Add:
308 case Instruction::FAdd:
309 case Instruction::Sub:
310 case Instruction::FSub:
311 case Instruction::Mul:
312 case Instruction::FMul:
313 case Instruction::UDiv:
314 case Instruction::SDiv:
315 case Instruction::FDiv:
316 case Instruction::URem:
317 case Instruction::SRem:
318 case Instruction::FRem:
319 case Instruction::Shl:
320 case Instruction::LShr:
321 case Instruction::AShr:
322 case Instruction::And:
323 case Instruction::Or :
324 case Instruction::Xor:
325 case Instruction::ICmp:
326 case Instruction::FCmp:
327 case Instruction::Trunc:
328 case Instruction::ZExt:
329 case Instruction::SExt:
330 case Instruction::FPToUI:
331 case Instruction::FPToSI:
332 case Instruction::UIToFP:
333 case Instruction::SIToFP:
334 case Instruction::FPTrunc:
335 case Instruction::FPExt:
336 case Instruction::PtrToInt:
337 case Instruction::IntToPtr:
338 case Instruction::BitCast:
339 case Instruction::Select:
340 case Instruction::ExtractElement:
341 case Instruction::InsertElement:
342 case Instruction::ShuffleVector:
343 case Instruction::ExtractValue:
344 case Instruction::InsertValue:
345 case Instruction::GetElementPtr:
346 exp = create_expression(I);
349 valueNumbering[V] = nextValueNumber;
350 return nextValueNumber++;
353 uint32_t& e = expressionNumbering[exp];
354 if (!e) e = nextValueNumber++;
355 valueNumbering[V] = e;
359 /// lookup - Returns the value number of the specified value. Fails if
360 /// the value has not yet been numbered.
361 uint32_t ValueTable::lookup(Value *V) const {
362 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
363 assert(VI != valueNumbering.end() && "Value not numbered?");
367 /// clear - Remove all entries from the ValueTable
368 void ValueTable::clear() {
369 valueNumbering.clear();
370 expressionNumbering.clear();
374 /// erase - Remove a value from the value numbering
375 void ValueTable::erase(Value *V) {
376 valueNumbering.erase(V);
379 /// verifyRemoved - Verify that the value is removed from all internal data
381 void ValueTable::verifyRemoved(const Value *V) const {
382 for (DenseMap<Value*, uint32_t>::const_iterator
383 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
384 assert(I->first != V && "Inst still occurs in value numbering map!");
388 //===----------------------------------------------------------------------===//
390 //===----------------------------------------------------------------------===//
394 class GVN : public FunctionPass {
395 bool runOnFunction(Function &F);
397 static char ID; // Pass identification, replacement for typeid
398 explicit GVN(bool noloads = false)
399 : FunctionPass(ID), NoLoads(noloads), MD(0) {
400 initializeGVNPass(*PassRegistry::getPassRegistry());
405 MemoryDependenceAnalysis *MD;
407 const TargetData* TD;
411 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
412 /// have that value number. Use findLeader to query it.
413 struct LeaderTableEntry {
416 LeaderTableEntry *Next;
418 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
419 BumpPtrAllocator TableAllocator;
421 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
422 /// its value number.
423 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
424 LeaderTableEntry& Curr = LeaderTable[N];
431 LeaderTableEntry* Node = TableAllocator.Allocate<LeaderTableEntry>();
434 Node->Next = Curr.Next;
438 /// removeFromLeaderTable - Scan the list of values corresponding to a given
439 /// value number, and remove the given value if encountered.
440 void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
441 LeaderTableEntry* Prev = 0;
442 LeaderTableEntry* Curr = &LeaderTable[N];
444 while (Curr->Val != V || Curr->BB != BB) {
450 Prev->Next = Curr->Next;
456 LeaderTableEntry* Next = Curr->Next;
457 Curr->Val = Next->Val;
459 Curr->Next = Next->Next;
464 // List of critical edges to be split between iterations.
465 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
467 // This transformation requires dominator postdominator info
468 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
469 AU.addRequired<DominatorTree>();
471 AU.addRequired<MemoryDependenceAnalysis>();
472 AU.addRequired<AliasAnalysis>();
474 AU.addPreserved<DominatorTree>();
475 AU.addPreserved<AliasAnalysis>();
479 // FIXME: eliminate or document these better
480 bool processLoad(LoadInst* L,
481 SmallVectorImpl<Instruction*> &toErase);
482 bool processInstruction(Instruction *I,
483 SmallVectorImpl<Instruction*> &toErase);
484 bool processNonLocalLoad(LoadInst* L,
485 SmallVectorImpl<Instruction*> &toErase);
486 bool processBlock(BasicBlock *BB);
487 void dump(DenseMap<uint32_t, Value*>& d);
488 bool iterateOnFunction(Function &F);
489 bool performPRE(Function& F);
490 Value *findLeader(BasicBlock *BB, uint32_t num);
491 void cleanupGlobalSets();
492 void verifyRemoved(const Instruction *I) const;
493 bool splitCriticalEdges();
499 // createGVNPass - The public interface to this file...
500 FunctionPass *llvm::createGVNPass(bool NoLoads) {
501 return new GVN(NoLoads);
504 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
505 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
506 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
507 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
508 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
510 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
512 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
513 E = d.end(); I != E; ++I) {
514 errs() << I->first << "\n";
520 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
521 /// we're analyzing is fully available in the specified block. As we go, keep
522 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
523 /// map is actually a tri-state map with the following values:
524 /// 0) we know the block *is not* fully available.
525 /// 1) we know the block *is* fully available.
526 /// 2) we do not know whether the block is fully available or not, but we are
527 /// currently speculating that it will be.
