1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
13 // Note that this pass does the value numbering itself; it does not use the
14 // ValueNumbering analysis passes.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "gvn"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/GlobalVariable.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/ConstantFolding.h"
25 #include "llvm/Analysis/Dominators.h"
26 #include "llvm/Analysis/InstructionSimplify.h"
27 #include "llvm/Analysis/Loads.h"
28 #include "llvm/Analysis/MemoryBuiltins.h"
29 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
30 #include "llvm/Analysis/PHITransAddr.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/Assembly/Writer.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/Target/TargetLibraryInfo.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/SSAUpdater.h"
37 #include "llvm/ADT/DenseMap.h"
38 #include "llvm/ADT/DepthFirstIterator.h"
39 #include "llvm/ADT/SmallPtrSet.h"
40 #include "llvm/ADT/Statistic.h"
41 #include "llvm/Support/Allocator.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/IRBuilder.h"
45 #include "llvm/Support/PatternMatch.h"
47 using namespace PatternMatch;
49 STATISTIC(NumGVNInstr, "Number of instructions deleted");
50 STATISTIC(NumGVNLoad, "Number of loads deleted");
51 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
52 STATISTIC(NumGVNBlocks, "Number of blocks merged");
53 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
54 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
55 STATISTIC(NumPRELoad, "Number of loads PRE'd");
57 static cl::opt<bool> EnablePRE("enable-pre",
58 cl::init(true), cl::Hidden);
59 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
61 //===----------------------------------------------------------------------===//
63 //===----------------------------------------------------------------------===//
65 /// This class holds the mapping between values and value numbers. It is used
66 /// as an efficient mechanism to determine the expression-wise equivalence of
72 SmallVector<uint32_t, 4> varargs;
74 Expression(uint32_t o = ~2U) : opcode(o) { }
76 bool operator==(const Expression &other) const {
77 if (opcode != other.opcode)
79 if (opcode == ~0U || opcode == ~1U)
81 if (type != other.type)
83 if (varargs != other.varargs)
90 DenseMap<Value*, uint32_t> valueNumbering;
91 DenseMap<Expression, uint32_t> expressionNumbering;
93 MemoryDependenceAnalysis *MD;
96 uint32_t nextValueNumber;
98 Expression create_expression(Instruction* I);
99 Expression create_cmp_expression(unsigned Opcode,
100 CmpInst::Predicate Predicate,
101 Value *LHS, Value *RHS);
102 Expression create_extractvalue_expression(ExtractValueInst* EI);
103 uint32_t lookup_or_add_call(CallInst* C);
105 ValueTable() : nextValueNumber(1) { }
106 uint32_t lookup_or_add(Value *V);
107 uint32_t lookup(Value *V) const;
108 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
109 Value *LHS, Value *RHS);
110 void add(Value *V, uint32_t num);
112 void erase(Value *v);
113 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
114 AliasAnalysis *getAliasAnalysis() const { return AA; }
115 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
116 void setDomTree(DominatorTree* D) { DT = D; }
117 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
118 void verifyRemoved(const Value *) const;
123 template <> struct DenseMapInfo<Expression> {
124 static inline Expression getEmptyKey() {
128 static inline Expression getTombstoneKey() {
132 static unsigned getHashValue(const Expression e) {
133 unsigned hash = e.opcode;
135 hash = ((unsigned)((uintptr_t)e.type >> 4) ^
136 (unsigned)((uintptr_t)e.type >> 9));
138 for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
139 E = e.varargs.end(); I != E; ++I)
140 hash = *I + hash * 37;
144 static bool isEqual(const Expression &LHS, const Expression &RHS) {
151 //===----------------------------------------------------------------------===//
152 // ValueTable Internal Functions
153 //===----------------------------------------------------------------------===//
155 Expression ValueTable::create_expression(Instruction *I) {
157 e.type = I->getType();
158 e.opcode = I->getOpcode();
159 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
161 e.varargs.push_back(lookup_or_add(*OI));
162 if (I->isCommutative()) {
163 // Ensure that commutative instructions that only differ by a permutation
164 // of their operands get the same value number by sorting the operand value
165 // numbers. Since all commutative instructions have two operands it is more
166 // efficient to sort by hand rather than using, say, std::sort.
167 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
168 if (e.varargs[0] > e.varargs[1])
169 std::swap(e.varargs[0], e.varargs[1]);
172 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
173 // Sort the operand value numbers so x<y and y>x get the same value number.
174 CmpInst::Predicate Predicate = C->getPredicate();
175 if (e.varargs[0] > e.varargs[1]) {
176 std::swap(e.varargs[0], e.varargs[1]);
177 Predicate = CmpInst::getSwappedPredicate(Predicate);
179 e.opcode = (C->getOpcode() << 8) | Predicate;
180 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
181 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
183 e.varargs.push_back(*II);
189 Expression ValueTable::create_cmp_expression(unsigned Opcode,
190 CmpInst::Predicate Predicate,
191 Value *LHS, Value *RHS) {
192 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
193 "Not a comparison!");
195 e.type = CmpInst::makeCmpResultType(LHS->getType());
196 e.varargs.push_back(lookup_or_add(LHS));
197 e.varargs.push_back(lookup_or_add(RHS));
199 // Sort the operand value numbers so x<y and y>x get the same value number.
200 if (e.varargs[0] > e.varargs[1]) {
201 std::swap(e.varargs[0], e.varargs[1]);
202 Predicate = CmpInst::getSwappedPredicate(Predicate);
204 e.opcode = (Opcode << 8) | Predicate;
208 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
209 assert(EI != 0 && "Not an ExtractValueInst?");
211 e.type = EI->getType();
214 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
215 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
216 // EI might be an extract from one of our recognised intrinsics. If it
217 // is we'll synthesize a semantically equivalent expression instead on
218 // an extract value expression.
219 switch (I->getIntrinsicID()) {
220 case Intrinsic::sadd_with_overflow:
221 case Intrinsic::uadd_with_overflow:
222 e.opcode = Instruction::Add;
224 case Intrinsic::ssub_with_overflow:
225 case Intrinsic::usub_with_overflow:
226 e.opcode = Instruction::Sub;
228 case Intrinsic::smul_with_overflow:
229 case Intrinsic::umul_with_overflow:
230 e.opcode = Instruction::Mul;
237 // Intrinsic recognized. Grab its args to finish building the expression.
238 assert(I->getNumArgOperands() == 2 &&
239 "Expect two args for recognised intrinsics.");
240 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
241 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
246 // Not a recognised intrinsic. Fall back to producing an extract value
248 e.opcode = EI->getOpcode();
249 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
251 e.varargs.push_back(lookup_or_add(*OI));
253 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
255 e.varargs.push_back(*II);
260 //===----------------------------------------------------------------------===//
261 // ValueTable External Functions
262 //===----------------------------------------------------------------------===//
264 /// add - Insert a value into the table with a specified value number.
265 void ValueTable::add(Value *V, uint32_t num) {
266 valueNumbering.insert(std::make_pair(V, num));
269 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
270 if (AA->doesNotAccessMemory(C)) {
271 Expression exp = create_expression(C);
272 uint32_t& e = expressionNumbering[exp];
273 if (!e) e = nextValueNumber++;
274 valueNumbering[C] = e;
276 } else if (AA->onlyReadsMemory(C)) {
277 Expression exp = create_expression(C);
278 uint32_t& e = expressionNumbering[exp];
280 e = nextValueNumber++;
281 valueNumbering[C] = e;
285 e = nextValueNumber++;
286 valueNumbering[C] = e;
290 MemDepResult local_dep = MD->getDependency(C);
292 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
293 valueNumbering[C] = nextValueNumber;
294 return nextValueNumber++;
297 if (local_dep.isDef()) {
298 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
300 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
301 valueNumbering[C] = nextValueNumber;
302 return nextValueNumber++;
305 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
306 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
307 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
309 valueNumbering[C] = nextValueNumber;
310 return nextValueNumber++;
314 uint32_t v = lookup_or_add(local_cdep);
315 valueNumbering[C] = v;
320 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
321 MD->getNonLocalCallDependency(CallSite(C));
322 // FIXME: Move the checking logic to MemDep!
