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/ADT/DenseMap.h"
21 #include "llvm/ADT/DepthFirstIterator.h"
22 #include "llvm/ADT/Hashing.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/AliasAnalysis.h"
26 #include "llvm/Analysis/ConstantFolding.h"
27 #include "llvm/Analysis/Dominators.h"
28 #include "llvm/Analysis/InstructionSimplify.h"
29 #include "llvm/Analysis/Loads.h"
30 #include "llvm/Analysis/MemoryBuiltins.h"
31 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
32 #include "llvm/Analysis/PHITransAddr.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/Assembly/Writer.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/GlobalVariable.h"
37 #include "llvm/IR/IRBuilder.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/IR/LLVMContext.h"
40 #include "llvm/IR/Metadata.h"
41 #include "llvm/Support/Allocator.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/PatternMatch.h"
45 #include "llvm/Target/TargetLibraryInfo.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Transforms/Utils/SSAUpdater.h"
49 using namespace PatternMatch;
51 STATISTIC(NumGVNInstr, "Number of instructions deleted");
52 STATISTIC(NumGVNLoad, "Number of loads deleted");
53 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
54 STATISTIC(NumGVNBlocks, "Number of blocks merged");
55 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
56 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
57 STATISTIC(NumPRELoad, "Number of loads PRE'd");
59 static cl::opt<bool> EnablePRE("enable-pre",
60 cl::init(true), cl::Hidden);
61 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
63 // Maximum allowed recursion depth.
64 static cl::opt<uint32_t>
65 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
66 cl::desc("Max recurse depth (default = 1000)"));
68 //===----------------------------------------------------------------------===//
70 //===----------------------------------------------------------------------===//
72 /// This class holds the mapping between values and value numbers. It is used
73 /// as an efficient mechanism to determine the expression-wise equivalence of
79 SmallVector<uint32_t, 4> varargs;
81 Expression(uint32_t o = ~2U) : opcode(o) { }
83 bool operator==(const Expression &other) const {
84 if (opcode != other.opcode)
86 if (opcode == ~0U || opcode == ~1U)
88 if (type != other.type)
90 if (varargs != other.varargs)
95 friend hash_code hash_value(const Expression &Value) {
96 return hash_combine(Value.opcode, Value.type,
97 hash_combine_range(Value.varargs.begin(),
98 Value.varargs.end()));
103 DenseMap<Value*, uint32_t> valueNumbering;
104 DenseMap<Expression, uint32_t> expressionNumbering;
106 MemoryDependenceAnalysis *MD;
109 uint32_t nextValueNumber;
111 Expression create_expression(Instruction* I);
112 Expression create_cmp_expression(unsigned Opcode,
113 CmpInst::Predicate Predicate,
114 Value *LHS, Value *RHS);
115 Expression create_extractvalue_expression(ExtractValueInst* EI);
116 uint32_t lookup_or_add_call(CallInst* C);
118 ValueTable() : nextValueNumber(1) { }
119 uint32_t lookup_or_add(Value *V);
120 uint32_t lookup(Value *V) const;
121 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
122 Value *LHS, Value *RHS);
123 void add(Value *V, uint32_t num);
125 void erase(Value *v);
126 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
127 AliasAnalysis *getAliasAnalysis() const { return AA; }
128 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
129 void setDomTree(DominatorTree* D) { DT = D; }
130 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
131 void verifyRemoved(const Value *) const;
136 template <> struct DenseMapInfo<Expression> {
137 static inline Expression getEmptyKey() {
141 static inline Expression getTombstoneKey() {
145 static unsigned getHashValue(const Expression e) {
146 using llvm::hash_value;
147 return static_cast<unsigned>(hash_value(e));
149 static bool isEqual(const Expression &LHS, const Expression &RHS) {
156 //===----------------------------------------------------------------------===//
157 // ValueTable Internal Functions
158 //===----------------------------------------------------------------------===//
160 Expression ValueTable::create_expression(Instruction *I) {
162 e.type = I->getType();
163 e.opcode = I->getOpcode();
164 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
166 e.varargs.push_back(lookup_or_add(*OI));
167 if (I->isCommutative()) {
168 // Ensure that commutative instructions that only differ by a permutation
169 // of their operands get the same value number by sorting the operand value
170 // numbers. Since all commutative instructions have two operands it is more
171 // efficient to sort by hand rather than using, say, std::sort.
172 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
173 if (e.varargs[0] > e.varargs[1])
174 std::swap(e.varargs[0], e.varargs[1]);
177 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
178 // Sort the operand value numbers so x<y and y>x get the same value number.
179 CmpInst::Predicate Predicate = C->getPredicate();
180 if (e.varargs[0] > e.varargs[1]) {
181 std::swap(e.varargs[0], e.varargs[1]);
182 Predicate = CmpInst::getSwappedPredicate(Predicate);
184 e.opcode = (C->getOpcode() << 8) | Predicate;
185 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
186 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
188 e.varargs.push_back(*II);
194 Expression ValueTable::create_cmp_expression(unsigned Opcode,
195 CmpInst::Predicate Predicate,
196 Value *LHS, Value *RHS) {
197 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
198 "Not a comparison!");
200 e.type = CmpInst::makeCmpResultType(LHS->getType());
201 e.varargs.push_back(lookup_or_add(LHS));
202 e.varargs.push_back(lookup_or_add(RHS));
204 // Sort the operand value numbers so x<y and y>x get the same value number.
205 if (e.varargs[0] > e.varargs[1]) {
206 std::swap(e.varargs[0], e.varargs[1]);
207 Predicate = CmpInst::getSwappedPredicate(Predicate);
209 e.opcode = (Opcode << 8) | Predicate;
213 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
214 assert(EI != 0 && "Not an ExtractValueInst?");
216 e.type = EI->getType();
219 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
220 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
221 // EI might be an extract from one of our recognised intrinsics. If it
222 // is we'll synthesize a semantically equivalent expression instead on
223 // an extract value expression.
224 switch (I->getIntrinsicID()) {
225 case Intrinsic::sadd_with_overflow:
226 case Intrinsic::uadd_with_overflow:
227 e.opcode = Instruction::Add;
229 case Intrinsic::ssub_with_overflow:
230 case Intrinsic::usub_with_overflow:
231 e.opcode = Instruction::Sub;
233 case Intrinsic::smul_with_overflow:
234 case Intrinsic::umul_with_overflow:
235 e.opcode = Instruction::Mul;
242 // Intrinsic recognized. Grab its args to finish building the expression.
243 assert(I->getNumArgOperands() == 2 &&
244 "Expect two args for recognised intrinsics.");
245 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
246 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
251 // Not a recognised intrinsic. Fall back to producing an extract value
253 e.opcode = EI->getOpcode();
254 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
256 e.varargs.push_back(lookup_or_add(*OI));
258 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
260 e.varargs.push_back(*II);
265 //===----------------------------------------------------------------------===//
266 // ValueTable External Functions
267 //===----------------------------------------------------------------------===//
269 /// add - Insert a value into the table with a specified value number.
270 void ValueTable::add(Value *V, uint32_t num) {
271 valueNumbering.insert(std::make_pair(V, num));
274 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
275 if (AA->doesNotAccessMemory(C)) {
276 Expression exp = create_expression(C);
277 uint32_t &e = expressionNumbering[exp];
278 if (!e) e = nextValueNumber++;
279 valueNumbering[C] = e;
281 } else if (AA->onlyReadsMemory(C)) {
282 Expression exp = create_expression(C);
283 uint32_t &e = expressionNumbering[exp];
285 e = nextValueNumber++;
286 valueNumbering[C] = e;
290 e = nextValueNumber++;
291 valueNumbering[C] = e;
295 MemDepResult local_dep = MD->getDependency(C);
297 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
298 valueNumbering[C] = nextValueNumber;
299 return nextValueNumber++;
302 if (local_dep.isDef()) {
303 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
305 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
306 valueNumbering[C] = nextValueNumber;
307 return nextValueNumber++;
310 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
311 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
312 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
314 valueNumbering[C] = nextValueNumber;
315 return nextValueNumber++;
319 uint32_t v = lookup_or_add(local_cdep);
320 valueNumbering[C] = v;
325 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
326 MD->getNonLocalCallDependency(CallSite(C));
327 // FIXME: Move the checking logic to MemDep!
330 // Check to see if we have a single dominating call instruction that is
332 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
333 const NonLocalDepEntry *I = &deps[i];
334 if (I->getResult().isNonLocal())
337 // We don't handle non-definitions. If we already have a call, reject
338 // instruction dependencies.