528 /// 3) we are speculating for this block and have used that to speculate for
530 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
531 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
532 // Optimistically assume that the block is fully available and check to see
533 // if we already know about this block in one lookup.
534 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
535 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
537 // If the entry already existed for this block, return the precomputed value.
539 // If this is a speculative "available" value, mark it as being used for
540 // speculation of other blocks.
541 if (IV.first->second == 2)
542 IV.first->second = 3;
543 return IV.first->second != 0;
546 // Otherwise, see if it is fully available in all predecessors.
547 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
549 // If this block has no predecessors, it isn't live-in here.
551 goto SpeculationFailure;
553 for (; PI != PE; ++PI)
554 // If the value isn't fully available in one of our predecessors, then it
555 // isn't fully available in this block either. Undo our previous
556 // optimistic assumption and bail out.
557 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
558 goto SpeculationFailure;
562 // SpeculationFailure - If we get here, we found out that this is not, after
563 // all, a fully-available block. We have a problem if we speculated on this and
564 // used the speculation to mark other blocks as available.
566 char &BBVal = FullyAvailableBlocks[BB];
568 // If we didn't speculate on this, just return with it set to false.
574 // If we did speculate on this value, we could have blocks set to 1 that are
575 // incorrect. Walk the (transitive) successors of this block and mark them as
577 SmallVector<BasicBlock*, 32> BBWorklist;
578 BBWorklist.push_back(BB);
581 BasicBlock *Entry = BBWorklist.pop_back_val();
582 // Note that this sets blocks to 0 (unavailable) if they happen to not
583 // already be in FullyAvailableBlocks. This is safe.
584 char &EntryVal = FullyAvailableBlocks[Entry];
585 if (EntryVal == 0) continue; // Already unavailable.
587 // Mark as unavailable.
590 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
591 BBWorklist.push_back(*I);
592 } while (!BBWorklist.empty());
598 /// CanCoerceMustAliasedValueToLoad - Return true if
599 /// CoerceAvailableValueToLoadType will succeed.
600 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
602 const TargetData &TD) {
603 // If the loaded or stored value is an first class array or struct, don't try
604 // to transform them. We need to be able to bitcast to integer.
605 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
606 StoredVal->getType()->isStructTy() ||
607 StoredVal->getType()->isArrayTy())
610 // The store has to be at least as big as the load.
611 if (TD.getTypeSizeInBits(StoredVal->getType()) <
612 TD.getTypeSizeInBits(LoadTy))
619 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
620 /// then a load from a must-aliased pointer of a different type, try to coerce
621 /// the stored value. LoadedTy is the type of the load we want to replace and
622 /// InsertPt is the place to insert new instructions.
624 /// If we can't do it, return null.
625 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
626 const Type *LoadedTy,
627 Instruction *InsertPt,
628 const TargetData &TD) {
629 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
632 const Type *StoredValTy = StoredVal->getType();
634 uint64_t StoreSize = TD.getTypeStoreSizeInBits(StoredValTy);
635 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
637 // If the store and reload are the same size, we can always reuse it.
638 if (StoreSize == LoadSize) {
639 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) {
640 // Pointer to Pointer -> use bitcast.
641 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
644 // Convert source pointers to integers, which can be bitcast.
645 if (StoredValTy->isPointerTy()) {
646 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
647 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
650 const Type *TypeToCastTo = LoadedTy;
651 if (TypeToCastTo->isPointerTy())
652 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
654 if (StoredValTy != TypeToCastTo)
655 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
657 // Cast to pointer if the load needs a pointer type.
658 if (LoadedTy->isPointerTy())
659 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
664 // If the loaded value is smaller than the available value, then we can
665 // extract out a piece from it. If the available value is too small, then we
666 // can't do anything.
667 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
669 // Convert source pointers to integers, which can be manipulated.
670 if (StoredValTy->isPointerTy()) {
671 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
672 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
675 // Convert vectors and fp to integer, which can be manipulated.
676 if (!StoredValTy->isIntegerTy()) {
677 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
678 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
681 // If this is a big-endian system, we need to shift the value down to the low
682 // bits so that a truncate will work.
683 if (TD.isBigEndian()) {
684 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
685 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
688 // Truncate the integer to the right size now.
689 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
690 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
692 if (LoadedTy == NewIntTy)
695 // If the result is a pointer, inttoptr.
696 if (LoadedTy->isPointerTy())
697 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
699 // Otherwise, bitcast.
700 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
703 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
704 /// memdep query of a load that ends up being a clobbering memory write (store,
705 /// memset, memcpy, memmove). This means that the write *may* provide bits used
706 /// by the load but we can't be sure because the pointers don't mustalias.
708 /// Check this case to see if there is anything more we can do before we give
709 /// up. This returns -1 if we have to give up, or a byte number in the stored
710 /// value of the piece that feeds the load.
711 static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr,
713 uint64_t WriteSizeInBits,
714 const TargetData &TD) {
715 // If the loaded or stored value is an first class array or struct, don't try
716 // to transform them. We need to be able to bitcast to integer.
717 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
720 int64_t StoreOffset = 0, LoadOffset = 0;
721 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD);
722 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
723 if (StoreBase != LoadBase)
726 // If the load and store are to the exact same address, they should have been
727 // a must alias. AA must have gotten confused.
728 // FIXME: Study to see if/when this happens. One case is forwarding a memset
729 // to a load from the base of the memset.