325 // Check to see if we have a single dominating call instruction that is
327 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
328 const NonLocalDepEntry *I = &deps[i];
329 if (I->getResult().isNonLocal())
332 // We don't handle non-definitions. If we already have a call, reject
333 // instruction dependencies.
334 if (!I->getResult().isDef() || cdep != 0) {
339 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
340 // FIXME: All duplicated with non-local case.
341 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
342 cdep = NonLocalDepCall;
351 valueNumbering[C] = nextValueNumber;
352 return nextValueNumber++;
355 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
356 valueNumbering[C] = nextValueNumber;
357 return nextValueNumber++;
359 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
360 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
361 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
363 valueNumbering[C] = nextValueNumber;
364 return nextValueNumber++;
368 uint32_t v = lookup_or_add(cdep);
369 valueNumbering[C] = v;
373 valueNumbering[C] = nextValueNumber;
374 return nextValueNumber++;
378 /// lookup_or_add - Returns the value number for the specified value, assigning
379 /// it a new number if it did not have one before.
380 uint32_t ValueTable::lookup_or_add(Value *V) {
381 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
382 if (VI != valueNumbering.end())
385 if (!isa<Instruction>(V)) {
386 valueNumbering[V] = nextValueNumber;
387 return nextValueNumber++;
390 Instruction* I = cast<Instruction>(V);
392 switch (I->getOpcode()) {
393 case Instruction::Call:
394 return lookup_or_add_call(cast<CallInst>(I));
395 case Instruction::Add:
396 case Instruction::FAdd:
397 case Instruction::Sub:
398 case Instruction::FSub:
399 case Instruction::Mul:
400 case Instruction::FMul:
401 case Instruction::UDiv:
402 case Instruction::SDiv:
403 case Instruction::FDiv:
404 case Instruction::URem:
405 case Instruction::SRem:
406 case Instruction::FRem:
407 case Instruction::Shl:
408 case Instruction::LShr:
409 case Instruction::AShr:
410 case Instruction::And:
411 case Instruction::Or :
412 case Instruction::Xor:
413 case Instruction::ICmp:
414 case Instruction::FCmp:
415 case Instruction::Trunc:
416 case Instruction::ZExt:
417 case Instruction::SExt:
418 case Instruction::FPToUI:
419 case Instruction::FPToSI:
420 case Instruction::UIToFP:
421 case Instruction::SIToFP:
422 case Instruction::FPTrunc:
423 case Instruction::FPExt:
424 case Instruction::PtrToInt:
425 case Instruction::IntToPtr:
426 case Instruction::BitCast:
427 case Instruction::Select:
428 case Instruction::ExtractElement:
429 case Instruction::InsertElement:
430 case Instruction::ShuffleVector:
431 case Instruction::InsertValue:
432 case Instruction::GetElementPtr:
433 exp = create_expression(I);
435 case Instruction::ExtractValue:
436 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
439 valueNumbering[V] = nextValueNumber;
440 return nextValueNumber++;
443 uint32_t& e = expressionNumbering[exp];
444 if (!e) e = nextValueNumber++;
445 valueNumbering[V] = e;
449 /// lookup - Returns the value number of the specified value. Fails if
450 /// the value has not yet been numbered.
451 uint32_t ValueTable::lookup(Value *V) const {
452 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
453 assert(VI != valueNumbering.end() && "Value not numbered?");
457 /// lookup_or_add_cmp - Returns the value number of the given comparison,
458 /// assigning it a new number if it did not have one before. Useful when
459 /// we deduced the result of a comparison, but don't immediately have an
460 /// instruction realizing that comparison to hand.
461 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
462 CmpInst::Predicate Predicate,
463 Value *LHS, Value *RHS) {
464 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
465 uint32_t& e = expressionNumbering[exp];
466 if (!e) e = nextValueNumber++;
470 /// clear - Remove all entries from the ValueTable.
471 void ValueTable::clear() {
472 valueNumbering.clear();
473 expressionNumbering.clear();
477 /// erase - Remove a value from the value numbering.
478 void ValueTable::erase(Value *V) {
479 valueNumbering.erase(V);
482 /// verifyRemoved - Verify that the value is removed from all internal data
484 void ValueTable::verifyRemoved(const Value *V) const {
485 for (DenseMap<Value*, uint32_t>::const_iterator
486 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
487 assert(I->first != V && "Inst still occurs in value numbering map!");
491 //===----------------------------------------------------------------------===//
493 //===----------------------------------------------------------------------===//
497 class GVN : public FunctionPass {
499 MemoryDependenceAnalysis *MD;
501 const TargetData *TD;
502 const TargetLibraryInfo *TLI;
506 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
507 /// have that value number. Use findLeader to query it.
508 struct LeaderTableEntry {
511 LeaderTableEntry *Next;
513 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
514 BumpPtrAllocator TableAllocator;
516 SmallVector<Instruction*, 8> InstrsToErase;
518 static char ID; // Pass identification, replacement for typeid
519 explicit GVN(bool noloads = false)
520 : FunctionPass(ID), NoLoads(noloads), MD(0) {
521 initializeGVNPass(*PassRegistry::getPassRegistry());
524 bool runOnFunction(Function &F);
526 /// markInstructionForDeletion - This removes the specified instruction from
527 /// our various maps and marks it for deletion.
528 void markInstructionForDeletion(Instruction *I) {
530 InstrsToErase.push_back(I);
533 const TargetData *getTargetData() const { return TD; }
534 DominatorTree &getDominatorTree() const { return *DT; }
535 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
536 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
538 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
539 /// its value number.
540 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
541 LeaderTableEntry &Curr = LeaderTable[N];
548 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
551 Node->Next = Curr.Next;
555 /// removeFromLeaderTable - Scan the list of values corresponding to a given
556 /// value number, and remove the given value if encountered.
557 void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
558 LeaderTableEntry* Prev = 0;
559 LeaderTableEntry* Curr = &LeaderTable[N];
561 while (Curr->Val != V || Curr->BB != BB) {
567 Prev->Next = Curr->Next;
573 LeaderTableEntry* Next = Curr->Next;
574 Curr->Val = Next->Val;
576 Curr->Next = Next->Next;
581 // List of critical edges to be split between iterations.
582 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
584 // This transformation requires dominator postdominator info
585 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
586 AU.addRequired<DominatorTree>();
587 AU.addRequired<TargetLibraryInfo>();
589 AU.addRequired<MemoryDependenceAnalysis>();
590 AU.addRequired<AliasAnalysis>();
592 AU.addPreserved<DominatorTree>();
593 AU.addPreserved<AliasAnalysis>();
598 // FIXME: eliminate or document these better
599 bool processLoad(LoadInst *L);
600 bool processInstruction(Instruction *I);
601 bool processNonLocalLoad(LoadInst *L);
602 bool processBlock(BasicBlock *BB);
603 void dump(DenseMap<uint32_t, Value*> &d);
604 bool iterateOnFunction(Function &F);
605 bool performPRE(Function &F);
606 Value *findLeader(BasicBlock *BB, uint32_t num);
607 void cleanupGlobalSets();
608 void verifyRemoved(const Instruction *I) const;
609 bool splitCriticalEdges();
610 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
612 bool propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root);
618 // createGVNPass - The public interface to this file...
619 FunctionPass *llvm::createGVNPass(bool NoLoads) {
620 return new GVN(NoLoads);
623 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
624 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
625 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
626 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
627 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
628 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
630 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
632 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
633 E = d.end(); I != E; ++I) {
634 errs() << I->first << "\n";
640 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
641 /// we're analyzing is fully available in the specified block. As we go, keep
642 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
643 /// map is actually a tri-state map with the following values:
644 /// 0) we know the block *is not* fully available.
645 /// 1) we know the block *is* fully available.
646 /// 2) we do not know whether the block is fully available or not, but we are
647 /// currently speculating that it will be.
648 /// 3) we are speculating for this block and have used that to speculate for
650 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
651 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
652 // Optimistically assume that the block is fully available and check to see
653 // if we already know about this block in one lookup.
654 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
655 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
657 // If the entry already existed for this block, return the precomputed value.
659 // If this is a speculative "available" value, mark it as being used for
660 // speculation of other blocks.