339 if (!I->getResult().isDef() || cdep != 0) {
344 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
345 // FIXME: All duplicated with non-local case.
346 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
347 cdep = NonLocalDepCall;
356 valueNumbering[C] = nextValueNumber;
357 return nextValueNumber++;
360 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
361 valueNumbering[C] = nextValueNumber;
362 return nextValueNumber++;
364 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
365 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
366 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
368 valueNumbering[C] = nextValueNumber;
369 return nextValueNumber++;
373 uint32_t v = lookup_or_add(cdep);
374 valueNumbering[C] = v;
378 valueNumbering[C] = nextValueNumber;
379 return nextValueNumber++;
383 /// lookup_or_add - Returns the value number for the specified value, assigning
384 /// it a new number if it did not have one before.
385 uint32_t ValueTable::lookup_or_add(Value *V) {
386 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
387 if (VI != valueNumbering.end())
390 if (!isa<Instruction>(V)) {
391 valueNumbering[V] = nextValueNumber;
392 return nextValueNumber++;
395 Instruction* I = cast<Instruction>(V);
397 switch (I->getOpcode()) {
398 case Instruction::Call:
399 return lookup_or_add_call(cast<CallInst>(I));
400 case Instruction::Add:
401 case Instruction::FAdd:
402 case Instruction::Sub:
403 case Instruction::FSub:
404 case Instruction::Mul:
405 case Instruction::FMul:
406 case Instruction::UDiv:
407 case Instruction::SDiv:
408 case Instruction::FDiv:
409 case Instruction::URem:
410 case Instruction::SRem:
411 case Instruction::FRem:
412 case Instruction::Shl:
413 case Instruction::LShr:
414 case Instruction::AShr:
415 case Instruction::And:
416 case Instruction::Or:
417 case Instruction::Xor:
418 case Instruction::ICmp:
419 case Instruction::FCmp:
420 case Instruction::Trunc:
421 case Instruction::ZExt:
422 case Instruction::SExt:
423 case Instruction::FPToUI:
424 case Instruction::FPToSI:
425 case Instruction::UIToFP:
426 case Instruction::SIToFP:
427 case Instruction::FPTrunc:
428 case Instruction::FPExt:
429 case Instruction::PtrToInt:
430 case Instruction::IntToPtr:
431 case Instruction::BitCast:
432 case Instruction::Select:
433 case Instruction::ExtractElement:
434 case Instruction::InsertElement:
435 case Instruction::ShuffleVector:
436 case Instruction::InsertValue:
437 case Instruction::GetElementPtr:
438 exp = create_expression(I);
440 case Instruction::ExtractValue:
441 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
444 valueNumbering[V] = nextValueNumber;
445 return nextValueNumber++;
448 uint32_t& e = expressionNumbering[exp];
449 if (!e) e = nextValueNumber++;
450 valueNumbering[V] = e;
454 /// lookup - Returns the value number of the specified value. Fails if
455 /// the value has not yet been numbered.
456 uint32_t ValueTable::lookup(Value *V) const {
457 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
458 assert(VI != valueNumbering.end() && "Value not numbered?");
462 /// lookup_or_add_cmp - Returns the value number of the given comparison,
463 /// assigning it a new number if it did not have one before. Useful when
464 /// we deduced the result of a comparison, but don't immediately have an
465 /// instruction realizing that comparison to hand.
466 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
467 CmpInst::Predicate Predicate,
468 Value *LHS, Value *RHS) {
469 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
470 uint32_t& e = expressionNumbering[exp];
471 if (!e) e = nextValueNumber++;
475 /// clear - Remove all entries from the ValueTable.
476 void ValueTable::clear() {
477 valueNumbering.clear();
478 expressionNumbering.clear();
482 /// erase - Remove a value from the value numbering.
483 void ValueTable::erase(Value *V) {
484 valueNumbering.erase(V);
487 /// verifyRemoved - Verify that the value is removed from all internal data
489 void ValueTable::verifyRemoved(const Value *V) const {
490 for (DenseMap<Value*, uint32_t>::const_iterator
491 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
492 assert(I->first != V && "Inst still occurs in value numbering map!");
496 //===----------------------------------------------------------------------===//
498 //===----------------------------------------------------------------------===//
502 class GVN : public FunctionPass {
504 MemoryDependenceAnalysis *MD;
506 const DataLayout *TD;
507 const TargetLibraryInfo *TLI;
511 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
512 /// have that value number. Use findLeader to query it.
513 struct LeaderTableEntry {
515 const BasicBlock *BB;
516 LeaderTableEntry *Next;
518 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
519 BumpPtrAllocator TableAllocator;
521 SmallVector<Instruction*, 8> InstrsToErase;
523 static char ID; // Pass identification, replacement for typeid
524 explicit GVN(bool noloads = false)
525 : FunctionPass(ID), NoLoads(noloads), MD(0) {
526 initializeGVNPass(*PassRegistry::getPassRegistry());
529 bool runOnFunction(Function &F);
531 /// markInstructionForDeletion - This removes the specified instruction from
532 /// our various maps and marks it for deletion.
533 void markInstructionForDeletion(Instruction *I) {
535 InstrsToErase.push_back(I);
538 const DataLayout *getDataLayout() const { return TD; }
539 DominatorTree &getDominatorTree() const { return *DT; }
540 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
541 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
543 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
544 /// its value number.
545 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
546 LeaderTableEntry &Curr = LeaderTable[N];
553 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
556 Node->Next = Curr.Next;
560 /// removeFromLeaderTable - Scan the list of values corresponding to a given
561 /// value number, and remove the given instruction if encountered.
562 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
563 LeaderTableEntry* Prev = 0;
564 LeaderTableEntry* Curr = &LeaderTable[N];
566 while (Curr->Val != I || Curr->BB != BB) {
572 Prev->Next = Curr->Next;
578 LeaderTableEntry* Next = Curr->Next;
579 Curr->Val = Next->Val;
581 Curr->Next = Next->Next;
586 // List of critical edges to be split between iterations.
587 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
589 // This transformation requires dominator postdominator info
590 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
591 AU.addRequired<DominatorTree>();
592 AU.addRequired<TargetLibraryInfo>();
594 AU.addRequired<MemoryDependenceAnalysis>();
595 AU.addRequired<AliasAnalysis>();
597 AU.addPreserved<DominatorTree>();
598 AU.addPreserved<AliasAnalysis>();
603 // FIXME: eliminate or document these better
604 bool processLoad(LoadInst *L);
605 bool processInstruction(Instruction *I);
606 bool processNonLocalLoad(LoadInst *L);
607 bool processBlock(BasicBlock *BB);
608 void dump(DenseMap<uint32_t, Value*> &d);
609 bool iterateOnFunction(Function &F);
610 bool performPRE(Function &F);
611 Value *findLeader(const BasicBlock *BB, uint32_t num);
612 void cleanupGlobalSets();
613 void verifyRemoved(const Instruction *I) const;
614 bool splitCriticalEdges();
615 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
616 const BasicBlockEdge &Root);
617 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
623 // createGVNPass - The public interface to this file...
624 FunctionPass *llvm::createGVNPass(bool NoLoads) {
625 return new GVN(NoLoads);
628 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
629 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
630 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
631 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
632 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
633 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
635 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
636 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
638 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
639 E = d.end(); I != E; ++I) {
640 errs() << I->first << "\n";
647 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
648 /// we're analyzing is fully available in the specified block. As we go, keep
649 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
650 /// map is actually a tri-state map with the following values:
651 /// 0) we know the block *is not* fully available.
652 /// 1) we know the block *is* fully available.
653 /// 2) we do not know whether the block is fully available or not, but we are
654 /// currently speculating that it will be.
655 /// 3) we are speculating for this block and have used that to speculate for
657 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
658 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
659 uint32_t RecurseDepth) {
660 if (RecurseDepth > MaxRecurseDepth)
663 // Optimistically assume that the block is fully available and check to see
664 // if we already know about this block in one lookup.
665 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
666 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
668 // If the entry already existed for this block, return the precomputed value.
670 // If this is a speculative "available" value, mark it as being used for
671 // speculation of other blocks.
672 if (IV.first->second == 2)
673 IV.first->second = 3;
674 return IV.first->second != 0;
677 // Otherwise, see if it is fully available in all predecessors.
678 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
680 // If this block has no predecessors, it isn't live-in here.
682 goto SpeculationFailure;
684 for (; PI != PE; ++PI)
685 // If the value isn't fully available in one of our predecessors, then it
686 // isn't fully available in this block either. Undo our previous
687 // optimistic assumption and bail out.