731 if (LoadOffset == StoreOffset) {
732 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
733 << "Base = " << *StoreBase << "\n"
734 << "Store Ptr = " << *WritePtr << "\n"
735 << "Store Offs = " << StoreOffset << "\n"
736 << "Load Ptr = " << *LoadPtr << "\n";
741 // If the load and store don't overlap at all, the store doesn't provide
742 // anything to the load. In this case, they really don't alias at all, AA
743 // must have gotten confused.
744 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
746 if ((WriteSizeInBits & 7) | (LoadSize & 7))
748 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
752 bool isAAFailure = false;
753 if (StoreOffset < LoadOffset)
754 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
756 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
760 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
761 << "Base = " << *StoreBase << "\n"
762 << "Store Ptr = " << *WritePtr << "\n"
763 << "Store Offs = " << StoreOffset << "\n"
764 << "Load Ptr = " << *LoadPtr << "\n";
770 // If the Load isn't completely contained within the stored bits, we don't
771 // have all the bits to feed it. We could do something crazy in the future
772 // (issue a smaller load then merge the bits in) but this seems unlikely to be
774 if (StoreOffset > LoadOffset ||
775 StoreOffset+StoreSize < LoadOffset+LoadSize)
778 // Okay, we can do this transformation. Return the number of bytes into the
779 // store that the load is.
780 return LoadOffset-StoreOffset;
783 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
784 /// memdep query of a load that ends up being a clobbering store.
785 static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr,
787 const TargetData &TD) {
788 // Cannot handle reading from store of first-class aggregate yet.
789 if (DepSI->getValueOperand()->getType()->isStructTy() ||
790 DepSI->getValueOperand()->getType()->isArrayTy())
793 Value *StorePtr = DepSI->getPointerOperand();
794 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
795 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
796 StorePtr, StoreSize, TD);
799 static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr,
801 const TargetData &TD) {
802 // If the mem operation is a non-constant size, we can't handle it.
803 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
804 if (SizeCst == 0) return -1;
805 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
807 // If this is memset, we just need to see if the offset is valid in the size
809 if (MI->getIntrinsicID() == Intrinsic::memset)
810 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
813 // If we have a memcpy/memmove, the only case we can handle is if this is a
814 // copy from constant memory. In that case, we can read directly from the
816 MemTransferInst *MTI = cast<MemTransferInst>(MI);
818 Constant *Src = dyn_cast<Constant>(MTI->getSource());
819 if (Src == 0) return -1;
821 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
822 if (GV == 0 || !GV->isConstant()) return -1;
824 // See if the access is within the bounds of the transfer.
825 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
826 MI->getDest(), MemSizeInBits, TD);
830 // Otherwise, see if we can constant fold a load from the constant with the
831 // offset applied as appropriate.
832 Src = ConstantExpr::getBitCast(Src,
833 llvm::Type::getInt8PtrTy(Src->getContext()));
834 Constant *OffsetCst =
835 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
836 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
837 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
838 if (ConstantFoldLoadFromConstPtr(Src, &TD))
844 /// GetStoreValueForLoad - This function is called when we have a
845 /// memdep query of a load that ends up being a clobbering store. This means
846 /// that the store *may* provide bits used by the load but we can't be sure
847 /// because the pointers don't mustalias. Check this case to see if there is
848 /// anything more we can do before we give up.
849 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
851 Instruction *InsertPt, const TargetData &TD){
852 LLVMContext &Ctx = SrcVal->getType()->getContext();
854 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
855 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
857 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
859 // Compute which bits of the stored value are being used by the load. Convert
860 // to an integer type to start with.
861 if (SrcVal->getType()->isPointerTy())
862 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp");
863 if (!SrcVal->getType()->isIntegerTy())
864 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8),
867 // Shift the bits to the least significant depending on endianness.
869 if (TD.isLittleEndian())
872 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
875 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp");
877 if (LoadSize != StoreSize)
878 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8),
881 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
884 /// GetMemInstValueForLoad - This function is called when we have a
885 /// memdep query of a load that ends up being a clobbering mem intrinsic.
886 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
887 const Type *LoadTy, Instruction *InsertPt,
888 const TargetData &TD){
889 LLVMContext &Ctx = LoadTy->getContext();
890 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
892 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
894 // We know that this method is only called when the mem transfer fully
895 // provides the bits for the load.
896 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
897 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
898 // independently of what the offset is.
899 Value *Val = MSI->getValue();
901 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
905 // Splat the value out to the right number of bits.
906 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
907 // If we can double the number of bytes set, do it.
908 if (NumBytesSet*2 <= LoadSize) {
909 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
910 Val = Builder.CreateOr(Val, ShVal);
915 // Otherwise insert one byte at a time.
916 Value *ShVal = Builder.CreateShl(Val, 1*8);
917 Val = Builder.CreateOr(OneElt, ShVal);
921 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
924 // Otherwise, this is a memcpy/memmove from a constant global.
925 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
926 Constant *Src = cast<Constant>(MTI->getSource());
928 // Otherwise, see if we can constant fold a load from the constant with the
929 // offset applied as appropriate.
930 Src = ConstantExpr::getBitCast(Src,
931 llvm::Type::getInt8PtrTy(Src->getContext()));
932 Constant *OffsetCst =
933 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
934 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
935 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
936 return ConstantFoldLoadFromConstPtr(Src, &TD);
941 struct AvailableValueInBlock {
942 /// BB - The basic block in question.
945 SimpleVal, // A simple offsetted value that is accessed.
946 MemIntrin // A memory intrinsic which is loaded from.
949 /// V - The value that is live out of the block.
950 PointerIntPair<Value *, 1, ValType> Val;
952 /// Offset - The byte offset in Val that is interesting for the load query.
955 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
956 unsigned Offset = 0) {
957 AvailableValueInBlock Res;
959 Res.Val.setPointer(V);
960 Res.Val.setInt(SimpleVal);
965 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
966 unsigned Offset = 0) {
967 AvailableValueInBlock Res;
969 Res.Val.setPointer(MI);
970 Res.Val.setInt(MemIntrin);
975 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
976 Value *getSimpleValue() const {
977 assert(isSimpleValue() && "Wrong accessor");
978 return Val.getPointer();
981 MemIntrinsic *getMemIntrinValue() const {
982 assert(!isSimpleValue() && "Wrong accessor");
983 return cast<MemIntrinsic>(Val.getPointer());
986 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
987 /// defined here to the specified type. This handles various coercion cases.
988 Value *MaterializeAdjustedValue(const Type *LoadTy,
989 const TargetData *TD) const {
991 if (isSimpleValue()) {
992 Res = getSimpleValue();
993 if (Res->getType() != LoadTy) {
994 assert(TD && "Need target data to handle type mismatch case");
995 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
998 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
999 << *getSimpleValue() << '\n'
1000 << *Res << '\n' << "\n\n\n");
1003 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1004 LoadTy, BB->getTerminator(), *TD);
1005 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1006 << " " << *getMemIntrinValue() << '\n'
1007 << *Res << '\n' << "\n\n\n");
1015 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1016 /// construct SSA form, allowing us to eliminate LI. This returns the value
1017 /// that should be used at LI's definition site.
1018 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1019 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1020 const TargetData *TD,
1021 const DominatorTree &DT,
1022 AliasAnalysis *AA) {
1023 // Check for the fully redundant, dominating load case. In this case, we can
1024 // just use the dominating value directly.
1025 if (ValuesPerBlock.size() == 1 &&
1026 DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent()))
1027 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD);
1029 // Otherwise, we have to construct SSA form.
1030 SmallVector<PHINode*, 8> NewPHIs;
1031 SSAUpdater SSAUpdate(&NewPHIs);
1032 SSAUpdate.Initialize(LI->getType(), LI->getName());
1034 const Type *LoadTy = LI->getType();
1036 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1037 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1038 BasicBlock *BB = AV.BB;
1040 if (SSAUpdate.HasValueForBlock(BB))
1043 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD));
1046 // Perform PHI construction.
1047 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1049 // If new PHI nodes were created, notify alias analysis.
1050 if (V->getType()->isPointerTy())
1051 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1052 AA->copyValue(LI, NewPHIs[i]);
1054 // Now that we've copied information to the new PHIs, scan through
1055 // them again and inform alias analysis that we've added potentially
1056 // escaping uses to any values that are operands to these PHIs.
1057 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1058 PHINode *P = NewPHIs[i];
1059 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii)
1060 AA->addEscapingUse(P->getOperandUse(2*ii));
1066 static bool isLifetimeStart(const Instruction *Inst) {
1067 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1068 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1072 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1073 /// non-local by performing PHI construction.
1074 bool GVN::processNonLocalLoad(LoadInst *LI,
1075 SmallVectorImpl<Instruction*> &toErase) {
1076 // Find the non-local dependencies of the load.
1077 SmallVector<NonLocalDepResult, 64> Deps;
1078 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1079 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1080 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
1081 // << Deps.size() << *LI << '\n');
1083 // If we had to process more than one hundred blocks to find the
1084 // dependencies, this load isn't worth worrying about. Optimizing
1085 // it will be too expensive.
1086 if (Deps.size() > 100)
1089 // If we had a phi translation failure, we'll have a single entry which is a
1090 // clobber in the current block. Reject this early.
1091 if (Deps.size() == 1 && Deps[0].getResult().isClobber()) {
1093 dbgs() << "GVN: non-local load ";
1094 WriteAsOperand(dbgs(), LI);
1095 dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n';
1100 // Filter out useless results (non-locals, etc). Keep track of the blocks
1101 // where we have a value available in repl, also keep track of whether we see
1102 // dependencies that produce an unknown value for the load (such as a call
1103 // that could potentially clobber the load).
1104 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
1105 SmallVector<BasicBlock*, 16> UnavailableBlocks;
1107 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
1108 BasicBlock *DepBB = Deps[i].getBB();
1109 MemDepResult DepInfo = Deps[i].getResult();
1111 if (DepInfo.isClobber()) {
1112 // The address being loaded in this non-local block may not be the same as
1113 // the pointer operand of the load if PHI translation occurs. Make sure
1114 // to consider the right address.
1115 Value *Address = Deps[i].getAddress();
1117 // If the dependence is to a store that writes to a superset of the bits
1118 // read by the load, we can extract the bits we need for the load from the
1120 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1121 if (TD && Address) {
1122 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1125 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1126 DepSI->getValueOperand(),
1133 // If the clobbering value is a memset/memcpy/memmove, see if we can
1134 // forward a value on from it.
1135 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1136 if (TD && Address) {
1137 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1140 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1147 UnavailableBlocks.push_back(DepBB);
1151 Instruction *DepInst = DepInfo.getInst();
1153 // Loading the allocation -> undef.
1154 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1155 // Loading immediately after lifetime begin -> undef.
1156 isLifetimeStart(DepInst)) {
1157 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1158 UndefValue::get(LI->getType())));
1162 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1163 // Reject loads and stores that are to the same address but are of
1164 // different types if we have to.
1165 if (S->getValueOperand()->getType() != LI->getType()) {
1166 // If the stored value is larger or equal to the loaded value, we can
1168 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1169 LI->getType(), *TD)) {
1170 UnavailableBlocks.push_back(DepBB);
1175 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1176 S->getValueOperand()));
1180 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1181 // If the types mismatch and we can't handle it, reject reuse of the load.
1182 if (LD->getType() != LI->getType()) {
1183 // If the stored value is larger or equal to the loaded value, we can
1185 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1186 UnavailableBlocks.push_back(DepBB);
1190 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD));
1194 UnavailableBlocks.push_back(DepBB);
1198 // If we have no predecessors that produce a known value for this load, exit
1200 if (ValuesPerBlock.empty()) return false;
1202 // If all of the instructions we depend on produce a known value for this
1203 // load, then it is fully redundant and we can use PHI insertion to compute
1204 // its value. Insert PHIs and remove the fully redundant value now.
1205 if (UnavailableBlocks.empty()) {
1206 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1208 // Perform PHI construction.
1209 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1210 VN.getAliasAnalysis());
1211 LI->replaceAllUsesWith(V);
1213 if (isa<PHINode>(V))
1215 if (V->getType()->isPointerTy())
1216 MD->invalidateCachedPointerInfo(V);
1218 toErase.push_back(LI);
1223 if (!EnablePRE || !EnableLoadPRE)
1226 // Okay, we have *some* definitions of the value. This means that the value
1227 // is available in some of our (transitive) predecessors. Lets think about
1228 // doing PRE of this load. This will involve inserting a new load into the
1229 // predecessor when it's not available. We could do this in general, but
1230 // prefer to not increase code size. As such, we only do this when we know
1231 // that we only have to insert *one* load (which means we're basically moving
1232 // the load, not inserting a new one).
1234 SmallPtrSet<BasicBlock *, 4> Blockers;
1235 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1236 Blockers.insert(UnavailableBlocks[i]);
1238 // Lets find first basic block with more than one predecessor. Walk backwards
1239 // through predecessors if needed.
1240 BasicBlock *LoadBB = LI->getParent();
1241 BasicBlock *TmpBB = LoadBB;
1243 bool isSinglePred = false;
1244 bool allSingleSucc = true;
1245 while (TmpBB->getSinglePredecessor()) {
1246 isSinglePred = true;
1247 TmpBB = TmpBB->getSinglePredecessor();
1248 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1250 if (Blockers.count(TmpBB))
1253 // If any of these blocks has more than one successor (i.e. if the edge we
1254 // just traversed was critical), then there are other paths through this
1255 // block along which the load may not be anticipated. Hoisting the load
1256 // above this block would be adding the load to execution paths along
1257 // which it was not previously executed.
1258 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1265 // FIXME: It is extremely unclear what this loop is doing, other than
1266 // artificially restricting loadpre.
1269 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1270 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1271 if (AV.isSimpleValue())
1272 // "Hot" Instruction is in some loop (because it dominates its dep.
1274 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1275 if (DT->dominates(LI, I)) {
1281 // We are interested only in "hot" instructions. We don't want to do any
1282 // mis-optimizations here.
1287 // Check to see how many predecessors have the loaded value fully
1289 DenseMap<BasicBlock*, Value*> PredLoads;
1290 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1291 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1292 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1293 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1294 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1296 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
1297 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1299 BasicBlock *Pred = *PI;
1300 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) {
1303 PredLoads[Pred] = 0;
1305 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1306 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1307 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1308 << Pred->getName() << "': " << *LI << '\n');
1311 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
1312 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
1315 if (!NeedToSplit.empty()) {
1316 toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
1320 // Decide whether PRE is profitable for this load.
1321 unsigned NumUnavailablePreds = PredLoads.size();
1322 assert(NumUnavailablePreds != 0 &&
1323 "Fully available value should be eliminated above!");
1325 // If this load is unavailable in multiple predecessors, reject it.
1326 // FIXME: If we could restructure the CFG, we could make a common pred with
1327 // all the preds that don't have an available LI and insert a new load into
1329 if (NumUnavailablePreds != 1)
1332 // Check if the load can safely be moved to all the unavailable predecessors.
1333 bool CanDoPRE = true;
1334 SmallVector<Instruction*, 8> NewInsts;
1335 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1336 E = PredLoads.end(); I != E; ++I) {
1337 BasicBlock *UnavailablePred = I->first;
1339 // Do PHI translation to get its value in the predecessor if necessary. The
1340 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1342 // If all preds have a single successor, then we know it is safe to insert
1343 // the load on the pred (?!?), so we can insert code to materialize the
1344 // pointer if it is not available.
1345 PHITransAddr Address(LI->getPointerOperand(), TD);
1347 if (allSingleSucc) {
1348 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1351 Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
1352 LoadPtr = Address.getAddr();
1355 // If we couldn't find or insert a computation of this phi translated value,
1358 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1359 << *LI->getPointerOperand() << "\n");
1364 // Make sure it is valid to move this load here. We have to watch out for:
1365 // @1 = getelementptr (i8* p, ...
1366 // test p and branch if == 0
1368 // It is valid to have the getelementptr before the test, even if p can
1369 // be 0, as getelementptr only does address arithmetic.
1370 // If we are not pushing the value through any multiple-successor blocks
1371 // we do not have this case. Otherwise, check that the load is safe to
1372 // put anywhere; this can be improved, but should be conservatively safe.
1373 if (!allSingleSucc &&
1374 // FIXME: REEVALUTE THIS.
1375 !isSafeToLoadUnconditionally(LoadPtr,
1376 UnavailablePred->getTerminator(),
1377 LI->getAlignment(), TD)) {
1382 I->second = LoadPtr;
1386 while (!NewInsts.empty()) {
1387 Instruction *I = NewInsts.pop_back_val();
1388 if (MD) MD->removeInstruction(I);
1389 I->eraseFromParent();
1394 // Okay, we can eliminate this load by inserting a reload in the predecessor
1395 // and using PHI construction to get the value in the other predecessors, do
1397 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1398 DEBUG(if (!NewInsts.empty())
1399 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1400 << *NewInsts.back() << '\n');
1402 // Assign value numbers to the new instructions.
1403 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1404 // FIXME: We really _ought_ to insert these value numbers into their
1405 // parent's availability map. However, in doing so, we risk getting into
1406 // ordering issues. If a block hasn't been processed yet, we would be
1407 // marking a value as AVAIL-IN, which isn't what we intend.
1408 VN.lookup_or_add(NewInsts[i]);
1411 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1412 E = PredLoads.end(); I != E; ++I) {
1413 BasicBlock *UnavailablePred = I->first;
1414 Value *LoadPtr = I->second;
1416 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1418 UnavailablePred->getTerminator());
1420 // Transfer the old load's TBAA tag to the new load.
1421 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1422 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1424 // Add the newly created load.
1425 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1427 MD->invalidateCachedPointerInfo(LoadPtr);
1428 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1431 // Perform PHI construction.
1432 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1433 VN.getAliasAnalysis());
1434 LI->replaceAllUsesWith(V);
1435 if (isa<PHINode>(V))
1437 if (V->getType()->isPointerTy())
1438 MD->invalidateCachedPointerInfo(V);
1440 toErase.push_back(LI);
1445 /// processLoad - Attempt to eliminate a load, first by eliminating it
1446 /// locally, and then attempting non-local elimination if that fails.
1447 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1451 if (L->isVolatile())
1454 // ... to a pointer that has been loaded from before...
1455 MemDepResult Dep = MD->getDependency(L);
1457 // If the value isn't available, don't do anything!
1458 if (Dep.isClobber()) {
1459 // Check to see if we have something like this:
1460 // store i32 123, i32* %P
1461 // %A = bitcast i32* %P to i8*
1462 // %B = gep i8* %A, i32 1
1465 // We could do that by recognizing if the clobber instructions are obviously
1466 // a common base + constant offset, and if the previous store (or memset)
1467 // completely covers this load. This sort of thing can happen in bitfield
1469 Value *AvailVal = 0;
1470 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst()))
1472 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1473 L->getPointerOperand(),
1476 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1477 L->getType(), L, *TD);
1480 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1481 // a value on from it.
1482 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1484 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1485 L->getPointerOperand(),
1488 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L,*TD);
1493 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1494 << *AvailVal << '\n' << *L << "\n\n\n");
1496 // Replace the load!
1497 L->replaceAllUsesWith(AvailVal);
1498 if (AvailVal->getType()->isPointerTy())
1499 MD->invalidateCachedPointerInfo(AvailVal);
1501 toErase.push_back(L);
1507 // fast print dep, using operator<< on instruction would be too slow
1508 dbgs() << "GVN: load ";
1509 WriteAsOperand(dbgs(), L);
1510 Instruction *I = Dep.getInst();
1511 dbgs() << " is clobbered by " << *I << '\n';
1516 // If it is defined in another block, try harder.
1517 if (Dep.isNonLocal())
1518 return processNonLocalLoad(L, toErase);
1520 Instruction *DepInst = Dep.getInst();
1521 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1522 Value *StoredVal = DepSI->getValueOperand();
1524 // The store and load are to a must-aliased pointer, but they may not
1525 // actually have the same type. See if we know how to reuse the stored
1526 // value (depending on its type).
1527 if (StoredVal->getType() != L->getType()) {
1529 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1534 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1535 << '\n' << *L << "\n\n\n");
1542 L->replaceAllUsesWith(StoredVal);
1543 if (StoredVal->getType()->isPointerTy())
1544 MD->invalidateCachedPointerInfo(StoredVal);
1546 toErase.push_back(L);
1551 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1552 Value *AvailableVal = DepLI;
1554 // The loads are of a must-aliased pointer, but they may not actually have
1555 // the same type. See if we know how to reuse the previously loaded value
1556 // (depending on its type).
1557 if (DepLI->getType() != L->getType()) {
1559 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L,*TD);
1560 if (AvailableVal == 0)
1563 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1564 << "\n" << *L << "\n\n\n");
1571 L->replaceAllUsesWith(AvailableVal);
1572 if (DepLI->getType()->isPointerTy())
1573 MD->invalidateCachedPointerInfo(DepLI);
1575 toErase.push_back(L);
1580 // If this load really doesn't depend on anything, then we must be loading an
1581 // undef value. This can happen when loading for a fresh allocation with no
1582 // intervening stores, for example.
1583 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1584 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1586 toErase.push_back(L);
1591 // If this load occurs either right after a lifetime begin,
1592 // then the loaded value is undefined.
1593 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
1594 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1595 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1597 toErase.push_back(L);
1606 // findLeader - In order to find a leader for a given value number at a
1607 // specific basic block, we first obtain the list of all Values for that number,
1608 // and then scan the list to find one whose block dominates the block in
1609 // question. This is fast because dominator tree queries consist of only
1610 // a few comparisons of DFS numbers.
1611 Value *GVN::findLeader(BasicBlock *BB, uint32_t num) {
1612 LeaderTableEntry Vals = LeaderTable[num];
1613 if (!Vals.Val) return 0;
1616 if (DT->dominates(Vals.BB, BB)) {
1618 if (isa<Constant>(Val)) return Val;
1621 LeaderTableEntry* Next = Vals.Next;
1623 if (DT->dominates(Next->BB, BB)) {
1624 if (isa<Constant>(Next->Val)) return Next->Val;
1625 if (!Val) Val = Next->Val;
1635 /// processInstruction - When calculating availability, handle an instruction
1636 /// by inserting it into the appropriate sets
1637 bool GVN::processInstruction(Instruction *I,
1638 SmallVectorImpl<Instruction*> &toErase) {
1639 // Ignore dbg info intrinsics.
1640 if (isa<DbgInfoIntrinsic>(I))
1643 // If the instruction can be easily simplified then do so now in preference
1644 // to value numbering it. Value numbering often exposes redundancies, for
1645 // example if it determines that %y is equal to %x then the instruction
1646 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1647 if (Value *V = SimplifyInstruction(I, TD, DT)) {
1648 I->replaceAllUsesWith(V);
1649 if (MD && V->getType()->isPointerTy())
1650 MD->invalidateCachedPointerInfo(V);
1652 toErase.push_back(I);
1656 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1657 bool Changed = processLoad(LI, toErase);
1660 unsigned Num = VN.lookup_or_add(LI);
1661 addToLeaderTable(Num, LI, LI->getParent());
1667 // For conditions branches, we can perform simple conditional propagation on
1668 // the condition value itself.
1669 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1670 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1673 Value *BranchCond = BI->getCondition();
1674 uint32_t CondVN = VN.lookup_or_add(BranchCond);
1676 BasicBlock *TrueSucc = BI->getSuccessor(0);
1677 BasicBlock *FalseSucc = BI->getSuccessor(1);
1679 if (TrueSucc->getSinglePredecessor())
1680 addToLeaderTable(CondVN,
1681 ConstantInt::getTrue(TrueSucc->getContext()),
1683 if (FalseSucc->getSinglePredecessor())
1684 addToLeaderTable(CondVN,
1685 ConstantInt::getFalse(TrueSucc->getContext()),
1691 // Instructions with void type don't return a value, so there's
1692 // no point in trying to find redudancies in them.
1693 if (I->getType()->isVoidTy()) return false;
1695 uint32_t NextNum = VN.getNextUnusedValueNumber();
1696 unsigned Num = VN.lookup_or_add(I);
1698 // Allocations are always uniquely numbered, so we can save time and memory
1699 // by fast failing them.
1700 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
1701 addToLeaderTable(Num, I, I->getParent());
1705 // If the number we were assigned was a brand new VN, then we don't
1706 // need to do a lookup to see if the number already exists
1707 // somewhere in the domtree: it can't!
1708 if (Num == NextNum) {
1709 addToLeaderTable(Num, I, I->getParent());
1713 // Perform fast-path value-number based elimination of values inherited from
1715 Value *repl = findLeader(I->getParent(), Num);
1717 // Failure, just remember this instance for future use.
1718 addToLeaderTable(Num, I, I->getParent());
1724 I->replaceAllUsesWith(repl);
1725 if (MD && repl->getType()->isPointerTy())
1726 MD->invalidateCachedPointerInfo(repl);
1727 toErase.push_back(I);
1731 /// runOnFunction - This is the main transformation entry point for a function.
1732 bool GVN::runOnFunction(Function& F) {
1734 MD = &getAnalysis<MemoryDependenceAnalysis>();
1735 DT = &getAnalysis<DominatorTree>();
1736 TD = getAnalysisIfAvailable<TargetData>();
1737 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1741 bool Changed = false;
1742 bool ShouldContinue = true;
1744 // Merge unconditional branches, allowing PRE to catch more
1745 // optimization opportunities.
1746 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1747 BasicBlock *BB = FI++;
1749 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1750 if (removedBlock) ++NumGVNBlocks;
1752 Changed |= removedBlock;
1755 unsigned Iteration = 0;
1756 while (ShouldContinue) {
1757 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
1758 ShouldContinue = iterateOnFunction(F);
1759 if (splitCriticalEdges())
1760 ShouldContinue = true;
1761 Changed |= ShouldContinue;
1766 bool PREChanged = true;
1767 while (PREChanged) {
1768 PREChanged = performPRE(F);
1769 Changed |= PREChanged;
1772 // FIXME: Should perform GVN again after PRE does something. PRE can move
1773 // computations into blocks where they become fully redundant. Note that
1774 // we can't do this until PRE's critical edge splitting updates memdep.
1775 // Actually, when this happens, we should just fully integrate PRE into GVN.
1777 cleanupGlobalSets();
1783 bool GVN::processBlock(BasicBlock *BB) {
1784 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
1785 // incrementing BI before processing an instruction).
1786 SmallVector<Instruction*, 8> toErase;
1787 bool ChangedFunction = false;
1789 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1791 ChangedFunction |= processInstruction(BI, toErase);
1792 if (toErase.empty()) {
1797 // If we need some instructions deleted, do it now.
1798 NumGVNInstr += toErase.size();
1800 // Avoid iterator invalidation.
1801 bool AtStart = BI == BB->begin();
1805 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1806 E = toErase.end(); I != E; ++I) {
1807 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
1808 if (MD) MD->removeInstruction(*I);
1809 (*I)->eraseFromParent();
1810 DEBUG(verifyRemoved(*I));
1820 return ChangedFunction;
1823 /// performPRE - Perform a purely local form of PRE that looks for diamond
1824 /// control flow patterns and attempts to perform simple PRE at the join point.
1825 bool GVN::performPRE(Function &F) {
1826 bool Changed = false;
1827 DenseMap<BasicBlock*, Value*> predMap;
1828 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
1829 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
1830 BasicBlock *CurrentBlock = *DI;
1832 // Nothing to PRE in the entry block.
1833 if (CurrentBlock == &F.getEntryBlock()) continue;
1835 for (BasicBlock::iterator BI = CurrentBlock->begin(),
1836 BE = CurrentBlock->end(); BI != BE; ) {
1837 Instruction *CurInst = BI++;
1839 if (isa<AllocaInst>(CurInst) ||
1840 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
1841 CurInst->getType()->isVoidTy() ||
1842 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
1843 isa<DbgInfoIntrinsic>(CurInst))
1846 // We don't currently value number ANY inline asm calls.
1847 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
1848 if (CallI->isInlineAsm())
1851 uint32_t ValNo = VN.lookup(CurInst);
1853 // Look for the predecessors for PRE opportunities. We're
1854 // only trying to solve the basic diamond case, where
1855 // a value is computed in the successor and one predecessor,
1856 // but not the other. We also explicitly disallow cases
1857 // where the successor is its own predecessor, because they're
1858 // more complicated to get right.
1859 unsigned NumWith = 0;
1860 unsigned NumWithout = 0;
1861 BasicBlock *PREPred = 0;
1864 for (pred_iterator PI = pred_begin(CurrentBlock),
1865 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
1866 BasicBlock *P = *PI;
1867 // We're not interested in PRE where the block is its
1868 // own predecessor, or in blocks with predecessors
1869 // that are not reachable.
1870 if (P == CurrentBlock) {
1873 } else if (!DT->dominates(&F.getEntryBlock(), P)) {
1878 Value* predV = findLeader(P, ValNo);
1882 } else if (predV == CurInst) {
1890 // Don't do PRE when it might increase code size, i.e. when
1891 // we would need to insert instructions in more than one pred.
1892 if (NumWithout != 1 || NumWith == 0)
1895 // Don't do PRE across indirect branch.
1896 if (isa<IndirectBrInst>(PREPred->getTerminator()))
1899 // We can't do PRE safely on a critical edge, so instead we schedule
1900 // the edge to be split and perform the PRE the next time we iterate
1902 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
1903 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
1904 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
1908 // Instantiate the expression in the predecessor that lacked it.
1909 // Because we are going top-down through the block, all value numbers
1910 // will be available in the predecessor by the time we need them. Any
1911 // that weren't originally present will have been instantiated earlier
1913 Instruction *PREInstr = CurInst->clone();
1914 bool success = true;
1915 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
1916 Value *Op = PREInstr->getOperand(i);
1917 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
1920 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
1921 PREInstr->setOperand(i, V);
1928 // Fail out if we encounter an operand that is not available in
1929 // the PRE predecessor. This is typically because of loads which
1930 // are not value numbered precisely.
1933 DEBUG(verifyRemoved(PREInstr));
1937 PREInstr->insertBefore(PREPred->getTerminator());
1938 PREInstr->setName(CurInst->getName() + ".pre");
1939 predMap[PREPred] = PREInstr;
1940 VN.add(PREInstr, ValNo);
1943 // Update the availability map to include the new instruction.
1944 addToLeaderTable(ValNo, PREInstr, PREPred);
1946 // Create a PHI to make the value available in this block.
1947 PHINode* Phi = PHINode::Create(CurInst->getType(),
1948 CurInst->getName() + ".pre-phi",
1949 CurrentBlock->begin());
1950 for (pred_iterator PI = pred_begin(CurrentBlock),
1951 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
1952 BasicBlock *P = *PI;
1953 Phi->addIncoming(predMap[P], P);
1957 addToLeaderTable(ValNo, Phi, CurrentBlock);
1959 CurInst->replaceAllUsesWith(Phi);
1960 if (Phi->getType()->isPointerTy()) {
1961 // Because we have added a PHI-use of the pointer value, it has now
1962 // "escaped" from alias analysis' perspective. We need to inform
1964 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii)
1965 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(2*ii));
1968 MD->invalidateCachedPointerInfo(Phi);
1971 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
1973 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
1974 if (MD) MD->removeInstruction(CurInst);
1975 CurInst->eraseFromParent();
1976 DEBUG(verifyRemoved(CurInst));
1981 if (splitCriticalEdges())
1987 /// splitCriticalEdges - Split critical edges found during the previous
1988 /// iteration that may enable further optimization.
1989 bool GVN::splitCriticalEdges() {
1990 if (toSplit.empty())
1993 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
1994 SplitCriticalEdge(Edge.first, Edge.second, this);
1995 } while (!toSplit.empty());
1996 if (MD) MD->invalidateCachedPredecessors();
2000 /// iterateOnFunction - Executes one iteration of GVN
2001 bool GVN::iterateOnFunction(Function &F) {
2002 cleanupGlobalSets();
2004 // Top-down walk of the dominator tree
2005 bool Changed = false;
2007 // Needed for value numbering with phi construction to work.
2008 ReversePostOrderTraversal<Function*> RPOT(&F);
2009 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2010 RE = RPOT.end(); RI != RE; ++RI)
2011 Changed |= processBlock(*RI);
2013 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2014 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2015 Changed |= processBlock(DI->getBlock());
2021 void GVN::cleanupGlobalSets() {
2023 LeaderTable.clear();
2024 TableAllocator.Reset();
2027 /// verifyRemoved - Verify that the specified instruction does not occur in our
2028 /// internal data structures.
2029 void GVN::verifyRemoved(const Instruction *Inst) const {
2030 VN.verifyRemoved(Inst);
2032 // Walk through the value number scope to make sure the instruction isn't
2033 // ferreted away in it.
2034 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2035 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2036 const LeaderTableEntry *Node = &I->second;
2037 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2039 while (Node->Next) {
2041 assert(Node->Val != Inst && "Inst still in value numbering scope!");