661 if (IV.first->second == 2)
662 IV.first->second = 3;
663 return IV.first->second != 0;
666 // Otherwise, see if it is fully available in all predecessors.
667 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
669 // If this block has no predecessors, it isn't live-in here.
671 goto SpeculationFailure;
673 for (; PI != PE; ++PI)
674 // If the value isn't fully available in one of our predecessors, then it
675 // isn't fully available in this block either. Undo our previous
676 // optimistic assumption and bail out.
677 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
678 goto SpeculationFailure;
682 // SpeculationFailure - If we get here, we found out that this is not, after
683 // all, a fully-available block. We have a problem if we speculated on this and
684 // used the speculation to mark other blocks as available.
686 char &BBVal = FullyAvailableBlocks[BB];
688 // If we didn't speculate on this, just return with it set to false.
694 // If we did speculate on this value, we could have blocks set to 1 that are
695 // incorrect. Walk the (transitive) successors of this block and mark them as
697 SmallVector<BasicBlock*, 32> BBWorklist;
698 BBWorklist.push_back(BB);
701 BasicBlock *Entry = BBWorklist.pop_back_val();
702 // Note that this sets blocks to 0 (unavailable) if they happen to not
703 // already be in FullyAvailableBlocks. This is safe.
704 char &EntryVal = FullyAvailableBlocks[Entry];
705 if (EntryVal == 0) continue; // Already unavailable.
707 // Mark as unavailable.
710 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
711 BBWorklist.push_back(*I);
712 } while (!BBWorklist.empty());
718 /// CanCoerceMustAliasedValueToLoad - Return true if
719 /// CoerceAvailableValueToLoadType will succeed.
720 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
722 const TargetData &TD) {
723 // If the loaded or stored value is an first class array or struct, don't try
724 // to transform them. We need to be able to bitcast to integer.
725 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
726 StoredVal->getType()->isStructTy() ||
727 StoredVal->getType()->isArrayTy())
730 // The store has to be at least as big as the load.
731 if (TD.getTypeSizeInBits(StoredVal->getType()) <
732 TD.getTypeSizeInBits(LoadTy))
739 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
740 /// then a load from a must-aliased pointer of a different type, try to coerce
741 /// the stored value. LoadedTy is the type of the load we want to replace and
742 /// InsertPt is the place to insert new instructions.
744 /// If we can't do it, return null.
745 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
747 Instruction *InsertPt,
748 const TargetData &TD) {
749 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
752 // If this is already the right type, just return it.
753 Type *StoredValTy = StoredVal->getType();
755 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
756 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
758 // If the store and reload are the same size, we can always reuse it.
759 if (StoreSize == LoadSize) {
760 // Pointer to Pointer -> use bitcast.
761 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy())
762 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
764 // Convert source pointers to integers, which can be bitcast.
765 if (StoredValTy->isPointerTy()) {
766 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
767 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
770 Type *TypeToCastTo = LoadedTy;
771 if (TypeToCastTo->isPointerTy())
772 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
774 if (StoredValTy != TypeToCastTo)
775 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
777 // Cast to pointer if the load needs a pointer type.
778 if (LoadedTy->isPointerTy())
779 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
784 // If the loaded value is smaller than the available value, then we can
785 // extract out a piece from it. If the available value is too small, then we
786 // can't do anything.
787 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
789 // Convert source pointers to integers, which can be manipulated.
790 if (StoredValTy->isPointerTy()) {
791 StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
792 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
795 // Convert vectors and fp to integer, which can be manipulated.
796 if (!StoredValTy->isIntegerTy()) {
797 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
798 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
801 // If this is a big-endian system, we need to shift the value down to the low
802 // bits so that a truncate will work.
803 if (TD.isBigEndian()) {
804 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
805 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
808 // Truncate the integer to the right size now.
809 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
810 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
812 if (LoadedTy == NewIntTy)
815 // If the result is a pointer, inttoptr.
816 if (LoadedTy->isPointerTy())
817 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
819 // Otherwise, bitcast.
820 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
823 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
824 /// memdep query of a load that ends up being a clobbering memory write (store,
825 /// memset, memcpy, memmove). This means that the write *may* provide bits used
826 /// by the load but we can't be sure because the pointers don't mustalias.
828 /// Check this case to see if there is anything more we can do before we give
829 /// up. This returns -1 if we have to give up, or a byte number in the stored
830 /// value of the piece that feeds the load.
831 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
833 uint64_t WriteSizeInBits,
834 const TargetData &TD) {
835 // If the loaded or stored value is a first class array or struct, don't try
836 // to transform them. We need to be able to bitcast to integer.
837 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
840 int64_t StoreOffset = 0, LoadOffset = 0;
841 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD);
842 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
843 if (StoreBase != LoadBase)
846 // If the load and store are to the exact same address, they should have been
847 // a must alias. AA must have gotten confused.
848 // FIXME: Study to see if/when this happens. One case is forwarding a memset
849 // to a load from the base of the memset.
851 if (LoadOffset == StoreOffset) {
852 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
853 << "Base = " << *StoreBase << "\n"
854 << "Store Ptr = " << *WritePtr << "\n"
855 << "Store Offs = " << StoreOffset << "\n"
856 << "Load Ptr = " << *LoadPtr << "\n";
861 // If the load and store don't overlap at all, the store doesn't provide
862 // anything to the load. In this case, they really don't alias at all, AA
863 // must have gotten confused.
864 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
866 if ((WriteSizeInBits & 7) | (LoadSize & 7))
868 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
872 bool isAAFailure = false;
873 if (StoreOffset < LoadOffset)
874 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
876 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
880 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
881 << "Base = " << *StoreBase << "\n"
882 << "Store Ptr = " << *WritePtr << "\n"
883 << "Store Offs = " << StoreOffset << "\n"
884 << "Load Ptr = " << *LoadPtr << "\n";
890 // If the Load isn't completely contained within the stored bits, we don't
891 // have all the bits to feed it. We could do something crazy in the future
892 // (issue a smaller load then merge the bits in) but this seems unlikely to be
894 if (StoreOffset > LoadOffset ||
895 StoreOffset+StoreSize < LoadOffset+LoadSize)
898 // Okay, we can do this transformation. Return the number of bytes into the
899 // store that the load is.
900 return LoadOffset-StoreOffset;
903 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
904 /// memdep query of a load that ends up being a clobbering store.
905 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
907 const TargetData &TD) {
908 // Cannot handle reading from store of first-class aggregate yet.
909 if (DepSI->getValueOperand()->getType()->isStructTy() ||
910 DepSI->getValueOperand()->getType()->isArrayTy())
913 Value *StorePtr = DepSI->getPointerOperand();
914 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
915 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
916 StorePtr, StoreSize, TD);
919 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
920 /// memdep query of a load that ends up being clobbered by another load. See if
921 /// the other load can feed into the second load.
922 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
923 LoadInst *DepLI, const TargetData &TD){
924 // Cannot handle reading from store of first-class aggregate yet.
925 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
928 Value *DepPtr = DepLI->getPointerOperand();
929 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType());
930 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD);
931 if (R != -1) return R;
933 // If we have a load/load clobber an DepLI can be widened to cover this load,
934 // then we should widen it!
935 int64_t LoadOffs = 0;
936 const Value *LoadBase =
937 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD);
938 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
940 unsigned Size = MemoryDependenceAnalysis::
941 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD);
942 if (Size == 0) return -1;
944 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD);
949 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
951 const TargetData &TD) {
952 // If the mem operation is a non-constant size, we can't handle it.
953 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
954 if (SizeCst == 0) return -1;
955 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
957 // If this is memset, we just need to see if the offset is valid in the size
959 if (MI->getIntrinsicID() == Intrinsic::memset)
960 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
963 // If we have a memcpy/memmove, the only case we can handle is if this is a
964 // copy from constant memory. In that case, we can read directly from the
966 MemTransferInst *MTI = cast<MemTransferInst>(MI);
968 Constant *Src = dyn_cast<Constant>(MTI->getSource());
969 if (Src == 0) return -1;
971 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
972 if (GV == 0 || !GV->isConstant()) return -1;
974 // See if the access is within the bounds of the transfer.
975 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
976 MI->getDest(), MemSizeInBits, TD);
980 // Otherwise, see if we can constant fold a load from the constant with the
981 // offset applied as appropriate.