688 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
689 goto SpeculationFailure;
693 // SpeculationFailure - If we get here, we found out that this is not, after
694 // all, a fully-available block. We have a problem if we speculated on this and
695 // used the speculation to mark other blocks as available.
697 char &BBVal = FullyAvailableBlocks[BB];
699 // If we didn't speculate on this, just return with it set to false.
705 // If we did speculate on this value, we could have blocks set to 1 that are
706 // incorrect. Walk the (transitive) successors of this block and mark them as
708 SmallVector<BasicBlock*, 32> BBWorklist;
709 BBWorklist.push_back(BB);
712 BasicBlock *Entry = BBWorklist.pop_back_val();
713 // Note that this sets blocks to 0 (unavailable) if they happen to not
714 // already be in FullyAvailableBlocks. This is safe.
715 char &EntryVal = FullyAvailableBlocks[Entry];
716 if (EntryVal == 0) continue; // Already unavailable.
718 // Mark as unavailable.
721 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
722 BBWorklist.push_back(*I);
723 } while (!BBWorklist.empty());
729 /// CanCoerceMustAliasedValueToLoad - Return true if
730 /// CoerceAvailableValueToLoadType will succeed.
731 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
733 const DataLayout &TD) {
734 // If the loaded or stored value is an first class array or struct, don't try
735 // to transform them. We need to be able to bitcast to integer.
736 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
737 StoredVal->getType()->isStructTy() ||
738 StoredVal->getType()->isArrayTy())
741 // The store has to be at least as big as the load.
742 if (TD.getTypeSizeInBits(StoredVal->getType()) <
743 TD.getTypeSizeInBits(LoadTy))
749 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
750 /// then a load from a must-aliased pointer of a different type, try to coerce
751 /// the stored value. LoadedTy is the type of the load we want to replace and
752 /// InsertPt is the place to insert new instructions.
754 /// If we can't do it, return null.
755 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
757 Instruction *InsertPt,
758 const DataLayout &TD) {
759 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
762 // If this is already the right type, just return it.
763 Type *StoredValTy = StoredVal->getType();
765 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
766 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
768 // If the store and reload are the same size, we can always reuse it.
769 if (StoreSize == LoadSize) {
770 // Pointer to Pointer -> use bitcast.
771 if (StoredValTy->getScalarType()->isPointerTy() &&
772 LoadedTy->getScalarType()->isPointerTy())
773 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
775 // Convert source pointers to integers, which can be bitcast.
776 if (StoredValTy->getScalarType()->isPointerTy()) {
777 StoredValTy = TD.getIntPtrType(StoredValTy);
778 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
781 Type *TypeToCastTo = LoadedTy;
782 if (TypeToCastTo->getScalarType()->isPointerTy())
783 TypeToCastTo = TD.getIntPtrType(TypeToCastTo);
785 if (StoredValTy != TypeToCastTo)
786 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
788 // Cast to pointer if the load needs a pointer type.
789 if (LoadedTy->getScalarType()->isPointerTy())
790 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
795 // If the loaded value is smaller than the available value, then we can
796 // extract out a piece from it. If the available value is too small, then we
797 // can't do anything.
798 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
800 // Convert source pointers to integers, which can be manipulated.
801 if (StoredValTy->getScalarType()->isPointerTy()) {
802 StoredValTy = TD.getIntPtrType(StoredValTy);
803 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
806 // Convert vectors and fp to integer, which can be manipulated.
807 if (!StoredValTy->isIntegerTy()) {
808 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
809 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
812 // If this is a big-endian system, we need to shift the value down to the low
813 // bits so that a truncate will work.
814 if (TD.isBigEndian()) {
815 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
816 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
819 // Truncate the integer to the right size now.
820 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
821 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
823 if (LoadedTy == NewIntTy)
826 // If the result is a pointer, inttoptr.
827 if (LoadedTy->getScalarType()->isPointerTy())
828 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
830 // Otherwise, bitcast.
831 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
834 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
835 /// memdep query of a load that ends up being a clobbering memory write (store,
836 /// memset, memcpy, memmove). This means that the write *may* provide bits used
837 /// by the load but we can't be sure because the pointers don't mustalias.
839 /// Check this case to see if there is anything more we can do before we give
840 /// up. This returns -1 if we have to give up, or a byte number in the stored
841 /// value of the piece that feeds the load.
842 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
844 uint64_t WriteSizeInBits,
845 const DataLayout &TD) {
846 // If the loaded or stored value is a first class array or struct, don't try
847 // to transform them. We need to be able to bitcast to integer.
848 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
851 int64_t StoreOffset = 0, LoadOffset = 0;
852 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&TD);
853 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &TD);
854 if (StoreBase != LoadBase)
857 // If the load and store are to the exact same address, they should have been
858 // a must alias. AA must have gotten confused.
859 // FIXME: Study to see if/when this happens. One case is forwarding a memset
860 // to a load from the base of the memset.
862 if (LoadOffset == StoreOffset) {
863 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
864 << "Base = " << *StoreBase << "\n"
865 << "Store Ptr = " << *WritePtr << "\n"
866 << "Store Offs = " << StoreOffset << "\n"
867 << "Load Ptr = " << *LoadPtr << "\n";
872 // If the load and store don't overlap at all, the store doesn't provide
873 // anything to the load. In this case, they really don't alias at all, AA
874 // must have gotten confused.
875 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
877 if ((WriteSizeInBits & 7) | (LoadSize & 7))
879 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
883 bool isAAFailure = false;
884 if (StoreOffset < LoadOffset)
885 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
887 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
891 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
892 << "Base = " << *StoreBase << "\n"
893 << "Store Ptr = " << *WritePtr << "\n"
894 << "Store Offs = " << StoreOffset << "\n"
895 << "Load Ptr = " << *LoadPtr << "\n";
901 // If the Load isn't completely contained within the stored bits, we don't
902 // have all the bits to feed it. We could do something crazy in the future
903 // (issue a smaller load then merge the bits in) but this seems unlikely to be
905 if (StoreOffset > LoadOffset ||
906 StoreOffset+StoreSize < LoadOffset+LoadSize)
909 // Okay, we can do this transformation. Return the number of bytes into the
910 // store that the load is.
911 return LoadOffset-StoreOffset;
914 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
915 /// memdep query of a load that ends up being a clobbering store.
916 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
918 const DataLayout &TD) {
919 // Cannot handle reading from store of first-class aggregate yet.
920 if (DepSI->getValueOperand()->getType()->isStructTy() ||
921 DepSI->getValueOperand()->getType()->isArrayTy())
924 Value *StorePtr = DepSI->getPointerOperand();
925 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
926 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
927 StorePtr, StoreSize, TD);
930 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
931 /// memdep query of a load that ends up being clobbered by another load. See if
932 /// the other load can feed into the second load.
933 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
934 LoadInst *DepLI, const DataLayout &TD){
935 // Cannot handle reading from store of first-class aggregate yet.
936 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
939 Value *DepPtr = DepLI->getPointerOperand();
940 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType());
941 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD);
942 if (R != -1) return R;
944 // If we have a load/load clobber an DepLI can be widened to cover this load,
945 // then we should widen it!
946 int64_t LoadOffs = 0;
947 const Value *LoadBase =
948 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &TD);
949 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
951 unsigned Size = MemoryDependenceAnalysis::
952 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD);
953 if (Size == 0) return -1;
955 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD);
960 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
962 const DataLayout &TD) {
963 // If the mem operation is a non-constant size, we can't handle it.
964 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
965 if (SizeCst == 0) return -1;
966 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
968 // If this is memset, we just need to see if the offset is valid in the size
970 if (MI->getIntrinsicID() == Intrinsic::memset)
971 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
974 // If we have a memcpy/memmove, the only case we can handle is if this is a
975 // copy from constant memory. In that case, we can read directly from the
977 MemTransferInst *MTI = cast<MemTransferInst>(MI);
979 Constant *Src = dyn_cast<Constant>(MTI->getSource());
980 if (Src == 0) return -1;
982 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
983 if (GV == 0 || !GV->isConstant()) return -1;
985 // See if the access is within the bounds of the transfer.
986 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
987 MI->getDest(), MemSizeInBits, TD);
991 // Otherwise, see if we can constant fold a load from the constant with the
992 // offset applied as appropriate.
993 Src = ConstantExpr::getBitCast(Src,
994 llvm::Type::getInt8PtrTy(Src->getContext()));
995 Constant *OffsetCst =
996 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
997 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
998 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
999 if (ConstantFoldLoadFromConstPtr(Src, &TD))
1005 /// GetStoreValueForLoad - This function is called when we have a
1006 /// memdep query of a load that ends up being a clobbering store. This means
1007 /// that the store provides bits used by the load but we the pointers don't
1008 /// mustalias. Check this case to see if there is anything more we can do
1009 /// before we give up.
1010 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1012 Instruction *InsertPt, const DataLayout &TD){
1013 LLVMContext &Ctx = SrcVal->getType()->getContext();
1015 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1016 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
1018 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1020 // Compute which bits of the stored value are being used by the load. Convert
1021 // to an integer type to start with.
1022 if (SrcVal->getType()->getScalarType()->isPointerTy())
1023 SrcVal = Builder.CreatePtrToInt(SrcVal,
1024 TD.getIntPtrType(SrcVal->getType()));
1025 if (!SrcVal->getType()->isIntegerTy())
1026 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1028 // Shift the bits to the least significant depending on endianness.
1030 if (TD.isLittleEndian())
1031 ShiftAmt = Offset*8;
1033 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1036 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1038 if (LoadSize != StoreSize)
1039 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1041 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1044 /// GetLoadValueForLoad - This function is called when we have a
1045 /// memdep query of a load that ends up being a clobbering load. This means
1046 /// that the load *may* provide bits used by the load but we can't be sure
1047 /// because the pointers don't mustalias. Check this case to see if there is
1048 /// anything more we can do before we give up.
1049 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1050 Type *LoadTy, Instruction *InsertPt,
1052 const DataLayout &TD = *gvn.getDataLayout();
1053 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1054 // widen SrcVal out to a larger load.
1055 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType());
1056 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
1057 if (Offset+LoadSize > SrcValSize) {
1058 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1059 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1060 // If we have a load/load clobber an DepLI can be widened to cover this
1061 // load, then we should widen it to the next power of 2 size big enough!
1062 unsigned NewLoadSize = Offset+LoadSize;
1063 if (!isPowerOf2_32(NewLoadSize))
1064 NewLoadSize = NextPowerOf2(NewLoadSize);
1066 Value *PtrVal = SrcVal->getPointerOperand();
1068 // Insert the new load after the old load. This ensures that subsequent
1069 // memdep queries will find the new load. We can't easily remove the old
1070 // load completely because it is already in the value numbering table.
1071 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1073 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1074 DestPTy = PointerType::get(DestPTy,
1075 cast<PointerType>(PtrVal->getType())->getAddressSpace());
1076 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1077 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1078 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1079 NewLoad->takeName(SrcVal);
1080 NewLoad->setAlignment(SrcVal->getAlignment());
1082 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1083 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1085 // Replace uses of the original load with the wider load. On a big endian
1086 // system, we need to shift down to get the relevant bits.
1087 Value *RV = NewLoad;
1088 if (TD.isBigEndian())
1089 RV = Builder.CreateLShr(RV,
1090 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1091 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1092 SrcVal->replaceAllUsesWith(RV);
1094 // We would like to use gvn.markInstructionForDeletion here, but we can't
1095 // because the load is already memoized into the leader map table that GVN
1096 // tracks. It is potentially possible to remove the load from the table,
1097 // but then there all of the operations based on it would need to be
1098 // rehashed. Just leave the dead load around.
1099 gvn.getMemDep().removeInstruction(SrcVal);
1103 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD);
1107 /// GetMemInstValueForLoad - This function is called when we have a
1108 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1109 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1110 Type *LoadTy, Instruction *InsertPt,
1111 const DataLayout &TD){
1112 LLVMContext &Ctx = LoadTy->getContext();
1113 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1115 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1117 // We know that this method is only called when the mem transfer fully
1118 // provides the bits for the load.
1119 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1120 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1121 // independently of what the offset is.
1122 Value *Val = MSI->getValue();
1124 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1126 Value *OneElt = Val;
1128 // Splat the value out to the right number of bits.
1129 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1130 // If we can double the number of bytes set, do it.
1131 if (NumBytesSet*2 <= LoadSize) {
1132 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1133 Val = Builder.CreateOr(Val, ShVal);
1138 // Otherwise insert one byte at a time.
1139 Value *ShVal = Builder.CreateShl(Val, 1*8);
1140 Val = Builder.CreateOr(OneElt, ShVal);
1144 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1147 // Otherwise, this is a memcpy/memmove from a constant global.
1148 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1149 Constant *Src = cast<Constant>(MTI->getSource());
1151 // Otherwise, see if we can constant fold a load from the constant with the
1152 // offset applied as appropriate.
1153 Src = ConstantExpr::getBitCast(Src,
1154 llvm::Type::getInt8PtrTy(Src->getContext()));
1155 Constant *OffsetCst =
1156 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1157 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1158 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1159 return ConstantFoldLoadFromConstPtr(Src, &TD);
1164 struct AvailableValueInBlock {
1165 /// BB - The basic block in question.
1168 SimpleVal, // A simple offsetted value that is accessed.
1169 LoadVal, // A value produced by a load.
1170 MemIntrin // A memory intrinsic which is loaded from.
1173 /// V - The value that is live out of the block.
1174 PointerIntPair<Value *, 2, ValType> Val;
1176 /// Offset - The byte offset in Val that is interesting for the load query.
1179 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1180 unsigned Offset = 0) {
1181 AvailableValueInBlock Res;
1183 Res.Val.setPointer(V);
1184 Res.Val.setInt(SimpleVal);
1185 Res.Offset = Offset;
1189 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1190 unsigned Offset = 0) {
1191 AvailableValueInBlock Res;
1193 Res.Val.setPointer(MI);
1194 Res.Val.setInt(MemIntrin);
1195 Res.Offset = Offset;
1199 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
1200 unsigned Offset = 0) {
1201 AvailableValueInBlock Res;
1203 Res.Val.setPointer(LI);
1204 Res.Val.setInt(LoadVal);
1205 Res.Offset = Offset;
1209 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
1210 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
1211 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
1213 Value *getSimpleValue() const {
1214 assert(isSimpleValue() && "Wrong accessor");
1215 return Val.getPointer();
1218 LoadInst *getCoercedLoadValue() const {
1219 assert(isCoercedLoadValue() && "Wrong accessor");
1220 return cast<LoadInst>(Val.getPointer());
1223 MemIntrinsic *getMemIntrinValue() const {
1224 assert(isMemIntrinValue() && "Wrong accessor");
1225 return cast<MemIntrinsic>(Val.getPointer());
1228 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
1229 /// defined here to the specified type. This handles various coercion cases.
1230 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1232 if (isSimpleValue()) {
1233 Res = getSimpleValue();
1234 if (Res->getType() != LoadTy) {
1235 const DataLayout *TD = gvn.getDataLayout();
1236 assert(TD && "Need target data to handle type mismatch case");
1237 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1240 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1241 << *getSimpleValue() << '\n'
1242 << *Res << '\n' << "\n\n\n");
1244 } else if (isCoercedLoadValue()) {
1245 LoadInst *Load = getCoercedLoadValue();
1246 if (Load->getType() == LoadTy && Offset == 0) {
1249 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1252 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1253 << *getCoercedLoadValue() << '\n'
1254 << *Res << '\n' << "\n\n\n");
1257 const DataLayout *TD = gvn.getDataLayout();
1258 assert(TD && "Need target data to handle type mismatch case");
1259 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1260 LoadTy, BB->getTerminator(), *TD);
1261 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1262 << " " << *getMemIntrinValue() << '\n'
1263 << *Res << '\n' << "\n\n\n");
1269 } // end anonymous namespace
1271 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1272 /// construct SSA form, allowing us to eliminate LI. This returns the value
1273 /// that should be used at LI's definition site.
1274 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1275 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1277 // Check for the fully redundant, dominating load case. In this case, we can
1278 // just use the dominating value directly.
1279 if (ValuesPerBlock.size() == 1 &&
1280 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1282 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1284 // Otherwise, we have to construct SSA form.
1285 SmallVector<PHINode*, 8> NewPHIs;
1286 SSAUpdater SSAUpdate(&NewPHIs);
1287 SSAUpdate.Initialize(LI->getType(), LI->getName());
1289 Type *LoadTy = LI->getType();
1291 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1292 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1293 BasicBlock *BB = AV.BB;
1295 if (SSAUpdate.HasValueForBlock(BB))
1298 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1301 // Perform PHI construction.
1302 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1304 // If new PHI nodes were created, notify alias analysis.
1305 if (V->getType()->getScalarType()->isPointerTy()) {
1306 AliasAnalysis *AA = gvn.getAliasAnalysis();
1308 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1309 AA->copyValue(LI, NewPHIs[i]);
1311 // Now that we've copied information to the new PHIs, scan through
1312 // them again and inform alias analysis that we've added potentially
1313 // escaping uses to any values that are operands to these PHIs.
1314 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1315 PHINode *P = NewPHIs[i];
1316 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1317 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1318 AA->addEscapingUse(P->getOperandUse(jj));
1326 static bool isLifetimeStart(const Instruction *Inst) {
1327 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1328 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1332 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1333 /// non-local by performing PHI construction.
1334 bool GVN::processNonLocalLoad(LoadInst *LI) {
1335 // Find the non-local dependencies of the load.
1336 SmallVector<NonLocalDepResult, 64> Deps;
1337 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1338 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1339 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
1340 // << Deps.size() << *LI << '\n');
1342 // If we had to process more than one hundred blocks to find the
1343 // dependencies, this load isn't worth worrying about. Optimizing
1344 // it will be too expensive.
1345 unsigned NumDeps = Deps.size();
1349 // If we had a phi translation failure, we'll have a single entry which is a
1350 // clobber in the current block. Reject this early.
1352 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1354 dbgs() << "GVN: non-local load ";
1355 WriteAsOperand(dbgs(), LI);
1356 dbgs() << " has unknown dependencies\n";
1361 // Filter out useless results (non-locals, etc). Keep track of the blocks
1362 // where we have a value available in repl, also keep track of whether we see
1363 // dependencies that produce an unknown value for the load (such as a call
1364 // that could potentially clobber the load).
1365 SmallVector<AvailableValueInBlock, 64> ValuesPerBlock;
1366 SmallVector<BasicBlock*, 64> UnavailableBlocks;
1368 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1369 BasicBlock *DepBB = Deps[i].getBB();
1370 MemDepResult DepInfo = Deps[i].getResult();
1372 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1373 UnavailableBlocks.push_back(DepBB);
1377 if (DepInfo.isClobber()) {
1378 // The address being loaded in this non-local block may not be the same as
1379 // the pointer operand of the load if PHI translation occurs. Make sure
1380 // to consider the right address.
1381 Value *Address = Deps[i].getAddress();
1383 // If the dependence is to a store that writes to a superset of the bits
1384 // read by the load, we can extract the bits we need for the load from the
1386 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1387 if (TD && Address) {
1388 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1391 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1392 DepSI->getValueOperand(),
1399 // Check to see if we have something like this:
1402 // if we have this, replace the later with an extraction from the former.
1403 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1404 // If this is a clobber and L is the first instruction in its block, then
1405 // we have the first instruction in the entry block.
1406 if (DepLI != LI && Address && TD) {
1407 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
1408 LI->getPointerOperand(),
1412 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1419 // If the clobbering value is a memset/memcpy/memmove, see if we can
1420 // forward a value on from it.
1421 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1422 if (TD && Address) {
1423 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1426 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1433 UnavailableBlocks.push_back(DepBB);
1437 // DepInfo.isDef() here
1439 Instruction *DepInst = DepInfo.getInst();
1441 // Loading the allocation -> undef.
1442 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1443 // Loading immediately after lifetime begin -> undef.
1444 isLifetimeStart(DepInst)) {
1445 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1446 UndefValue::get(LI->getType())));
1450 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1451 // Reject loads and stores that are to the same address but are of
1452 // different types if we have to.
1453 if (S->getValueOperand()->getType() != LI->getType()) {
1454 // If the stored value is larger or equal to the loaded value, we can
1456 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1457 LI->getType(), *TD)) {
1458 UnavailableBlocks.push_back(DepBB);
1463 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1464 S->getValueOperand()));
1468 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1469 // If the types mismatch and we can't handle it, reject reuse of the load.
1470 if (LD->getType() != LI->getType()) {
1471 // If the stored value is larger or equal to the loaded value, we can
1473 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1474 UnavailableBlocks.push_back(DepBB);
1478 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1482 UnavailableBlocks.push_back(DepBB);
1486 // If we have no predecessors that produce a known value for this load, exit
1488 if (ValuesPerBlock.empty()) return false;
1490 // If all of the instructions we depend on produce a known value for this
1491 // load, then it is fully redundant and we can use PHI insertion to compute
1492 // its value. Insert PHIs and remove the fully redundant value now.
1493 if (UnavailableBlocks.empty()) {
1494 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1496 // Perform PHI construction.
1497 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1498 LI->replaceAllUsesWith(V);
1500 if (isa<PHINode>(V))
1502 if (V->getType()->getScalarType()->isPointerTy())
1503 MD->invalidateCachedPointerInfo(V);
1504 markInstructionForDeletion(LI);
1509 if (!EnablePRE || !EnableLoadPRE)
1512 // Okay, we have *some* definitions of the value. This means that the value
1513 // is available in some of our (transitive) predecessors. Lets think about
1514 // doing PRE of this load. This will involve inserting a new load into the
1515 // predecessor when it's not available. We could do this in general, but
1516 // prefer to not increase code size. As such, we only do this when we know
1517 // that we only have to insert *one* load (which means we're basically moving
1518 // the load, not inserting a new one).
1520 SmallPtrSet<BasicBlock *, 4> Blockers;
1521 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1522 Blockers.insert(UnavailableBlocks[i]);
1524 // Let's find the first basic block with more than one predecessor. Walk
1525 // backwards through predecessors if needed.
1526 BasicBlock *LoadBB = LI->getParent();
1527 BasicBlock *TmpBB = LoadBB;
1529 bool isSinglePred = false;
1530 bool allSingleSucc = true;
1531 while (TmpBB->getSinglePredecessor()) {
1532 isSinglePred = true;
1533 TmpBB = TmpBB->getSinglePredecessor();
1534 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1536 if (Blockers.count(TmpBB))
1539 // If any of these blocks has more than one successor (i.e. if the edge we
1540 // just traversed was critical), then there are other paths through this
1541 // block along which the load may not be anticipated. Hoisting the load
1542 // above this block would be adding the load to execution paths along
1543 // which it was not previously executed.
1544 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1551 // FIXME: It is extremely unclear what this loop is doing, other than
1552 // artificially restricting loadpre.
1555 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1556 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1557 if (AV.isSimpleValue())
1558 // "Hot" Instruction is in some loop (because it dominates its dep.
1560 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1561 if (DT->dominates(LI, I)) {
1567 // We are interested only in "hot" instructions. We don't want to do any
1568 // mis-optimizations here.
1573 // Check to see how many predecessors have the loaded value fully
1575 DenseMap<BasicBlock*, Value*> PredLoads;
1576 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1577 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1578 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1579 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1580 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1582 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
1583 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1585 BasicBlock *Pred = *PI;
1586 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1589 PredLoads[Pred] = 0;
1591 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1592 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1593 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1594 << Pred->getName() << "': " << *LI << '\n');
1598 if (LoadBB->isLandingPad()) {
1600 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1601 << Pred->getName() << "': " << *LI << '\n');
1605 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
1606 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
1610 if (!NeedToSplit.empty()) {
1611 toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
1615 // Decide whether PRE is profitable for this load.
1616 unsigned NumUnavailablePreds = PredLoads.size();
1617 assert(NumUnavailablePreds != 0 &&
1618 "Fully available value should be eliminated above!");
1620 // If this load is unavailable in multiple predecessors, reject it.
1621 // FIXME: If we could restructure the CFG, we could make a common pred with
1622 // all the preds that don't have an available LI and insert a new load into
1624 if (NumUnavailablePreds != 1)
1627 // Check if the load can safely be moved to all the unavailable predecessors.
1628 bool CanDoPRE = true;
1629 SmallVector<Instruction*, 8> NewInsts;
1630 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1631 E = PredLoads.end(); I != E; ++I) {
1632 BasicBlock *UnavailablePred = I->first;
1634 // Do PHI translation to get its value in the predecessor if necessary. The
1635 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1637 // If all preds have a single successor, then we know it is safe to insert
1638 // the load on the pred (?!?), so we can insert code to materialize the
1639 // pointer if it is not available.
1640 PHITransAddr Address(LI->getPointerOperand(), TD);
1642 if (allSingleSucc) {
1643 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1646 Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
1647 LoadPtr = Address.getAddr();
1650 // If we couldn't find or insert a computation of this phi translated value,
1653 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1654 << *LI->getPointerOperand() << "\n");
1659 // Make sure it is valid to move this load here. We have to watch out for:
1660 // @1 = getelementptr (i8* p, ...
1661 // test p and branch if == 0
1663 // It is valid to have the getelementptr before the test, even if p can
1664 // be 0, as getelementptr only does address arithmetic.
1665 // If we are not pushing the value through any multiple-successor blocks
1666 // we do not have this case. Otherwise, check that the load is safe to
1667 // put anywhere; this can be improved, but should be conservatively safe.
1668 if (!allSingleSucc &&
1669 // FIXME: REEVALUTE THIS.
1670 !isSafeToLoadUnconditionally(LoadPtr,
1671 UnavailablePred->getTerminator(),
1672 LI->getAlignment(), TD)) {
1677 I->second = LoadPtr;
1681 while (!NewInsts.empty()) {
1682 Instruction *I = NewInsts.pop_back_val();
1683 if (MD) MD->removeInstruction(I);
1684 I->eraseFromParent();
1689 // Okay, we can eliminate this load by inserting a reload in the predecessor
1690 // and using PHI construction to get the value in the other predecessors, do
1692 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1693 DEBUG(if (!NewInsts.empty())
1694 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1695 << *NewInsts.back() << '\n');
1697 // Assign value numbers to the new instructions.
1698 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1699 // FIXME: We really _ought_ to insert these value numbers into their
1700 // parent's availability map. However, in doing so, we risk getting into
1701 // ordering issues. If a block hasn't been processed yet, we would be
1702 // marking a value as AVAIL-IN, which isn't what we intend.
1703 VN.lookup_or_add(NewInsts[i]);
1706 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1707 E = PredLoads.end(); I != E; ++I) {
1708 BasicBlock *UnavailablePred = I->first;
1709 Value *LoadPtr = I->second;
1711 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1713 UnavailablePred->getTerminator());
1715 // Transfer the old load's TBAA tag to the new load.
1716 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1717 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1719 // Transfer DebugLoc.
1720 NewLoad->setDebugLoc(LI->getDebugLoc());
1722 // Add the newly created load.
1723 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1725 MD->invalidateCachedPointerInfo(LoadPtr);
1726 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1729 // Perform PHI construction.
1730 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1731 LI->replaceAllUsesWith(V);
1732 if (isa<PHINode>(V))
1734 if (V->getType()->getScalarType()->isPointerTy())
1735 MD->invalidateCachedPointerInfo(V);
1736 markInstructionForDeletion(LI);
1741 static void patchReplacementInstruction(Value *Repl, Instruction *I) {
1742 // Patch the replacement so that it is not more restrictive than the value
1744 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1745 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1746 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) &&
1747 isa<OverflowingBinaryOperator>(ReplOp)) {
1748 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap())
1749 ReplOp->setHasNoSignedWrap(false);
1750 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap())
1751 ReplOp->setHasNoUnsignedWrap(false);
1753 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1754 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
1755 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata);
1756 for (int i = 0, n = Metadata.size(); i < n; ++i) {
1757 unsigned Kind = Metadata[i].first;
1758 MDNode *IMD = I->getMetadata(Kind);
1759 MDNode *ReplMD = Metadata[i].second;
1762 ReplInst->setMetadata(Kind, NULL); // Remove unknown metadata
1764 case LLVMContext::MD_dbg:
1765 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1766 case LLVMContext::MD_tbaa:
1767 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD));
1769 case LLVMContext::MD_range:
1770 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD));
1772 case LLVMContext::MD_prof:
1773 llvm_unreachable("MD_prof in a non terminator instruction");
1775 case LLVMContext::MD_fpmath:
1776 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD));
1783 static void patchAndReplaceAllUsesWith(Value *Repl, Instruction *I) {
1784 patchReplacementInstruction(Repl, I);
1785 I->replaceAllUsesWith(Repl);
1788 /// processLoad - Attempt to eliminate a load, first by eliminating it
1789 /// locally, and then attempting non-local elimination if that fails.
1790 bool GVN::processLoad(LoadInst *L) {
1797 if (L->use_empty()) {
1798 markInstructionForDeletion(L);
1802 // ... to a pointer that has been loaded from before...
1803 MemDepResult Dep = MD->getDependency(L);
1805 // If we have a clobber and target data is around, see if this is a clobber
1806 // that we can fix up through code synthesis.
1807 if (Dep.isClobber() && TD) {
1808 // Check to see if we have something like this:
1809 // store i32 123, i32* %P
1810 // %A = bitcast i32* %P to i8*
1811 // %B = gep i8* %A, i32 1
1814 // We could do that by recognizing if the clobber instructions are obviously
1815 // a common base + constant offset, and if the previous store (or memset)
1816 // completely covers this load. This sort of thing can happen in bitfield
1818 Value *AvailVal = 0;
1819 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1820 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1821 L->getPointerOperand(),
1824 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1825 L->getType(), L, *TD);
1828 // Check to see if we have something like this:
1831 // if we have this, replace the later with an extraction from the former.
1832 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1833 // If this is a clobber and L is the first instruction in its block, then
1834 // we have the first instruction in the entry block.
1838 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1839 L->getPointerOperand(),
1842 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1845 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1846 // a value on from it.
1847 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1848 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1849 L->getPointerOperand(),
1852 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
1856 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1857 << *AvailVal << '\n' << *L << "\n\n\n");
1859 // Replace the load!
1860 L->replaceAllUsesWith(AvailVal);
1861 if (AvailVal->getType()->getScalarType()->isPointerTy())
1862 MD->invalidateCachedPointerInfo(AvailVal);
1863 markInstructionForDeletion(L);
1869 // If the value isn't available, don't do anything!
1870 if (Dep.isClobber()) {
1872 // fast print dep, using operator<< on instruction is too slow.
1873 dbgs() << "GVN: load ";
1874 WriteAsOperand(dbgs(), L);
1875 Instruction *I = Dep.getInst();
1876 dbgs() << " is clobbered by " << *I << '\n';
1881 // If it is defined in another block, try harder.
1882 if (Dep.isNonLocal())
1883 return processNonLocalLoad(L);
1887 // fast print dep, using operator<< on instruction is too slow.
1888 dbgs() << "GVN: load ";
1889 WriteAsOperand(dbgs(), L);
1890 dbgs() << " has unknown dependence\n";
1895 Instruction *DepInst = Dep.getInst();
1896 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1897 Value *StoredVal = DepSI->getValueOperand();
1899 // The store and load are to a must-aliased pointer, but they may not
1900 // actually have the same type. See if we know how to reuse the stored
1901 // value (depending on its type).
1902 if (StoredVal->getType() != L->getType()) {
1904 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1909 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1910 << '\n' << *L << "\n\n\n");
1917 L->replaceAllUsesWith(StoredVal);
1918 if (StoredVal->getType()->getScalarType()->isPointerTy())
1919 MD->invalidateCachedPointerInfo(StoredVal);
1920 markInstructionForDeletion(L);
1925 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1926 Value *AvailableVal = DepLI;
1928 // The loads are of a must-aliased pointer, but they may not actually have
1929 // the same type. See if we know how to reuse the previously loaded value
1930 // (depending on its type).
1931 if (DepLI->getType() != L->getType()) {
1933 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1935 if (AvailableVal == 0)
1938 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1939 << "\n" << *L << "\n\n\n");
1946 patchAndReplaceAllUsesWith(AvailableVal, L);
1947 if (DepLI->getType()->getScalarType()->isPointerTy())
1948 MD->invalidateCachedPointerInfo(DepLI);
1949 markInstructionForDeletion(L);
1954 // If this load really doesn't depend on anything, then we must be loading an
1955 // undef value. This can happen when loading for a fresh allocation with no
1956 // intervening stores, for example.
1957 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1958 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1959 markInstructionForDeletion(L);
1964 // If this load occurs either right after a lifetime begin,
1965 // then the loaded value is undefined.
1966 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1967 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1968 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1969 markInstructionForDeletion(L);
1978 // findLeader - In order to find a leader for a given value number at a
1979 // specific basic block, we first obtain the list of all Values for that number,
1980 // and then scan the list to find one whose block dominates the block in
1981 // question. This is fast because dominator tree queries consist of only
1982 // a few comparisons of DFS numbers.
1983 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1984 LeaderTableEntry Vals = LeaderTable[num];
1985 if (!Vals.Val) return 0;
1988 if (DT->dominates(Vals.BB, BB)) {
1990 if (isa<Constant>(Val)) return Val;
1993 LeaderTableEntry* Next = Vals.Next;
1995 if (DT->dominates(Next->BB, BB)) {
1996 if (isa<Constant>(Next->Val)) return Next->Val;
1997 if (!Val) Val = Next->Val;
2006 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
2007 /// use is dominated by the given basic block. Returns the number of uses that
2009 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
2010 const BasicBlockEdge &Root) {
2012 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2014 Use &U = (UI++).getUse();
2016 if (DT->dominates(Root, U)) {
2024 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
2025 /// true if every path from the entry block to 'Dst' passes via this edge. In
2026 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2027 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2028 DominatorTree *DT) {
2029 // While in theory it is interesting to consider the case in which Dst has
2030 // more than one predecessor, because Dst might be part of a loop which is
2031 // only reachable from Src, in practice it is pointless since at the time
2032 // GVN runs all such loops have preheaders, which means that Dst will have
2033 // been changed to have only one predecessor, namely Src.
2034 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2035 const BasicBlock *Src = E.getStart();
2036 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2041 /// propagateEquality - The given values are known to be equal in every block
2042 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2043 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2044 bool GVN::propagateEquality(Value *LHS, Value *RHS,
2045 const BasicBlockEdge &Root) {
2046 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2047 Worklist.push_back(std::make_pair(LHS, RHS));
2048 bool Changed = false;
2049 // For speed, compute a conservative fast approximation to
2050 // DT->dominates(Root, Root.getEnd());
2051 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2053 while (!Worklist.empty()) {
2054 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2055 LHS = Item.first; RHS = Item.second;
2057 if (LHS == RHS) continue;
2058 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2060 // Don't try to propagate equalities between constants.
2061 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2063 // Prefer a constant on the right-hand side, or an Argument if no constants.
2064 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2065 std::swap(LHS, RHS);
2066 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2068 // If there is no obvious reason to prefer the left-hand side over the right-
2069 // hand side, ensure the longest lived term is on the right-hand side, so the
2070 // shortest lived term will be replaced by the longest lived. This tends to
2071 // expose more simplifications.
2072 uint32_t LVN = VN.lookup_or_add(LHS);
2073 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2074 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2075 // Move the 'oldest' value to the right-hand side, using the value number as
2077 uint32_t RVN = VN.lookup_or_add(RHS);
2079 std::swap(LHS, RHS);
2084 // If value numbering later sees that an instruction in the scope is equal
2085 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2086 // the invariant that instructions only occur in the leader table for their
2087 // own value number (this is used by removeFromLeaderTable), do not do this
2088 // if RHS is an instruction (if an instruction in the scope is morphed into
2089 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2090 // using the leader table is about compiling faster, not optimizing better).
2091 // The leader table only tracks basic blocks, not edges. Only add to if we
2092 // have the simple case where the edge dominates the end.
2093 if (RootDominatesEnd && !isa<Instruction>(RHS))
2094 addToLeaderTable(LVN, RHS, Root.getEnd());
2096 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2097 // LHS always has at least one use that is not dominated by Root, this will
2098 // never do anything if LHS has only one use.
2099 if (!LHS->hasOneUse()) {
2100 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2101 Changed |= NumReplacements > 0;
2102 NumGVNEqProp += NumReplacements;
2105 // Now try to deduce additional equalities from this one. For example, if the
2106 // known equality was "(A != B)" == "false" then it follows that A and B are
2107 // equal in the scope. Only boolean equalities with an explicit true or false
2108 // RHS are currently supported.
2109 if (!RHS->getType()->isIntegerTy(1))
2110 // Not a boolean equality - bail out.
2112 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2114 // RHS neither 'true' nor 'false' - bail out.
2116 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2117 bool isKnownTrue = CI->isAllOnesValue();
2118 bool isKnownFalse = !isKnownTrue;
2120 // If "A && B" is known true then both A and B are known true. If "A || B"
2121 // is known false then both A and B are known false.
2123 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2124 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2125 Worklist.push_back(std::make_pair(A, RHS));
2126 Worklist.push_back(std::make_pair(B, RHS));
2130 // If we are propagating an equality like "(A == B)" == "true" then also
2131 // propagate the equality A == B. When propagating a comparison such as
2132 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2133 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
2134 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2136 // If "A == B" is known true, or "A != B" is known false, then replace
2137 // A with B everywhere in the scope.
2138 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2139 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2140 Worklist.push_back(std::make_pair(Op0, Op1));
2142 // If "A >= B" is known true, replace "A < B" with false everywhere.
2143 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2144 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2145 // Since we don't have the instruction "A < B" immediately to hand, work out
2146 // the value number that it would have and use that to find an appropriate
2147 // instruction (if any).
2148 uint32_t NextNum = VN.getNextUnusedValueNumber();
2149 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2150 // If the number we were assigned was brand new then there is no point in
2151 // looking for an instruction realizing it: there cannot be one!
2152 if (Num < NextNum) {
2153 Value *NotCmp = findLeader(Root.getEnd(), Num);
2154 if (NotCmp && isa<Instruction>(NotCmp)) {
2155 unsigned NumReplacements =
2156 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2157 Changed |= NumReplacements > 0;
2158 NumGVNEqProp += NumReplacements;
2161 // Ensure that any instruction in scope that gets the "A < B" value number
2162 // is replaced with false.
2163 // The leader table only tracks basic blocks, not edges. Only add to if we
2164 // have the simple case where the edge dominates the end.
2165 if (RootDominatesEnd)
2166 addToLeaderTable(Num, NotVal, Root.getEnd());
2175 /// processInstruction - When calculating availability, handle an instruction
2176 /// by inserting it into the appropriate sets
2177 bool GVN::processInstruction(Instruction *I) {
2178 // Ignore dbg info intrinsics.
2179 if (isa<DbgInfoIntrinsic>(I))
2182 // If the instruction can be easily simplified then do so now in preference
2183 // to value numbering it. Value numbering often exposes redundancies, for
2184 // example if it determines that %y is equal to %x then the instruction
2185 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2186 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) {
2187 I->replaceAllUsesWith(V);
2188 if (MD && V->getType()->getScalarType()->isPointerTy())
2189 MD->invalidateCachedPointerInfo(V);
2190 markInstructionForDeletion(I);
2195 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2196 if (processLoad(LI))
2199 unsigned Num = VN.lookup_or_add(LI);
2200 addToLeaderTable(Num, LI, LI->getParent());
2204 // For conditional branches, we can perform simple conditional propagation on
2205 // the condition value itself.
2206 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2207 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
2210 Value *BranchCond = BI->getCondition();
2212 BasicBlock *TrueSucc = BI->getSuccessor(0);
2213 BasicBlock *FalseSucc = BI->getSuccessor(1);
2214 // Avoid multiple edges early.
2215 if (TrueSucc == FalseSucc)
2218 BasicBlock *Parent = BI->getParent();
2219 bool Changed = false;
2221 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2222 BasicBlockEdge TrueE(Parent, TrueSucc);
2223 Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
2225 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2226 BasicBlockEdge FalseE(Parent, FalseSucc);
2227 Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
2232 // For switches, propagate the case values into the case destinations.
2233 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2234 Value *SwitchCond = SI->getCondition();
2235 BasicBlock *Parent = SI->getParent();
2236 bool Changed = false;
2238 // Remember how many outgoing edges there are to every successor.
2239 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2240 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2241 ++SwitchEdges[SI->getSuccessor(i)];
2243 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2245 BasicBlock *Dst = i.getCaseSuccessor();
2246 // If there is only a single edge, propagate the case value into it.
2247 if (SwitchEdges.lookup(Dst) == 1) {
2248 BasicBlockEdge E(Parent, Dst);
2249 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
2255 // Instructions with void type don't return a value, so there's
2256 // no point in trying to find redundancies in them.
2257 if (I->getType()->isVoidTy()) return false;
2259 uint32_t NextNum = VN.getNextUnusedValueNumber();
2260 unsigned Num = VN.lookup_or_add(I);
2262 // Allocations are always uniquely numbered, so we can save time and memory
2263 // by fast failing them.
2264 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2265 addToLeaderTable(Num, I, I->getParent());
2269 // If the number we were assigned was a brand new VN, then we don't
2270 // need to do a lookup to see if the number already exists
2271 // somewhere in the domtree: it can't!
2272 if (Num >= NextNum) {
2273 addToLeaderTable(Num, I, I->getParent());
2277 // Perform fast-path value-number based elimination of values inherited from
2279 Value *repl = findLeader(I->getParent(), Num);
2281 // Failure, just remember this instance for future use.
2282 addToLeaderTable(Num, I, I->getParent());
2287 patchAndReplaceAllUsesWith(repl, I);
2288 if (MD && repl->getType()->getScalarType()->isPointerTy())
2289 MD->invalidateCachedPointerInfo(repl);
2290 markInstructionForDeletion(I);
2294 /// runOnFunction - This is the main transformation entry point for a function.
2295 bool GVN::runOnFunction(Function& F) {
2297 MD = &getAnalysis<MemoryDependenceAnalysis>();
2298 DT = &getAnalysis<DominatorTree>();
2299 TD = getAnalysisIfAvailable<DataLayout>();
2300 TLI = &getAnalysis<TargetLibraryInfo>();
2301 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2305 bool Changed = false;
2306 bool ShouldContinue = true;
2308 // Merge unconditional branches, allowing PRE to catch more
2309 // optimization opportunities.
2310 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2311 BasicBlock *BB = FI++;
2313 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2314 if (removedBlock) ++NumGVNBlocks;
2316 Changed |= removedBlock;
2319 unsigned Iteration = 0;
2320 while (ShouldContinue) {
2321 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2322 ShouldContinue = iterateOnFunction(F);
2323 if (splitCriticalEdges())
2324 ShouldContinue = true;
2325 Changed |= ShouldContinue;
2330 bool PREChanged = true;
2331 while (PREChanged) {
2332 PREChanged = performPRE(F);
2333 Changed |= PREChanged;
2336 // FIXME: Should perform GVN again after PRE does something. PRE can move
2337 // computations into blocks where they become fully redundant. Note that
2338 // we can't do this until PRE's critical edge splitting updates memdep.
2339 // Actually, when this happens, we should just fully integrate PRE into GVN.
2341 cleanupGlobalSets();
2347 bool GVN::processBlock(BasicBlock *BB) {
2348 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2349 // (and incrementing BI before processing an instruction).
2350 assert(InstrsToErase.empty() &&
2351 "We expect InstrsToErase to be empty across iterations");
2352 bool ChangedFunction = false;
2354 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2356 ChangedFunction |= processInstruction(BI);
2357 if (InstrsToErase.empty()) {
2362 // If we need some instructions deleted, do it now.
2363 NumGVNInstr += InstrsToErase.size();
2365 // Avoid iterator invalidation.
2366 bool AtStart = BI == BB->begin();
2370 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(),
2371 E = InstrsToErase.end(); I != E; ++I) {
2372 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2373 if (MD) MD->removeInstruction(*I);
2374 (*I)->eraseFromParent();
2375 DEBUG(verifyRemoved(*I));
2377 InstrsToErase.clear();
2385 return ChangedFunction;
2388 /// performPRE - Perform a purely local form of PRE that looks for diamond
2389 /// control flow patterns and attempts to perform simple PRE at the join point.
2390 bool GVN::performPRE(Function &F) {
2391 bool Changed = false;
2392 DenseMap<BasicBlock*, Value*> predMap;
2393 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2394 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2395 BasicBlock *CurrentBlock = *DI;
2397 // Nothing to PRE in the entry block.
2398 if (CurrentBlock == &F.getEntryBlock()) continue;
2400 // Don't perform PRE on a landing pad.
2401 if (CurrentBlock->isLandingPad()) continue;
2403 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2404 BE = CurrentBlock->end(); BI != BE; ) {
2405 Instruction *CurInst = BI++;
2407 if (isa<AllocaInst>(CurInst) ||
2408 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2409 CurInst->getType()->isVoidTy() ||
2410 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2411 isa<DbgInfoIntrinsic>(CurInst))
2414 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2415 // sinking the compare again, and it would force the code generator to
2416 // move the i1 from processor flags or predicate registers into a general
2417 // purpose register.
2418 if (isa<CmpInst>(CurInst))
2421 // We don't currently value number ANY inline asm calls.
2422 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2423 if (CallI->isInlineAsm())
2426 uint32_t ValNo = VN.lookup(CurInst);
2428 // Look for the predecessors for PRE opportunities. We're
2429 // only trying to solve the basic diamond case, where
2430 // a value is computed in the successor and one predecessor,
2431 // but not the other. We also explicitly disallow cases
2432 // where the successor is its own predecessor, because they're
2433 // more complicated to get right.
2434 unsigned NumWith = 0;
2435 unsigned NumWithout = 0;
2436 BasicBlock *PREPred = 0;
2439 for (pred_iterator PI = pred_begin(CurrentBlock),
2440 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2441 BasicBlock *P = *PI;
2442 // We're not interested in PRE where the block is its
2443 // own predecessor, or in blocks with predecessors
2444 // that are not reachable.
2445 if (P == CurrentBlock) {
2448 } else if (!DT->dominates(&F.getEntryBlock(), P)) {
2453 Value* predV = findLeader(P, ValNo);
2457 } else if (predV == CurInst) {
2465 // Don't do PRE when it might increase code size, i.e. when
2466 // we would need to insert instructions in more than one pred.
2467 if (NumWithout != 1 || NumWith == 0)
2470 // Don't do PRE across indirect branch.
2471 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2474 // We can't do PRE safely on a critical edge, so instead we schedule
2475 // the edge to be split and perform the PRE the next time we iterate
2477 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2478 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2479 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2483 // Instantiate the expression in the predecessor that lacked it.
2484 // Because we are going top-down through the block, all value numbers
2485 // will be available in the predecessor by the time we need them. Any
2486 // that weren't originally present will have been instantiated earlier
2488 Instruction *PREInstr = CurInst->clone();
2489 bool success = true;
2490 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2491 Value *Op = PREInstr->getOperand(i);
2492 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2495 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2496 PREInstr->setOperand(i, V);
2503 // Fail out if we encounter an operand that is not available in
2504 // the PRE predecessor. This is typically because of loads which
2505 // are not value numbered precisely.
2508 DEBUG(verifyRemoved(PREInstr));
2512 PREInstr->insertBefore(PREPred->getTerminator());
2513 PREInstr->setName(CurInst->getName() + ".pre");
2514 PREInstr->setDebugLoc(CurInst->getDebugLoc());
2515 predMap[PREPred] = PREInstr;
2516 VN.add(PREInstr, ValNo);
2519 // Update the availability map to include the new instruction.
2520 addToLeaderTable(ValNo, PREInstr, PREPred);
2522 // Create a PHI to make the value available in this block.
2523 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2524 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE),
2525 CurInst->getName() + ".pre-phi",
2526 CurrentBlock->begin());
2527 for (pred_iterator PI = PB; PI != PE; ++PI) {
2528 BasicBlock *P = *PI;
2529 Phi->addIncoming(predMap[P], P);
2533 addToLeaderTable(ValNo, Phi, CurrentBlock);
2534 Phi->setDebugLoc(CurInst->getDebugLoc());
2535 CurInst->replaceAllUsesWith(Phi);
2536 if (Phi->getType()->getScalarType()->isPointerTy()) {
2537 // Because we have added a PHI-use of the pointer value, it has now
2538 // "escaped" from alias analysis' perspective. We need to inform
2540 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
2542 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2543 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2547 MD->invalidateCachedPointerInfo(Phi);
2550 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2552 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2553 if (MD) MD->removeInstruction(CurInst);
2554 CurInst->eraseFromParent();
2555 DEBUG(verifyRemoved(CurInst));
2560 if (splitCriticalEdges())
2566 /// splitCriticalEdges - Split critical edges found during the previous
2567 /// iteration that may enable further optimization.
2568 bool GVN::splitCriticalEdges() {
2569 if (toSplit.empty())
2572 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2573 SplitCriticalEdge(Edge.first, Edge.second, this);
2574 } while (!toSplit.empty());
2575 if (MD) MD->invalidateCachedPredecessors();
2579 /// iterateOnFunction - Executes one iteration of GVN
2580 bool GVN::iterateOnFunction(Function &F) {
2581 cleanupGlobalSets();
2583 // Top-down walk of the dominator tree
2584 bool Changed = false;
2586 // Needed for value numbering with phi construction to work.
2587 ReversePostOrderTraversal<Function*> RPOT(&F);
2588 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2589 RE = RPOT.end(); RI != RE; ++RI)
2590 Changed |= processBlock(*RI);
2592 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2593 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2594 Changed |= processBlock(DI->getBlock());
2600 void GVN::cleanupGlobalSets() {
2602 LeaderTable.clear();
2603 TableAllocator.Reset();
2606 /// verifyRemoved - