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 // Pointer to Pointer -> use bitcast.
640 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy())
641 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
643 // Convert source pointers to integers, which can be bitcast.
644 if (StoredValTy->isPointerTy()) {
645 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
646 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
649 const Type *TypeToCastTo = LoadedTy;
650 if (TypeToCastTo->isPointerTy())
651 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
653 if (StoredValTy != TypeToCastTo)
654 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
656 // Cast to pointer if the load needs a pointer type.
657 if (LoadedTy->isPointerTy())
658 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
663 // If the loaded value is smaller than the available value, then we can
664 // extract out a piece from it. If the available value is too small, then we
665 // can't do anything.
666 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
668 // Convert source pointers to integers, which can be manipulated.
669 if (StoredValTy->isPointerTy()) {
670 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
671 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
674 // Convert vectors and fp to integer, which can be manipulated.
675 if (!StoredValTy->isIntegerTy()) {
676 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
677 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
680 // If this is a big-endian system, we need to shift the value down to the low
681 // bits so that a truncate will work.
682 if (TD.isBigEndian()) {
683 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
684 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
687 // Truncate the integer to the right size now.
688 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
689 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
691 if (LoadedTy == NewIntTy)
694 // If the result is a pointer, inttoptr.
695 if (LoadedTy->isPointerTy())
696 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
698 // Otherwise, bitcast.
699 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
702 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
703 /// memdep query of a load that ends up being a clobbering memory write (store,
704 /// memset, memcpy, memmove). This means that the write *may* provide bits used
705 /// by the load but we can't be sure because the pointers don't mustalias.
707 /// Check this case to see if there is anything more we can do before we give
708 /// up. This returns -1 if we have to give up, or a byte number in the stored
709 /// value of the piece that feeds the load.
710 static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr,
712 uint64_t WriteSizeInBits,
713 const TargetData &TD) {
714 // If the loaded or stored value is an first class array or struct, don't try
715 // to transform them. We need to be able to bitcast to integer.
716 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
719 int64_t StoreOffset = 0, LoadOffset = 0;
720 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD);
721 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
722 if (StoreBase != LoadBase)
725 // If the load and store are to the exact same address, they should have been
726 // a must alias. AA must have gotten confused.
727 // FIXME: Study to see if/when this happens. One case is forwarding a memset
728 // to a load from the base of the memset.
730 if (LoadOffset == StoreOffset) {
731 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
732 << "Base = " << *StoreBase << "\n"
733 << "Store Ptr = " << *WritePtr << "\n"
734 << "Store Offs = " << StoreOffset << "\n"
735 << "Load Ptr = " << *LoadPtr << "\n";
740 // If the load and store don't overlap at all, the store doesn't provide
741 // anything to the load. In this case, they really don't alias at all, AA
742 // must have gotten confused.
743 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
745 if ((WriteSizeInBits & 7) | (LoadSize & 7))
747 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
751 bool isAAFailure = false;
752 if (StoreOffset < LoadOffset)
753 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
755 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
759 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
760 << "Base = " << *StoreBase << "\n"
761 << "Store Ptr = " << *WritePtr << "\n"
762 << "Store Offs = " << StoreOffset << "\n"
763 << "Load Ptr = " << *LoadPtr << "\n";
769 // If the Load isn't completely contained within the stored bits, we don't
770 // have all the bits to feed it. We could do something crazy in the future
771 // (issue a smaller load then merge the bits in) but this seems unlikely to be
773 if (StoreOffset > LoadOffset ||
774 StoreOffset+StoreSize < LoadOffset+LoadSize)
777 // Okay, we can do this transformation. Return the number of bytes into the
778 // store that the load is.
779 return LoadOffset-StoreOffset;
782 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
783 /// memdep query of a load that ends up being a clobbering store.
784 static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr,
786 const TargetData &TD) {
787 // Cannot handle reading from store of first-class aggregate yet.
788 if (DepSI->getValueOperand()->getType()->isStructTy() ||
789 DepSI->getValueOperand()->getType()->isArrayTy())
792 Value *StorePtr = DepSI->getPointerOperand();
793 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
794 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
795 StorePtr, StoreSize, TD);
798 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
799 /// memdep query of a load that ends up being clobbered by another load. See if
800 /// the other load can feed into the second load.
801 static int AnalyzeLoadFromClobberingLoad(const Type *LoadTy, Value *LoadPtr,
802 LoadInst *DepLI, const TargetData &TD){
803 // Cannot handle reading from store of first-class aggregate yet.
804 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
807 Value *DepPtr = DepLI->getPointerOperand();
808 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType());
809 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD);
814 static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr,
816 const TargetData &TD) {
817 // If the mem operation is a non-constant size, we can't handle it.
818 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
819 if (SizeCst == 0) return -1;
820 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
822 // If this is memset, we just need to see if the offset is valid in the size
824 if (MI->getIntrinsicID() == Intrinsic::memset)
825 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
828 // If we have a memcpy/memmove, the only case we can handle is if this is a
829 // copy from constant memory. In that case, we can read directly from the
831 MemTransferInst *MTI = cast<MemTransferInst>(MI);
833 Constant *Src = dyn_cast<Constant>(MTI->getSource());
834 if (Src == 0) return -1;
836 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
837 if (GV == 0 || !GV->isConstant()) return -1;
839 // See if the access is within the bounds of the transfer.
840 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
841 MI->getDest(), MemSizeInBits, TD);
845 // Otherwise, see if we can constant fold a load from the constant with the
846 // offset applied as appropriate.
847 Src = ConstantExpr::getBitCast(Src,
848 llvm::Type::getInt8PtrTy(Src->getContext()));
849 Constant *OffsetCst =
850 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
851 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
852 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
853 if (ConstantFoldLoadFromConstPtr(Src, &TD))
859 /// GetStoreValueForLoad - This function is called when we have a
860 /// memdep query of a load that ends up being a clobbering store. This means
861 /// that the store *may* provide bits used by the load but we can't be sure
862 /// because the pointers don't mustalias. Check this case to see if there is
863 /// anything more we can do before we give up.
864 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
866 Instruction *InsertPt, const TargetData &TD){
867 LLVMContext &Ctx = SrcVal->getType()->getContext();
869 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
870 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
872 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
874 // Compute which bits of the stored value are being used by the load. Convert
875 // to an integer type to start with.
876 if (SrcVal->getType()->isPointerTy())
877 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp");
878 if (!SrcVal->getType()->isIntegerTy())
879 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8),
882 // Shift the bits to the least significant depending on endianness.
884 if (TD.isLittleEndian())
887 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
890 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp");
892 if (LoadSize != StoreSize)
893 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8),
896 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
899 /// GetMemInstValueForLoad - This function is called when we have a
900 /// memdep query of a load that ends up being a clobbering mem intrinsic.
901 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
902 const Type *LoadTy, Instruction *InsertPt,
903 const TargetData &TD){
904 LLVMContext &Ctx = LoadTy->getContext();
905 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
907 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
909 // We know that this method is only called when the mem transfer fully
910 // provides the bits for the load.
911 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
912 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
913 // independently of what the offset is.
914 Value *Val = MSI->getValue();
916 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
920 // Splat the value out to the right number of bits.
921 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
922 // If we can double the number of bytes set, do it.
923 if (NumBytesSet*2 <= LoadSize) {
924 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
925 Val = Builder.CreateOr(Val, ShVal);
930 // Otherwise insert one byte at a time.
931 Value *ShVal = Builder.CreateShl(Val, 1*8);
932 Val = Builder.CreateOr(OneElt, ShVal);
936 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
939 // Otherwise, this is a memcpy/memmove from a constant global.
940 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
941 Constant *Src = cast<Constant>(MTI->getSource());
943 // Otherwise, see if we can constant fold a load from the constant with the
944 // offset applied as appropriate.
945 Src = ConstantExpr::getBitCast(Src,
946 llvm::Type::getInt8PtrTy(Src->getContext()));
947 Constant *OffsetCst =
948 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
949 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1);
950 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
951 return ConstantFoldLoadFromConstPtr(Src, &TD);
956 struct AvailableValueInBlock {
957 /// BB - The basic block in question.
960 SimpleVal, // A simple offsetted value that is accessed.
961 MemIntrin // A memory intrinsic which is loaded from.
964 /// V - The value that is live out of the block.
965 PointerIntPair<Value *, 1, ValType> Val;
967 /// Offset - The byte offset in Val that is interesting for the load query.
970 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
971 unsigned Offset = 0) {
972 AvailableValueInBlock Res;
974 Res.Val.setPointer(V);
975 Res.Val.setInt(SimpleVal);
980 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
981 unsigned Offset = 0) {
982 AvailableValueInBlock Res;
984 Res.Val.setPointer(MI);
985 Res.Val.setInt(MemIntrin);
990 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
991 Value *getSimpleValue() const {
992 assert(isSimpleValue() && "Wrong accessor");
993 return Val.getPointer();
996 MemIntrinsic *getMemIntrinValue() const {
997 assert(!isSimpleValue() && "Wrong accessor");
998 return cast<MemIntrinsic>(Val.getPointer());
1001 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
1002 /// defined here to the specified type. This handles various coercion cases.
1003 Value *MaterializeAdjustedValue(const Type *LoadTy,
1004 const TargetData *TD) const {
1006 if (isSimpleValue()) {
1007 Res = getSimpleValue();
1008 if (Res->getType() != LoadTy) {
1009 assert(TD && "Need target data to handle type mismatch case");
1010 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1013 DEBUG(errs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1014 << *getSimpleValue() << '\n'
1015 << *Res << '\n' << "\n\n\n");
1018 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1019 LoadTy, BB->getTerminator(), *TD);
1020 DEBUG(errs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1021 << " " << *getMemIntrinValue() << '\n'
1022 << *Res << '\n' << "\n\n\n");
1030 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1031 /// construct SSA form, allowing us to eliminate LI. This returns the value
1032 /// that should be used at LI's definition site.
1033 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1034 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1035 const TargetData *TD,
1036 const DominatorTree &DT,
1037 AliasAnalysis *AA) {
1038 // Check for the fully redundant, dominating load case. In this case, we can
1039 // just use the dominating value directly.
1040 if (ValuesPerBlock.size() == 1 &&
1041 DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent()))
1042 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD);
1044 // Otherwise, we have to construct SSA form.
1045 SmallVector<PHINode*, 8> NewPHIs;
1046 SSAUpdater SSAUpdate(&NewPHIs);
1047 SSAUpdate.Initialize(LI->getType(), LI->getName());
1049 const Type *LoadTy = LI->getType();
1051 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1052 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1053 BasicBlock *BB = AV.BB;
1055 if (SSAUpdate.HasValueForBlock(BB))
1058 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD));
1061 // Perform PHI construction.
1062 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1064 // If new PHI nodes were created, notify alias analysis.
1065 if (V->getType()->isPointerTy())
1066 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1067 AA->copyValue(LI, NewPHIs[i]);
1069 // Now that we've copied information to the new PHIs, scan through
1070 // them again and inform alias analysis that we've added potentially
1071 // escaping uses to any values that are operands to these PHIs.
1072 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1073 PHINode *P = NewPHIs[i];
1074 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii)
1075 AA->addEscapingUse(P->getOperandUse(2*ii));
1081 static bool isLifetimeStart(const Instruction *Inst) {
1082 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1083 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1087 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1088 /// non-local by performing PHI construction.
1089 bool GVN::processNonLocalLoad(LoadInst *LI,
1090 SmallVectorImpl<Instruction*> &toErase) {
1091 // Find the non-local dependencies of the load.
1092 SmallVector<NonLocalDepResult, 64> Deps;
1093 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1094 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1095 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
1096 // << Deps.size() << *LI << '\n');
1098 // If we had to process more than one hundred blocks to find the
1099 // dependencies, this load isn't worth worrying about. Optimizing
1100 // it will be too expensive.
1101 if (Deps.size() > 100)
1104 // If we had a phi translation failure, we'll have a single entry which is a
1105 // clobber in the current block. Reject this early.
1106 if (Deps.size() == 1 && Deps[0].getResult().isClobber() &&
1107 Deps[0].getResult().getInst()->getParent() == LI->getParent()) {
1109 dbgs() << "GVN: non-local load ";
1110 WriteAsOperand(dbgs(), LI);
1111 dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n';
1116 // Filter out useless results (non-locals, etc). Keep track of the blocks
1117 // where we have a value available in repl, also keep track of whether we see
1118 // dependencies that produce an unknown value for the load (such as a call
1119 // that could potentially clobber the load).
1120 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
1121 SmallVector<BasicBlock*, 16> UnavailableBlocks;
1123 for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
1124 BasicBlock *DepBB = Deps[i].getBB();
1125 MemDepResult DepInfo = Deps[i].getResult();
1127 if (DepInfo.isClobber()) {
1128 // The address being loaded in this non-local block may not be the same as
1129 // the pointer operand of the load if PHI translation occurs. Make sure
1130 // to consider the right address.
1131 Value *Address = Deps[i].getAddress();
1133 // If the dependence is to a store that writes to a superset of the bits
1134 // read by the load, we can extract the bits we need for the load from the
1136 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1137 if (TD && Address) {
1138 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1141 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1142 DepSI->getValueOperand(),
1149 // Check to see if we have something like this:
1152 // if we have this, replace the later with an extraction from the former.
1153 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1154 // If this is a clobber and L is the first instruction in its block, then
1155 // we have the first instruction in the entry block.
1156 if (DepLI != LI && Address && TD) {
1157 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
1158 LI->getPointerOperand(),
1162 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, DepLI,
1169 // If the clobbering value is a memset/memcpy/memmove, see if we can
1170 // forward a value on from it.
1171 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1172 if (TD && Address) {
1173 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1176 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1183 UnavailableBlocks.push_back(DepBB);
1187 Instruction *DepInst = DepInfo.getInst();
1189 // Loading the allocation -> undef.
1190 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1191 // Loading immediately after lifetime begin -> undef.
1192 isLifetimeStart(DepInst)) {
1193 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1194 UndefValue::get(LI->getType())));
1198 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1199 // Reject loads and stores that are to the same address but are of
1200 // different types if we have to.
1201 if (S->getValueOperand()->getType() != LI->getType()) {
1202 // If the stored value is larger or equal to the loaded value, we can
1204 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1205 LI->getType(), *TD)) {
1206 UnavailableBlocks.push_back(DepBB);
1211 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1212 S->getValueOperand()));
1216 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1217 // If the types mismatch and we can't handle it, reject reuse of the load.
1218 if (LD->getType() != LI->getType()) {
1219 // If the stored value is larger or equal to the loaded value, we can
1221 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1222 UnavailableBlocks.push_back(DepBB);
1226 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, LD));
1230 UnavailableBlocks.push_back(DepBB);
1234 // If we have no predecessors that produce a known value for this load, exit
1236 if (ValuesPerBlock.empty()) return false;
1238 // If all of the instructions we depend on produce a known value for this
1239 // load, then it is fully redundant and we can use PHI insertion to compute
1240 // its value. Insert PHIs and remove the fully redundant value now.
1241 if (UnavailableBlocks.empty()) {
1242 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1244 // Perform PHI construction.
1245 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1246 VN.getAliasAnalysis());
1247 LI->replaceAllUsesWith(V);
1249 if (isa<PHINode>(V))
1251 if (V->getType()->isPointerTy())
1252 MD->invalidateCachedPointerInfo(V);
1254 toErase.push_back(LI);
1259 if (!EnablePRE || !EnableLoadPRE)
1262 // Okay, we have *some* definitions of the value. This means that the value
1263 // is available in some of our (transitive) predecessors. Lets think about
1264 // doing PRE of this load. This will involve inserting a new load into the
1265 // predecessor when it's not available. We could do this in general, but
1266 // prefer to not increase code size. As such, we only do this when we know
1267 // that we only have to insert *one* load (which means we're basically moving
1268 // the load, not inserting a new one).
1270 SmallPtrSet<BasicBlock *, 4> Blockers;
1271 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1272 Blockers.insert(UnavailableBlocks[i]);
1274 // Lets find first basic block with more than one predecessor. Walk backwards
1275 // through predecessors if needed.
1276 BasicBlock *LoadBB = LI->getParent();
1277 BasicBlock *TmpBB = LoadBB;
1279 bool isSinglePred = false;
1280 bool allSingleSucc = true;
1281 while (TmpBB->getSinglePredecessor()) {
1282 isSinglePred = true;
1283 TmpBB = TmpBB->getSinglePredecessor();
1284 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1286 if (Blockers.count(TmpBB))
1289 // If any of these blocks has more than one successor (i.e. if the edge we
1290 // just traversed was critical), then there are other paths through this
1291 // block along which the load may not be anticipated. Hoisting the load
1292 // above this block would be adding the load to execution paths along
1293 // which it was not previously executed.
1294 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1301 // FIXME: It is extremely unclear what this loop is doing, other than
1302 // artificially restricting loadpre.
1305 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1306 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1307 if (AV.isSimpleValue())
1308 // "Hot" Instruction is in some loop (because it dominates its dep.
1310 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1311 if (DT->dominates(LI, I)) {
1317 // We are interested only in "hot" instructions. We don't want to do any
1318 // mis-optimizations here.
1323 // Check to see how many predecessors have the loaded value fully
1325 DenseMap<BasicBlock*, Value*> PredLoads;
1326 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1327 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1328 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1329 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1330 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1332 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
1333 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1335 BasicBlock *Pred = *PI;
1336 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) {
1339 PredLoads[Pred] = 0;
1341 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1342 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1343 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1344 << Pred->getName() << "': " << *LI << '\n');
1347 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
1348 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
1351 if (!NeedToSplit.empty()) {
1352 toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
1356 // Decide whether PRE is profitable for this load.
1357 unsigned NumUnavailablePreds = PredLoads.size();
1358 assert(NumUnavailablePreds != 0 &&
1359 "Fully available value should be eliminated above!");
1361 // If this load is unavailable in multiple predecessors, reject it.
1362 // FIXME: If we could restructure the CFG, we could make a common pred with
1363 // all the preds that don't have an available LI and insert a new load into
1365 if (NumUnavailablePreds != 1)
1368 // Check if the load can safely be moved to all the unavailable predecessors.
1369 bool CanDoPRE = true;
1370 SmallVector<Instruction*, 8> NewInsts;
1371 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1372 E = PredLoads.end(); I != E; ++I) {
1373 BasicBlock *UnavailablePred = I->first;
1375 // Do PHI translation to get its value in the predecessor if necessary. The
1376 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1378 // If all preds have a single successor, then we know it is safe to insert
1379 // the load on the pred (?!?), so we can insert code to materialize the
1380 // pointer if it is not available.
1381 PHITransAddr Address(LI->getPointerOperand(), TD);
1383 if (allSingleSucc) {
1384 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1387 Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
1388 LoadPtr = Address.getAddr();
1391 // If we couldn't find or insert a computation of this phi translated value,
1394 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1395 << *LI->getPointerOperand() << "\n");
1400 // Make sure it is valid to move this load here. We have to watch out for:
1401 // @1 = getelementptr (i8* p, ...
1402 // test p and branch if == 0
1404 // It is valid to have the getelementptr before the test, even if p can
1405 // be 0, as getelementptr only does address arithmetic.
1406 // If we are not pushing the value through any multiple-successor blocks
1407 // we do not have this case. Otherwise, check that the load is safe to
1408 // put anywhere; this can be improved, but should be conservatively safe.
1409 if (!allSingleSucc &&
1410 // FIXME: REEVALUTE THIS.
1411 !isSafeToLoadUnconditionally(LoadPtr,
1412 UnavailablePred->getTerminator(),
1413 LI->getAlignment(), TD)) {
1418 I->second = LoadPtr;
1422 while (!NewInsts.empty()) {
1423 Instruction *I = NewInsts.pop_back_val();
1424 if (MD) MD->removeInstruction(I);
1425 I->eraseFromParent();
1430 // Okay, we can eliminate this load by inserting a reload in the predecessor
1431 // and using PHI construction to get the value in the other predecessors, do
1433 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1434 DEBUG(if (!NewInsts.empty())
1435 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1436 << *NewInsts.back() << '\n');
1438 // Assign value numbers to the new instructions.
1439 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1440 // FIXME: We really _ought_ to insert these value numbers into their
1441 // parent's availability map. However, in doing so, we risk getting into
1442 // ordering issues. If a block hasn't been processed yet, we would be
1443 // marking a value as AVAIL-IN, which isn't what we intend.
1444 VN.lookup_or_add(NewInsts[i]);
1447 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1448 E = PredLoads.end(); I != E; ++I) {
1449 BasicBlock *UnavailablePred = I->first;
1450 Value *LoadPtr = I->second;
1452 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1454 UnavailablePred->getTerminator());
1456 // Transfer the old load's TBAA tag to the new load.
1457 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1458 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1460 // Add the newly created load.
1461 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1463 MD->invalidateCachedPointerInfo(LoadPtr);
1464 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1467 // Perform PHI construction.
1468 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT,
1469 VN.getAliasAnalysis());
1470 LI->replaceAllUsesWith(V);
1471 if (isa<PHINode>(V))
1473 if (V->getType()->isPointerTy())
1474 MD->invalidateCachedPointerInfo(V);
1476 toErase.push_back(LI);
1481 /// processLoad - Attempt to eliminate a load, first by eliminating it
1482 /// locally, and then attempting non-local elimination if that fails.
1483 bool GVN::processLoad(LoadInst *L, SmallVectorImpl<Instruction*> &toErase) {
1487 if (L->isVolatile())
1490 // ... to a pointer that has been loaded from before...
1491 MemDepResult Dep = MD->getDependency(L);
1493 // If we have a clobber and target data is around, see if this is a clobber
1494 // that we can fix up through code synthesis.
1495 if (Dep.isClobber() && TD) {
1496 // Check to see if we have something like this:
1497 // store i32 123, i32* %P
1498 // %A = bitcast i32* %P to i8*
1499 // %B = gep i8* %A, i32 1
1502 // We could do that by recognizing if the clobber instructions are obviously
1503 // a common base + constant offset, and if the previous store (or memset)
1504 // completely covers this load. This sort of thing can happen in bitfield
1506 Value *AvailVal = 0;
1507 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1508 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1509 L->getPointerOperand(),
1512 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1513 L->getType(), L, *TD);
1516 // Check to see if we have something like this:
1519 // if we have this, replace the later with an extraction from the former.
1520 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1521 // If this is a clobber and L is the first instruction in its block, then
1522 // we have the first instruction in the entry block.
1526 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1527 L->getPointerOperand(),
1530 AvailVal = GetStoreValueForLoad(DepLI, Offset, L->getType(), L, *TD);
1533 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1534 // a value on from it.
1535 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1536 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1537 L->getPointerOperand(),
1540 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
1544 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1545 << *AvailVal << '\n' << *L << "\n\n\n");
1547 // Replace the load!
1548 L->replaceAllUsesWith(AvailVal);
1549 if (AvailVal->getType()->isPointerTy())
1550 MD->invalidateCachedPointerInfo(AvailVal);
1552 toErase.push_back(L);
1558 // If the value isn't available, don't do anything!
1559 if (Dep.isClobber()) {
1561 // fast print dep, using operator<< on instruction is too slow.
1562 dbgs() << "GVN: load ";
1563 WriteAsOperand(dbgs(), L);
1564 Instruction *I = Dep.getInst();
1565 dbgs() << " is clobbered by " << *I << '\n';
1570 // If it is defined in another block, try harder.
1571 if (Dep.isNonLocal())
1572 return processNonLocalLoad(L, toErase);
1574 Instruction *DepInst = Dep.getInst();
1575 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1576 Value *StoredVal = DepSI->getValueOperand();
1578 // The store and load are to a must-aliased pointer, but they may not
1579 // actually have the same type. See if we know how to reuse the stored
1580 // value (depending on its type).
1581 if (StoredVal->getType() != L->getType()) {
1583 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1588 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1589 << '\n' << *L << "\n\n\n");
1596 L->replaceAllUsesWith(StoredVal);
1597 if (StoredVal->getType()->isPointerTy())
1598 MD->invalidateCachedPointerInfo(StoredVal);
1600 toErase.push_back(L);
1605 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1606 Value *AvailableVal = DepLI;
1608 // The loads are of a must-aliased pointer, but they may not actually have
1609 // the same type. See if we know how to reuse the previously loaded value
1610 // (depending on its type).
1611 if (DepLI->getType() != L->getType()) {
1613 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1615 if (AvailableVal == 0)
1618 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1619 << "\n" << *L << "\n\n\n");
1626 L->replaceAllUsesWith(AvailableVal);
1627 if (DepLI->getType()->isPointerTy())
1628 MD->invalidateCachedPointerInfo(DepLI);
1630 toErase.push_back(L);
1635 // If this load really doesn't depend on anything, then we must be loading an
1636 // undef value. This can happen when loading for a fresh allocation with no
1637 // intervening stores, for example.
1638 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1639 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1641 toErase.push_back(L);
1646 // If this load occurs either right after a lifetime begin,
1647 // then the loaded value is undefined.
1648 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) {
1649 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1650 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1652 toErase.push_back(L);
1661 // findLeader - In order to find a leader for a given value number at a
1662 // specific basic block, we first obtain the list of all Values for that number,
1663 // and then scan the list to find one whose block dominates the block in
1664 // question. This is fast because dominator tree queries consist of only
1665 // a few comparisons of DFS numbers.
1666 Value *GVN::findLeader(BasicBlock *BB, uint32_t num) {
1667 LeaderTableEntry Vals = LeaderTable[num];
1668 if (!Vals.Val) return 0;
1671 if (DT->dominates(Vals.BB, BB)) {
1673 if (isa<Constant>(Val)) return Val;
1676 LeaderTableEntry* Next = Vals.Next;
1678 if (DT->dominates(Next->BB, BB)) {
1679 if (isa<Constant>(Next->Val)) return Next->Val;
1680 if (!Val) Val = Next->Val;
1690 /// processInstruction - When calculating availability, handle an instruction
1691 /// by inserting it into the appropriate sets
1692 bool GVN::processInstruction(Instruction *I,
1693 SmallVectorImpl<Instruction*> &toErase) {
1694 // Ignore dbg info intrinsics.
1695 if (isa<DbgInfoIntrinsic>(I))
1698 // If the instruction can be easily simplified then do so now in preference
1699 // to value numbering it. Value numbering often exposes redundancies, for
1700 // example if it determines that %y is equal to %x then the instruction
1701 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
1702 if (Value *V = SimplifyInstruction(I, TD, DT)) {
1703 I->replaceAllUsesWith(V);
1704 if (MD && V->getType()->isPointerTy())
1705 MD->invalidateCachedPointerInfo(V);
1707 toErase.push_back(I);
1711 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
1712 bool Changed = processLoad(LI, toErase);
1715 unsigned Num = VN.lookup_or_add(LI);
1716 addToLeaderTable(Num, LI, LI->getParent());
1722 // For conditions branches, we can perform simple conditional propagation on
1723 // the condition value itself.
1724 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
1725 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
1728 Value *BranchCond = BI->getCondition();
1729 uint32_t CondVN = VN.lookup_or_add(BranchCond);
1731 BasicBlock *TrueSucc = BI->getSuccessor(0);
1732 BasicBlock *FalseSucc = BI->getSuccessor(1);
1734 if (TrueSucc->getSinglePredecessor())
1735 addToLeaderTable(CondVN,
1736 ConstantInt::getTrue(TrueSucc->getContext()),
1738 if (FalseSucc->getSinglePredecessor())
1739 addToLeaderTable(CondVN,
1740 ConstantInt::getFalse(TrueSucc->getContext()),
1746 // Instructions with void type don't return a value, so there's
1747 // no point in trying to find redudancies in them.
1748 if (I->getType()->isVoidTy()) return false;
1750 uint32_t NextNum = VN.getNextUnusedValueNumber();
1751 unsigned Num = VN.lookup_or_add(I);
1753 // Allocations are always uniquely numbered, so we can save time and memory
1754 // by fast failing them.
1755 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
1756 addToLeaderTable(Num, I, I->getParent());
1760 // If the number we were assigned was a brand new VN, then we don't
1761 // need to do a lookup to see if the number already exists
1762 // somewhere in the domtree: it can't!
1763 if (Num == NextNum) {
1764 addToLeaderTable(Num, I, I->getParent());
1768 // Perform fast-path value-number based elimination of values inherited from
1770 Value *repl = findLeader(I->getParent(), Num);
1772 // Failure, just remember this instance for future use.
1773 addToLeaderTable(Num, I, I->getParent());
1779 I->replaceAllUsesWith(repl);
1780 if (MD && repl->getType()->isPointerTy())
1781 MD->invalidateCachedPointerInfo(repl);
1782 toErase.push_back(I);
1786 /// runOnFunction - This is the main transformation entry point for a function.
1787 bool GVN::runOnFunction(Function& F) {
1789 MD = &getAnalysis<MemoryDependenceAnalysis>();
1790 DT = &getAnalysis<DominatorTree>();
1791 TD = getAnalysisIfAvailable<TargetData>();
1792 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
1796 bool Changed = false;
1797 bool ShouldContinue = true;
1799 // Merge unconditional branches, allowing PRE to catch more
1800 // optimization opportunities.
1801 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
1802 BasicBlock *BB = FI++;
1804 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
1805 if (removedBlock) ++NumGVNBlocks;
1807 Changed |= removedBlock;
1810 unsigned Iteration = 0;
1811 while (ShouldContinue) {
1812 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
1813 ShouldContinue = iterateOnFunction(F);
1814 if (splitCriticalEdges())
1815 ShouldContinue = true;
1816 Changed |= ShouldContinue;
1821 bool PREChanged = true;
1822 while (PREChanged) {
1823 PREChanged = performPRE(F);
1824 Changed |= PREChanged;
1827 // FIXME: Should perform GVN again after PRE does something. PRE can move
1828 // computations into blocks where they become fully redundant. Note that
1829 // we can't do this until PRE's critical edge splitting updates memdep.
1830 // Actually, when this happens, we should just fully integrate PRE into GVN.
1832 cleanupGlobalSets();
1838 bool GVN::processBlock(BasicBlock *BB) {
1839 // FIXME: Kill off toErase by doing erasing eagerly in a helper function (and
1840 // incrementing BI before processing an instruction).
1841 SmallVector<Instruction*, 8> toErase;
1842 bool ChangedFunction = false;
1844 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
1846 ChangedFunction |= processInstruction(BI, toErase);
1847 if (toErase.empty()) {
1852 // If we need some instructions deleted, do it now.
1853 NumGVNInstr += toErase.size();
1855 // Avoid iterator invalidation.
1856 bool AtStart = BI == BB->begin();
1860 for (SmallVector<Instruction*, 4>::iterator I = toErase.begin(),
1861 E = toErase.end(); I != E; ++I) {
1862 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
1863 if (MD) MD->removeInstruction(*I);
1864 (*I)->eraseFromParent();
1865 DEBUG(verifyRemoved(*I));
1875 return ChangedFunction;
1878 /// performPRE - Perform a purely local form of PRE that looks for diamond
1879 /// control flow patterns and attempts to perform simple PRE at the join point.
1880 bool GVN::performPRE(Function &F) {
1881 bool Changed = false;
1882 DenseMap<BasicBlock*, Value*> predMap;
1883 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
1884 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
1885 BasicBlock *CurrentBlock = *DI;
1887 // Nothing to PRE in the entry block.
1888 if (CurrentBlock == &F.getEntryBlock()) continue;
1890 for (BasicBlock::iterator BI = CurrentBlock->begin(),
1891 BE = CurrentBlock->end(); BI != BE; ) {
1892 Instruction *CurInst = BI++;
1894 if (isa<AllocaInst>(CurInst) ||
1895 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
1896 CurInst->getType()->isVoidTy() ||
1897 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
1898 isa<DbgInfoIntrinsic>(CurInst))
1901 // We don't currently value number ANY inline asm calls.
1902 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
1903 if (CallI->isInlineAsm())
1906 uint32_t ValNo = VN.lookup(CurInst);
1908 // Look for the predecessors for PRE opportunities. We're
1909 // only trying to solve the basic diamond case, where
1910 // a value is computed in the successor and one predecessor,
1911 // but not the other. We also explicitly disallow cases
1912 // where the successor is its own predecessor, because they're
1913 // more complicated to get right.
1914 unsigned NumWith = 0;
1915 unsigned NumWithout = 0;
1916 BasicBlock *PREPred = 0;
1919 for (pred_iterator PI = pred_begin(CurrentBlock),
1920 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
1921 BasicBlock *P = *PI;
1922 // We're not interested in PRE where the block is its
1923 // own predecessor, or in blocks with predecessors
1924 // that are not reachable.
1925 if (P == CurrentBlock) {
1928 } else if (!DT->dominates(&F.getEntryBlock(), P)) {
1933 Value* predV = findLeader(P, ValNo);
1937 } else if (predV == CurInst) {
1945 // Don't do PRE when it might increase code size, i.e. when
1946 // we would need to insert instructions in more than one pred.
1947 if (NumWithout != 1 || NumWith == 0)
1950 // Don't do PRE across indirect branch.
1951 if (isa<IndirectBrInst>(PREPred->getTerminator()))
1954 // We can't do PRE safely on a critical edge, so instead we schedule
1955 // the edge to be split and perform the PRE the next time we iterate
1957 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
1958 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
1959 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
1963 // Instantiate the expression in the predecessor that lacked it.
1964 // Because we are going top-down through the block, all value numbers
1965 // will be available in the predecessor by the time we need them. Any
1966 // that weren't originally present will have been instantiated earlier
1968 Instruction *PREInstr = CurInst->clone();
1969 bool success = true;
1970 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
1971 Value *Op = PREInstr->getOperand(i);
1972 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
1975 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
1976 PREInstr->setOperand(i, V);
1983 // Fail out if we encounter an operand that is not available in
1984 // the PRE predecessor. This is typically because of loads which
1985 // are not value numbered precisely.
1988 DEBUG(verifyRemoved(PREInstr));
1992 PREInstr->insertBefore(PREPred->getTerminator());
1993 PREInstr->setName(CurInst->getName() + ".pre");
1994 predMap[PREPred] = PREInstr;
1995 VN.add(PREInstr, ValNo);
1998 // Update the availability map to include the new instruction.
1999 addToLeaderTable(ValNo, PREInstr, PREPred);
2001 // Create a PHI to make the value available in this block.
2002 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2003 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE),
2004 CurInst->getName() + ".pre-phi",
2005 CurrentBlock->begin());
2006 for (pred_iterator PI = PB; PI != PE; ++PI) {
2007 BasicBlock *P = *PI;
2008 Phi->addIncoming(predMap[P], P);
2012 addToLeaderTable(ValNo, Phi, CurrentBlock);
2014 CurInst->replaceAllUsesWith(Phi);
2015 if (Phi->getType()->isPointerTy()) {
2016 // Because we have added a PHI-use of the pointer value, it has now
2017 // "escaped" from alias analysis' perspective. We need to inform
2019 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii)
2020 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(2*ii));
2023 MD->invalidateCachedPointerInfo(Phi);
2026 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2028 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2029 if (MD) MD->removeInstruction(CurInst);
2030 CurInst->eraseFromParent();
2031 DEBUG(verifyRemoved(CurInst));
2036 if (splitCriticalEdges())
2042 /// splitCriticalEdges - Split critical edges found during the previous
2043 /// iteration that may enable further optimization.
2044 bool GVN::splitCriticalEdges() {
2045 if (toSplit.empty())
2048 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2049 SplitCriticalEdge(Edge.first, Edge.second, this);
2050 } while (!toSplit.empty());
2051 if (MD) MD->invalidateCachedPredecessors();
2055 /// iterateOnFunction - Executes one iteration of GVN
2056 bool GVN::iterateOnFunction(Function &F) {
2057 cleanupGlobalSets();
2059 // Top-down walk of the dominator tree
2060 bool Changed = false;
2062 // Needed for value numbering with phi construction to work.
2063 ReversePostOrderTraversal<Function*> RPOT(&F);
2064 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2065 RE = RPOT.end(); RI != RE; ++RI)
2066 Changed |= processBlock(*RI);
2068 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2069 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2070 Changed |= processBlock(DI->getBlock());
2076 void GVN::cleanupGlobalSets() {
2078 LeaderTable.clear();
2079 TableAllocator.Reset();
2082 /// verifyRemoved - Verify that the specified instruction does not occur in our
2083 /// internal data structures.
2084 void GVN::verifyRemoved(const Instruction *Inst) const {
2085 VN.verifyRemoved(Inst);
2087 // Walk through the value number scope to make sure the instruction isn't
2088 // ferreted away in it.
2089 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2090 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2091 const LeaderTableEntry *Node = &I->second;
2092 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2094 while (Node->Next) {
2096 assert(Node->Val != Inst && "Inst still in value numbering scope!");