982 Src = ConstantExpr::getBitCast(Src,
983 llvm::Type::getInt8PtrTy(Src->getContext()));
984 Constant *OffsetCst =
985 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
986 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
987 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
988 if (ConstantFoldLoadFromConstPtr(Src, &TD))
994 /// GetStoreValueForLoad - This function is called when we have a
995 /// memdep query of a load that ends up being a clobbering store. This means
996 /// that the store provides bits used by the load but we the pointers don't
997 /// mustalias. Check this case to see if there is anything more we can do
998 /// before we give up.
999 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1001 Instruction *InsertPt, const TargetData &TD){
1002 LLVMContext &Ctx = SrcVal->getType()->getContext();
1004 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1005 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
1007 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1009 // Compute which bits of the stored value are being used by the load. Convert
1010 // to an integer type to start with.
1011 if (SrcVal->getType()->isPointerTy())
1012 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx));
1013 if (!SrcVal->getType()->isIntegerTy())
1014 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1016 // Shift the bits to the least significant depending on endianness.
1018 if (TD.isLittleEndian())
1019 ShiftAmt = Offset*8;
1021 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1024 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1026 if (LoadSize != StoreSize)
1027 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1029 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1032 /// GetLoadValueForLoad - This function is called when we have a
1033 /// memdep query of a load that ends up being a clobbering load. This means
1034 /// that the load *may* provide bits used by the load but we can't be sure
1035 /// because the pointers don't mustalias. Check this case to see if there is
1036 /// anything more we can do before we give up.
1037 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1038 Type *LoadTy, Instruction *InsertPt,
1040 const TargetData &TD = *gvn.getTargetData();
1041 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1042 // widen SrcVal out to a larger load.
1043 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType());
1044 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
1045 if (Offset+LoadSize > SrcValSize) {
1046 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1047 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1048 // If we have a load/load clobber an DepLI can be widened to cover this
1049 // load, then we should widen it to the next power of 2 size big enough!
1050 unsigned NewLoadSize = Offset+LoadSize;
1051 if (!isPowerOf2_32(NewLoadSize))
1052 NewLoadSize = NextPowerOf2(NewLoadSize);
1054 Value *PtrVal = SrcVal->getPointerOperand();
1056 // Insert the new load after the old load. This ensures that subsequent
1057 // memdep queries will find the new load. We can't easily remove the old
1058 // load completely because it is already in the value numbering table.
1059 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1061 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1062 DestPTy = PointerType::get(DestPTy,
1063 cast<PointerType>(PtrVal->getType())->getAddressSpace());
1064 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1065 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1066 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1067 NewLoad->takeName(SrcVal);
1068 NewLoad->setAlignment(SrcVal->getAlignment());
1070 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1071 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1073 // Replace uses of the original load with the wider load. On a big endian
1074 // system, we need to shift down to get the relevant bits.
1075 Value *RV = NewLoad;
1076 if (TD.isBigEndian())
1077 RV = Builder.CreateLShr(RV,
1078 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1079 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1080 SrcVal->replaceAllUsesWith(RV);
1082 // We would like to use gvn.markInstructionForDeletion here, but we can't
1083 // because the load is already memoized into the leader map table that GVN
1084 // tracks. It is potentially possible to remove the load from the table,
1085 // but then there all of the operations based on it would need to be
1086 // rehashed. Just leave the dead load around.
1087 gvn.getMemDep().removeInstruction(SrcVal);
1091 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD);
1095 /// GetMemInstValueForLoad - This function is called when we have a
1096 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1097 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1098 Type *LoadTy, Instruction *InsertPt,
1099 const TargetData &TD){
1100 LLVMContext &Ctx = LoadTy->getContext();
1101 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1103 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1105 // We know that this method is only called when the mem transfer fully
1106 // provides the bits for the load.
1107 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1108 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1109 // independently of what the offset is.
1110 Value *Val = MSI->getValue();
1112 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1114 Value *OneElt = Val;
1116 // Splat the value out to the right number of bits.
1117 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1118 // If we can double the number of bytes set, do it.
1119 if (NumBytesSet*2 <= LoadSize) {
1120 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1121 Val = Builder.CreateOr(Val, ShVal);
1126 // Otherwise insert one byte at a time.
1127 Value *ShVal = Builder.CreateShl(Val, 1*8);
1128 Val = Builder.CreateOr(OneElt, ShVal);
1132 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1135 // Otherwise, this is a memcpy/memmove from a constant global.
1136 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1137 Constant *Src = cast<Constant>(MTI->getSource());
1139 // Otherwise, see if we can constant fold a load from the constant with the
1140 // offset applied as appropriate.
1141 Src = ConstantExpr::getBitCast(Src,
1142 llvm::Type::getInt8PtrTy(Src->getContext()));
1143 Constant *OffsetCst =
1144 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1145 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1146 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1147 return ConstantFoldLoadFromConstPtr(Src, &TD);
1152 struct AvailableValueInBlock {
1153 /// BB - The basic block in question.
1156 SimpleVal, // A simple offsetted value that is accessed.
1157 LoadVal, // A value produced by a load.
1158 MemIntrin // A memory intrinsic which is loaded from.
1161 /// V - The value that is live out of the block.
1162 PointerIntPair<Value *, 2, ValType> Val;
1164 /// Offset - The byte offset in Val that is interesting for the load query.
1167 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1168 unsigned Offset = 0) {
1169 AvailableValueInBlock Res;
1171 Res.Val.setPointer(V);
1172 Res.Val.setInt(SimpleVal);
1173 Res.Offset = Offset;
1177 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1178 unsigned Offset = 0) {
1179 AvailableValueInBlock Res;
1181 Res.Val.setPointer(MI);
1182 Res.Val.setInt(MemIntrin);
1183 Res.Offset = Offset;
1187 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
1188 unsigned Offset = 0) {
1189 AvailableValueInBlock Res;
1191 Res.Val.setPointer(LI);
1192 Res.Val.setInt(LoadVal);
1193 Res.Offset = Offset;
1197 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
1198 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
1199 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
1201 Value *getSimpleValue() const {
1202 assert(isSimpleValue() && "Wrong accessor");
1203 return Val.getPointer();
1206 LoadInst *getCoercedLoadValue() const {
1207 assert(isCoercedLoadValue() && "Wrong accessor");
1208 return cast<LoadInst>(Val.getPointer());
1211 MemIntrinsic *getMemIntrinValue() const {
1212 assert(isMemIntrinValue() && "Wrong accessor");
1213 return cast<MemIntrinsic>(Val.getPointer());
1216 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
1217 /// defined here to the specified type. This handles various coercion cases.
1218 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1220 if (isSimpleValue()) {
1221 Res = getSimpleValue();
1222 if (Res->getType() != LoadTy) {
1223 const TargetData *TD = gvn.getTargetData();
1224 assert(TD && "Need target data to handle type mismatch case");
1225 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1228 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1229 << *getSimpleValue() << '\n'
1230 << *Res << '\n' << "\n\n\n");
1232 } else if (isCoercedLoadValue()) {
1233 LoadInst *Load = getCoercedLoadValue();
1234 if (Load->getType() == LoadTy && Offset == 0) {
1237 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1240 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1241 << *getCoercedLoadValue() << '\n'
1242 << *Res << '\n' << "\n\n\n");
1245 const TargetData *TD = gvn.getTargetData();
1246 assert(TD && "Need target data to handle type mismatch case");
1247 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1248 LoadTy, BB->getTerminator(), *TD);
1249 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1250 << " " << *getMemIntrinValue() << '\n'
1251 << *Res << '\n' << "\n\n\n");
1257 } // end anonymous namespace
1259 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1260 /// construct SSA form, allowing us to eliminate LI. This returns the value
1261 /// that should be used at LI's definition site.
1262 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1263 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1265 // Check for the fully redundant, dominating load case. In this case, we can
1266 // just use the dominating value directly.
1267 if (ValuesPerBlock.size() == 1 &&
1268 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1270 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1272 // Otherwise, we have to construct SSA form.
1273 SmallVector<PHINode*, 8> NewPHIs;
1274 SSAUpdater SSAUpdate(&NewPHIs);
1275 SSAUpdate.Initialize(LI->getType(), LI->getName());
1277 Type *LoadTy = LI->getType();
1279 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1280 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1281 BasicBlock *BB = AV.BB;
1283 if (SSAUpdate.HasValueForBlock(BB))
1286 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1289 // Perform PHI construction.
1290 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1292 // If new PHI nodes were created, notify alias analysis.
1293 if (V->getType()->isPointerTy()) {
1294 AliasAnalysis *AA = gvn.getAliasAnalysis();
1296 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1297 AA->copyValue(LI, NewPHIs[i]);
1299 // Now that we've copied information to the new PHIs, scan through
1300 // them again and inform alias analysis that we've added potentially
1301 // escaping uses to any values that are operands to these PHIs.
1302 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1303 PHINode *P = NewPHIs[i];
1304 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1305 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1306 AA->addEscapingUse(P->getOperandUse(jj));
1314 static bool isLifetimeStart(const Instruction *Inst) {
1315 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1316 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1320 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1321 /// non-local by performing PHI construction.
1322 bool GVN::processNonLocalLoad(LoadInst *LI) {
1323 // Find the non-local dependencies of the load.
1324 SmallVector<NonLocalDepResult, 64> Deps;
1325 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1326 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1327 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
1328 // << Deps.size() << *LI << '\n');
1330 // If we had to process more than one hundred blocks to find the
1331 // dependencies, this load isn't worth worrying about. Optimizing
1332 // it will be too expensive.
1333 unsigned NumDeps = Deps.size();
1337 // If we had a phi translation failure, we'll have a single entry which is a
1338 // clobber in the current block. Reject this early.
1340 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1342 dbgs() << "GVN: non-local load ";
1343 WriteAsOperand(dbgs(), LI);
1344 dbgs() << " has unknown dependencies\n";
1349 // Filter out useless results (non-locals, etc). Keep track of the blocks
1350 // where we have a value available in repl, also keep track of whether we see
1351 // dependencies that produce an unknown value for the load (such as a call
1352 // that could potentially clobber the load).
1353 SmallVector<AvailableValueInBlock, 64> ValuesPerBlock;
1354 SmallVector<BasicBlock*, 64> UnavailableBlocks;
1356 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1357 BasicBlock *DepBB = Deps[i].getBB();
1358 MemDepResult DepInfo = Deps[i].getResult();
1360 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1361 UnavailableBlocks.push_back(DepBB);
1365 if (DepInfo.isClobber()) {
1366 // The address being loaded in this non-local block may not be the same as
1367 // the pointer operand of the load if PHI translation occurs. Make sure
1368 // to consider the right address.
1369 Value *Address = Deps[i].getAddress();
1371 // If the dependence is to a store that writes to a superset of the bits
1372 // read by the load, we can extract the bits we need for the load from the
1374 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1375 if (TD && Address) {
1376 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1379 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1380 DepSI->getValueOperand(),
1387 // Check to see if we have something like this:
1390 // if we have this, replace the later with an extraction from the former.
1391 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1392 // If this is a clobber and L is the first instruction in its block, then
1393 // we have the first instruction in the entry block.
1394 if (DepLI != LI && Address && TD) {
1395 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
1396 LI->getPointerOperand(),
1400 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1407 // If the clobbering value is a memset/memcpy/memmove, see if we can
1408 // forward a value on from it.
1409 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1410 if (TD && Address) {
1411 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1414 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1421 UnavailableBlocks.push_back(DepBB);
1425 // DepInfo.isDef() here
1427 Instruction *DepInst = DepInfo.getInst();
1429 // Loading the allocation -> undef.
1430 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1431 // Loading immediately after lifetime begin -> undef.
1432 isLifetimeStart(DepInst)) {
1433 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1434 UndefValue::get(LI->getType())));
1438 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1439 // Reject loads and stores that are to the same address but are of
1440 // different types if we have to.
1441 if (S->getValueOperand()->getType() != LI->getType()) {
1442 // If the stored value is larger or equal to the loaded value, we can
1444 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1445 LI->getType(), *TD)) {
1446 UnavailableBlocks.push_back(DepBB);
1451 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1452 S->getValueOperand()));
1456 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1457 // If the types mismatch and we can't handle it, reject reuse of the load.
1458 if (LD->getType() != LI->getType()) {
1459 // If the stored value is larger or equal to the loaded value, we can
1461 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1462 UnavailableBlocks.push_back(DepBB);
1466 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1470 UnavailableBlocks.push_back(DepBB);
1474 // If we have no predecessors that produce a known value for this load, exit
1476 if (ValuesPerBlock.empty()) return false;
1478 // If all of the instructions we depend on produce a known value for this
1479 // load, then it is fully redundant and we can use PHI insertion to compute
1480 // its value. Insert PHIs and remove the fully redundant value now.
1481 if (UnavailableBlocks.empty()) {
1482 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1484 // Perform PHI construction.
1485 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1486 LI->replaceAllUsesWith(V);
1488 if (isa<PHINode>(V))
1490 if (V->getType()->isPointerTy())
1491 MD->invalidateCachedPointerInfo(V);
1492 markInstructionForDeletion(LI);
1497 if (!EnablePRE || !EnableLoadPRE)
1500 // Okay, we have *some* definitions of the value. This means that the value
1501 // is available in some of our (transitive) predecessors. Lets think about
1502 // doing PRE of this load. This will involve inserting a new load into the
1503 // predecessor when it's not available. We could do this in general, but
1504 // prefer to not increase code size. As such, we only do this when we know
1505 // that we only have to insert *one* load (which means we're basically moving
1506 // the load, not inserting a new one).
1508 SmallPtrSet<BasicBlock *, 4> Blockers;
1509 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1510 Blockers.insert(UnavailableBlocks[i]);
1512 // Let's find the first basic block with more than one predecessor. Walk
1513 // backwards through predecessors if needed.
1514 BasicBlock *LoadBB = LI->getParent();
1515 BasicBlock *TmpBB = LoadBB;
1517 bool isSinglePred = false;
1518 bool allSingleSucc = true;
1519 while (TmpBB->getSinglePredecessor()) {
1520 isSinglePred = true;
1521 TmpBB = TmpBB->getSinglePredecessor();
1522 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1524 if (Blockers.count(TmpBB))
1527 // If any of these blocks has more than one successor (i.e. if the edge we
1528 // just traversed was critical), then there are other paths through this
1529 // block along which the load may not be anticipated. Hoisting the load
1530 // above this block would be adding the load to execution paths along
1531 // which it was not previously executed.
1532 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1539 // FIXME: It is extremely unclear what this loop is doing, other than
1540 // artificially restricting loadpre.
1543 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1544 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1545 if (AV.isSimpleValue())
1546 // "Hot" Instruction is in some loop (because it dominates its dep.
1548 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1549 if (DT->dominates(LI, I)) {
1555 // We are interested only in "hot" instructions. We don't want to do any
1556 // mis-optimizations here.
1561 // Check to see how many predecessors have the loaded value fully
1563 DenseMap<BasicBlock*, Value*> PredLoads;
1564 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1565 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1566 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1567 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1568 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1570 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
1571 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1573 BasicBlock *Pred = *PI;
1574 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) {
1577 PredLoads[Pred] = 0;
1579 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1580 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1581 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1582 << Pred->getName() << "': " << *LI << '\n');
1586 if (LoadBB->isLandingPad()) {
1588 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1589 << Pred->getName() << "': " << *LI << '\n');
1593 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
1594 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
1598 if (!NeedToSplit.empty()) {
1599 toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
1603 // Decide whether PRE is profitable for this load.
1604 unsigned NumUnavailablePreds = PredLoads.size();
1605 assert(NumUnavailablePreds != 0 &&
1606 "Fully available value should be eliminated above!");
1608 // If this load is unavailable in multiple predecessors, reject it.
1609 // FIXME: If we could restructure the CFG, we could make a common pred with
1610 // all the preds that don't have an available LI and insert a new load into
1612 if (NumUnavailablePreds != 1)
1615 // Check if the load can safely be moved to all the unavailable predecessors.
1616 bool CanDoPRE = true;
1617 SmallVector<Instruction*, 8> NewInsts;
1618 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1619 E = PredLoads.end(); I != E; ++I) {
1620 BasicBlock *UnavailablePred = I->first;
1622 // Do PHI translation to get its value in the predecessor if necessary. The
1623 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1625 // If all preds have a single successor, then we know it is safe to insert
1626 // the load on the pred (?!?), so we can insert code to materialize the
1627 // pointer if it is not available.
1628 PHITransAddr Address(LI->getPointerOperand(), TD);
1630 if (allSingleSucc) {
1631 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1634 Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
1635 LoadPtr = Address.getAddr();
1638 // If we couldn't find or insert a computation of this phi translated value,
1641 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1642 << *LI->getPointerOperand() << "\n");
1647 // Make sure it is valid to move this load here. We have to watch out for:
1648 // @1 = getelementptr (i8* p, ...
1649 // test p and branch if == 0
1651 // It is valid to have the getelementptr before the test, even if p can
1652 // be 0, as getelementptr only does address arithmetic.
1653 // If we are not pushing the value through any multiple-successor blocks
1654 // we do not have this case. Otherwise, check that the load is safe to
1655 // put anywhere; this can be improved, but should be conservatively safe.
1656 if (!allSingleSucc &&
1657 // FIXME: REEVALUTE THIS.
1658 !isSafeToLoadUnconditionally(LoadPtr,
1659 UnavailablePred->getTerminator(),
1660 LI->getAlignment(), TD)) {
1665 I->second = LoadPtr;
1669 while (!NewInsts.empty()) {
1670 Instruction *I = NewInsts.pop_back_val();
1671 if (MD) MD->removeInstruction(I);
1672 I->eraseFromParent();
1677 // Okay, we can eliminate this load by inserting a reload in the predecessor
1678 // and using PHI construction to get the value in the other predecessors, do
1680 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1681 DEBUG(if (!NewInsts.empty())
1682 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1683 << *NewInsts.back() << '\n');
1685 // Assign value numbers to the new instructions.
1686 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1687 // FIXME: We really _ought_ to insert these value numbers into their
1688 // parent's availability map. However, in doing so, we risk getting into
1689 // ordering issues. If a block hasn't been processed yet, we would be
1690 // marking a value as AVAIL-IN, which isn't what we intend.
1691 VN.lookup_or_add(NewInsts[i]);
1694 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1695 E = PredLoads.end(); I != E; ++I) {
1696 BasicBlock *UnavailablePred = I->first;
1697 Value *LoadPtr = I->second;
1699 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1701 UnavailablePred->getTerminator());
1703 // Transfer the old load's TBAA tag to the new load.
1704 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1705 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1707 // Transfer DebugLoc.
1708 NewLoad->setDebugLoc(LI->getDebugLoc());
1710 // Add the newly created load.
1711 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1713 MD->invalidateCachedPointerInfo(LoadPtr);
1714 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1717 // Perform PHI construction.
1718 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1719 LI->replaceAllUsesWith(V);
1720 if (isa<PHINode>(V))
1722 if (V->getType()->isPointerTy())
1723 MD->invalidateCachedPointerInfo(V);
1724 markInstructionForDeletion(LI);
1729 /// processLoad - Attempt to eliminate a load, first by eliminating it
1730 /// locally, and then attempting non-local elimination if that fails.
1731 bool GVN::processLoad(LoadInst *L) {
1738 if (L->use_empty()) {
1739 markInstructionForDeletion(L);
1743 // ... to a pointer that has been loaded from before...
1744 MemDepResult Dep = MD->getDependency(L);
1746 // If we have a clobber and target data is around, see if this is a clobber
1747 // that we can fix up through code synthesis.
1748 if (Dep.isClobber() && TD) {
1749 // Check to see if we have something like this:
1750 // store i32 123, i32* %P
1751 // %A = bitcast i32* %P to i8*
1752 // %B = gep i8* %A, i32 1
1755 // We could do that by recognizing if the clobber instructions are obviously
1756 // a common base + constant offset, and if the previous store (or memset)
1757 // completely covers this load. This sort of thing can happen in bitfield
1759 Value *AvailVal = 0;
1760 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1761 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1762 L->getPointerOperand(),
1765 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1766 L->getType(), L, *TD);
1769 // Check to see if we have something like this:
1772 // if we have this, replace the later with an extraction from the former.
1773 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1774 // If this is a clobber and L is the first instruction in its block, then
1775 // we have the first instruction in the entry block.
1779 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1780 L->getPointerOperand(),
1783 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1786 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1787 // a value on from it.
1788 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1789 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1790 L->getPointerOperand(),
1793 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
1797 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1798 << *AvailVal << '\n' << *L << "\n\n\n");
1800 // Replace the load!
1801 L->replaceAllUsesWith(AvailVal);
1802 if (AvailVal->getType()->isPointerTy())
1803 MD->invalidateCachedPointerInfo(AvailVal);
1804 markInstructionForDeletion(L);
1810 // If the value isn't available, don't do anything!
1811 if (Dep.isClobber()) {
1813 // fast print dep, using operator<< on instruction is too slow.
1814 dbgs() << "GVN: load ";
1815 WriteAsOperand(dbgs(), L);
1816 Instruction *I = Dep.getInst();
1817 dbgs() << " is clobbered by " << *I << '\n';
1822 // If it is defined in another block, try harder.
1823 if (Dep.isNonLocal())
1824 return processNonLocalLoad(L);
1828 // fast print dep, using operator<< on instruction is too slow.
1829 dbgs() << "GVN: load ";
1830 WriteAsOperand(dbgs(), L);
1831 dbgs() << " has unknown dependence\n";
1836 Instruction *DepInst = Dep.getInst();
1837 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1838 Value *StoredVal = DepSI->getValueOperand();
1840 // The store and load are to a must-aliased pointer, but they may not
1841 // actually have the same type. See if we know how to reuse the stored
1842 // value (depending on its type).
1843 if (StoredVal->getType() != L->getType()) {
1845 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1850 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1851 << '\n' << *L << "\n\n\n");
1858 L->replaceAllUsesWith(StoredVal);
1859 if (StoredVal->getType()->isPointerTy())
1860 MD->invalidateCachedPointerInfo(StoredVal);
1861 markInstructionForDeletion(L);
1866 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1867 Value *AvailableVal = DepLI;
1869 // The loads are of a must-aliased pointer, but they may not actually have
1870 // the same type. See if we know how to reuse the previously loaded value
1871 // (depending on its type).
1872 if (DepLI->getType() != L->getType()) {
1874 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1876 if (AvailableVal == 0)
1879 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1880 << "\n" << *L << "\n\n\n");
1887 L->replaceAllUsesWith(AvailableVal);
1888 if (DepLI->getType()->isPointerTy())
1889 MD->invalidateCachedPointerInfo(DepLI);
1890 markInstructionForDeletion(L);
1895 // If this load really doesn't depend on anything, then we must be loading an
1896 // undef value. This can happen when loading for a fresh allocation with no
1897 // intervening stores, for example.
1898 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1899 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1900 markInstructionForDeletion(L);
1905 // If this load occurs either right after a lifetime begin,
1906 // then the loaded value is undefined.
1907 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1908 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1909 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1910 markInstructionForDeletion(L);
1919 // findLeader - In order to find a leader for a given value number at a
1920 // specific basic block, we first obtain the list of all Values for that number,
1921 // and then scan the list to find one whose block dominates the block in
1922 // question. This is fast because dominator tree queries consist of only
1923 // a few comparisons of DFS numbers.
1924 Value *GVN::findLeader(BasicBlock *BB, uint32_t num) {
1925 LeaderTableEntry Vals = LeaderTable[num];
1926 if (!Vals.Val) return 0;
1929 if (DT->dominates(Vals.BB, BB)) {
1931 if (isa<Constant>(Val)) return Val;
1934 LeaderTableEntry* Next = Vals.Next;
1936 if (DT->dominates(Next->BB, BB)) {
1937 if (isa<Constant>(Next->Val)) return Next->Val;
1938 if (!Val) Val = Next->Val;
1947 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
1948 /// use is dominated by the given basic block. Returns the number of uses that
1950 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
1953 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1955 Use &U = (UI++).getUse();
1957 // If From occurs as a phi node operand then the use implicitly lives in the
1958 // corresponding incoming block. Otherwise it is the block containing the
1959 // user that must be dominated by Root.
1960 BasicBlock *UsingBlock;
1961 if (PHINode *PN = dyn_cast<PHINode>(U.getUser()))
1962 UsingBlock = PN->getIncomingBlock(U);
1964 UsingBlock = cast<Instruction>(U.getUser())->getParent();
1966 if (DT->dominates(Root, UsingBlock)) {
1974 /// propagateEquality - The given values are known to be equal in every block
1975 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1976 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1977 bool GVN::propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root) {
1978 if (LHS == RHS) return false;
1979 assert(LHS->getType() == RHS->getType() && "Equal but types differ!");
1981 // Don't try to propagate equalities between constants.
1982 if (isa<Constant>(LHS) && isa<Constant>(RHS))
1985 // Prefer a constant on the right-hand side, or an Argument if no constants.
1986 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1987 std::swap(LHS, RHS);
1988 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1990 // If there is no obvious reason to prefer the left-hand side over the right-
1991 // hand side, ensure the longest lived term is on the right-hand side, so the
1992 // shortest lived term will be replaced by the longest lived. This tends to
1993 // expose more simplifications.
1994 uint32_t LVN = VN.lookup_or_add(LHS);
1995 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1996 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1997 // Move the 'oldest' value to the right-hand side, using the value number as
1999 uint32_t RVN = VN.lookup_or_add(RHS);
2001 std::swap(LHS, RHS);
2006 // If value numbering later deduces that an instruction in the scope is equal
2007 // to 'LHS' then ensure it will be turned into 'RHS'.
2008 addToLeaderTable(LVN, RHS, Root);
2010 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2011 // LHS always has at least one use that is not dominated by Root, this will
2012 // never do anything if LHS has only one use.
2013 bool Changed = false;
2014 if (!LHS->hasOneUse()) {
2015 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2016 Changed |= NumReplacements > 0;
2017 NumGVNEqProp += NumReplacements;
2020 // Now try to deduce additional equalities from this one. For example, if the
2021 // known equality was "(A != B)" == "false" then it follows that A and B are
2022 // equal in the scope. Only boolean equalities with an explicit true or false
2023 // RHS are currently supported.
2024 if (!RHS->getType()->isIntegerTy(1))
2025 // Not a boolean equality - bail out.
2027 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2029 // RHS neither 'true' nor 'false' - bail out.
2031 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2032 bool isKnownTrue = CI->isAllOnesValue();
2033 bool isKnownFalse = !isKnownTrue;
2035 // If "A && B" is known true then both A and B are known true. If "A || B"
2036 // is known false then both A and B are known false.
2038 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2039 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2040 Changed |= propagateEquality(A, RHS, Root);
2041 Changed |= propagateEquality(B, RHS, Root);
2045 // If we are propagating an equality like "(A == B)" == "true" then also
2046 // propagate the equality A == B. When propagating a comparison such as
2047 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2048 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
2049 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2051 // If "A == B" is known true, or "A != B" is known false, then replace
2052 // A with B everywhere in the scope.
2053 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2054 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2055 Changed |= propagateEquality(Op0, Op1, Root);
2057 // If "A >= B" is known true, replace "A < B" with false everywhere.
2058 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2059 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2060 // Since we don't have the instruction "A < B" immediately to hand, work out
2061 // the value number that it would have and use that to find an appropriate
2062 // instruction (if any).
2063 uint32_t NextNum = VN.getNextUnusedValueNumber();
2064 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2065 // If the number we were assigned was brand new then there is no point in
2066 // looking for an instruction realizing it: there cannot be one!
2067 if (Num < NextNum) {
2068 Value *NotCmp = findLeader(Root, Num);
2069 if (NotCmp && isa<Instruction>(NotCmp)) {
2070 unsigned NumReplacements =
2071 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2072 Changed |= NumReplacements > 0;
2073 NumGVNEqProp += NumReplacements;
2076 // Ensure that any instruction in scope that gets the "A < B" value number
2077 // is replaced with false.
2078 addToLeaderTable(Num, NotVal, Root);
2086 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
2087 /// true if every path from the entry block to 'Dst' passes via this edge. In
2088 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2089 static bool isOnlyReachableViaThisEdge(BasicBlock *Src, BasicBlock *Dst,
2090 DominatorTree *DT) {
2091 // While in theory it is interesting to consider the case in which Dst has
2092 // more than one predecessor, because Dst might be part of a loop which is
2093 // only reachable from Src, in practice it is pointless since at the time
2094 // GVN runs all such loops have preheaders, which means that Dst will have
2095 // been changed to have only one predecessor, namely Src.
2096 BasicBlock *Pred = Dst->getSinglePredecessor();
2097 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2102 /// processInstruction - When calculating availability, handle an instruction
2103 /// by inserting it into the appropriate sets
2104 bool GVN::processInstruction(Instruction *I) {
2105 // Ignore dbg info intrinsics.
2106 if (isa<DbgInfoIntrinsic>(I))
2109 // If the instruction can be easily simplified then do so now in preference
2110 // to value numbering it. Value numbering often exposes redundancies, for
2111 // example if it determines that %y is equal to %x then the instruction
2112 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2113 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) {
2114 I->replaceAllUsesWith(V);
2115 if (MD && V->getType()->isPointerTy())
2116 MD->invalidateCachedPointerInfo(V);
2117 markInstructionForDeletion(I);
2122 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2123 if (processLoad(LI))
2126 unsigned Num = VN.lookup_or_add(LI);
2127 addToLeaderTable(Num, LI, LI->getParent());
2131 // For conditional branches, we can perform simple conditional propagation on
2132 // the condition value itself.
2133 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2134 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
2137 Value *BranchCond = BI->getCondition();
2139 BasicBlock *TrueSucc = BI->getSuccessor(0);
2140 BasicBlock *FalseSucc = BI->getSuccessor(1);
2141 BasicBlock *Parent = BI->getParent();
2142 bool Changed = false;
2144 if (isOnlyReachableViaThisEdge(Parent, TrueSucc, DT))
2145 Changed |= propagateEquality(BranchCond,
2146 ConstantInt::getTrue(TrueSucc->getContext()),
2149 if (isOnlyReachableViaThisEdge(Parent, FalseSucc, DT))
2150 Changed |= propagateEquality(BranchCond,
2151 ConstantInt::getFalse(FalseSucc->getContext()),
2157 // For switches, propagate the case values into the case destinations.
2158 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2159 Value *SwitchCond = SI->getCondition();
2160 BasicBlock *Parent = SI->getParent();
2161 bool Changed = false;
2162 for (unsigned i = 0, e = SI->getNumCases(); i != e; ++i) {
2163 BasicBlock *Dst = SI->getCaseSuccessor(i);
2164 if (isOnlyReachableViaThisEdge(Parent, Dst, DT))
2165 Changed |= propagateEquality(SwitchCond, SI->getCaseValue(i), Dst);
2170 // Instructions with void type don't return a value, so there's
2171 // no point in trying to find redundancies in them.
2172 if (I->getType()->isVoidTy()) return false;
2174 uint32_t NextNum = VN.getNextUnusedValueNumber();
2175 unsigned Num = VN.lookup_or_add(I);
2177 // Allocations are always uniquely numbered, so we can save time and memory
2178 // by fast failing them.
2179 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2180 addToLeaderTable(Num, I, I->getParent());
2184 // If the number we were assigned was a brand new VN, then we don't
2185 // need to do a lookup to see if the number already exists
2186 // somewhere in the domtree: it can't!
2187 if (Num >= NextNum) {
2188 addToLeaderTable(Num, I, I->getParent());
2192 // Perform fast-path value-number based elimination of values inherited from
2194 Value *repl = findLeader(I->getParent(), Num);
2196 // Failure, just remember this instance for future use.
2197 addToLeaderTable(Num, I, I->getParent());
2202 I->replaceAllUsesWith(repl);
2203 if (MD && repl->getType()->isPointerTy())
2204 MD->invalidateCachedPointerInfo(repl);
2205 markInstructionForDeletion(I);
2209 /// runOnFunction - This is the main transformation entry point for a function.
2210 bool GVN::runOnFunction(Function& F) {
2212 MD = &getAnalysis<MemoryDependenceAnalysis>();
2213 DT = &getAnalysis<DominatorTree>();
2214 TD = getAnalysisIfAvailable<TargetData>();
2215 TLI = &getAnalysis<TargetLibraryInfo>();
2216 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2220 bool Changed = false;
2221 bool ShouldContinue = true;
2223 // Merge unconditional branches, allowing PRE to catch more
2224 // optimization opportunities.
2225 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2226 BasicBlock *BB = FI++;
2228 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2229 if (removedBlock) ++NumGVNBlocks;
2231 Changed |= removedBlock;
2234 unsigned Iteration = 0;
2235 while (ShouldContinue) {
2236 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2237 ShouldContinue = iterateOnFunction(F);
2238 if (splitCriticalEdges())
2239 ShouldContinue = true;
2240 Changed |= ShouldContinue;
2245 bool PREChanged = true;
2246 while (PREChanged) {
2247 PREChanged = performPRE(F);
2248 Changed |= PREChanged;
2251 // FIXME: Should perform GVN again after PRE does something. PRE can move
2252 // computations into blocks where they become fully redundant. Note that
2253 // we can't do this until PRE's critical edge splitting updates memdep.
2254 // Actually, when this happens, we should just fully integrate PRE into GVN.
2256 cleanupGlobalSets();
2262 bool GVN::processBlock(BasicBlock *BB) {
2263 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2264 // (and incrementing BI before processing an instruction).
2265 assert(InstrsToErase.empty() &&
2266 "We expect InstrsToErase to be empty across iterations");
2267 bool ChangedFunction = false;
2269 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2271 ChangedFunction |= processInstruction(BI);
2272 if (InstrsToErase.empty()) {
2277 // If we need some instructions deleted, do it now.
2278 NumGVNInstr += InstrsToErase.size();
2280 // Avoid iterator invalidation.
2281 bool AtStart = BI == BB->begin();
2285 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(),
2286 E = InstrsToErase.end(); I != E; ++I) {
2287 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2288 if (MD) MD->removeInstruction(*I);
2289 (*I)->eraseFromParent();
2290 DEBUG(verifyRemoved(*I));
2292 InstrsToErase.clear();
2300 return ChangedFunction;
2303 /// performPRE - Perform a purely local form of PRE that looks for diamond
2304 /// control flow patterns and attempts to perform simple PRE at the join point.
2305 bool GVN::performPRE(Function &F) {
2306 bool Changed = false;
2307 DenseMap<BasicBlock*, Value*> predMap;
2308 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2309 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2310 BasicBlock *CurrentBlock = *DI;
2312 // Nothing to PRE in the entry block.
2313 if (CurrentBlock == &F.getEntryBlock()) continue;
2315 // Don't perform PRE on a landing pad.
2316 if (CurrentBlock->isLandingPad()) continue;
2318 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2319 BE = CurrentBlock->end(); BI != BE; ) {
2320 Instruction *CurInst = BI++;
2322 if (isa<AllocaInst>(CurInst) ||
2323 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2324 CurInst->getType()->isVoidTy() ||
2325 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2326 isa<DbgInfoIntrinsic>(CurInst))
2329 // We don't currently value number ANY inline asm calls.
2330 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2331 if (CallI->isInlineAsm())
2334 uint32_t ValNo = VN.lookup(CurInst);
2336 // Look for the predecessors for PRE opportunities. We're
2337 // only trying to solve the basic diamond case, where
2338 // a value is computed in the successor and one predecessor,
2339 // but not the other. We also explicitly disallow cases
2340 // where the successor is its own predecessor, because they're
2341 // more complicated to get right.
2342 unsigned NumWith = 0;
2343 unsigned NumWithout = 0;
2344 BasicBlock *PREPred = 0;
2347 for (pred_iterator PI = pred_begin(CurrentBlock),
2348 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2349 BasicBlock *P = *PI;
2350 // We're not interested in PRE where the block is its
2351 // own predecessor, or in blocks with predecessors
2352 // that are not reachable.
2353 if (P == CurrentBlock) {
2356 } else if (!DT->dominates(&F.getEntryBlock(), P)) {
2361 Value* predV = findLeader(P, ValNo);
2365 } else if (predV == CurInst) {
2373 // Don't do PRE when it might increase code size, i.e. when
2374 // we would need to insert instructions in more than one pred.
2375 if (NumWithout != 1 || NumWith == 0)
2378 // Don't do PRE across indirect branch.
2379 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2382 // We can't do PRE safely on a critical edge, so instead we schedule
2383 // the edge to be split and perform the PRE the next time we iterate
2385 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2386 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2387 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2391 // Instantiate the expression in the predecessor that lacked it.
2392 // Because we are going top-down through the block, all value numbers
2393 // will be available in the predecessor by the time we need them. Any
2394 // that weren't originally present will have been instantiated earlier
2396 Instruction *PREInstr = CurInst->clone();
2397 bool success = true;
2398 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2399 Value *Op = PREInstr->getOperand(i);
2400 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2403 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2404 PREInstr->setOperand(i, V);
2411 // Fail out if we encounter an operand that is not available in
2412 // the PRE predecessor. This is typically because of loads which
2413 // are not value numbered precisely.
2416 DEBUG(verifyRemoved(PREInstr));
2420 PREInstr->insertBefore(PREPred->getTerminator());
2421 PREInstr->setName(CurInst->getName() + ".pre");
2422 PREInstr->setDebugLoc(CurInst->getDebugLoc());
2423 predMap[PREPred] = PREInstr;
2424 VN.add(PREInstr, ValNo);
2427 // Update the availability map to include the new instruction.
2428 addToLeaderTable(ValNo, PREInstr, PREPred);
2430 // Create a PHI to make the value available in this block.
2431 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2432 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE),
2433 CurInst->getName() + ".pre-phi",
2434 CurrentBlock->begin());
2435 for (pred_iterator PI = PB; PI != PE; ++PI) {
2436 BasicBlock *P = *PI;
2437 Phi->addIncoming(predMap[P], P);
2441 addToLeaderTable(ValNo, Phi, CurrentBlock);
2442 Phi->setDebugLoc(CurInst->getDebugLoc());
2443 CurInst->replaceAllUsesWith(Phi);
2444 if (Phi->getType()->isPointerTy()) {
2445 // Because we have added a PHI-use of the pointer value, it has now
2446 // "escaped" from alias analysis' perspective. We need to inform
2448 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
2450 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2451 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2455 MD->invalidateCachedPointerInfo(Phi);
2458 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2460 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2461 if (MD) MD->removeInstruction(CurInst);
2462 CurInst->eraseFromParent();
2463 DEBUG(verifyRemoved(CurInst));
2468 if (splitCriticalEdges())
2474 /// splitCriticalEdges - Split critical edges found during the previous
2475 /// iteration that may enable further optimization.
2476 bool GVN::splitCriticalEdges() {
2477 if (toSplit.empty())
2480 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2481 SplitCriticalEdge(Edge.first, Edge.second, this);
2482 } while (!toSplit.empty());
2483 if (MD) MD->invalidateCachedPredecessors();
2487 /// iterateOnFunction - Executes one iteration of GVN
2488 bool GVN::iterateOnFunction(Function &F) {
2489 cleanupGlobalSets();
2491 // Top-down walk of the dominator tree
2492 bool Changed = false;
2494 // Needed for value numbering with phi construction to work.
2495 ReversePostOrderTraversal<Function*> RPOT(&F);
2496 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2497 RE = RPOT.end(); RI != RE; ++RI)
2498 Changed |= processBlock(*RI);
2500 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2501 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2502 Changed |= processBlock(DI->getBlock());
2508 void GVN::cleanupGlobalSets() {
2510 LeaderTable.clear();
2511 TableAllocator.Reset();
2514 /// verifyRemoved - Verify that the specified instruction does not occur in our
2515 /// internal data structures.
2516 void GVN::verifyRemoved(const Instruction *Inst) const {
2517 VN.verifyRemoved(Inst);
2519 // Walk through the value number scope to make sure the instruction isn't
2520 // ferreted away in it.
2521 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2522 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2523 const LeaderTableEntry *Node = &I->second;
2524 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2526 while (Node->Next) {
2528 assert(Node->Val != Inst && "Inst still in value numbering scope!");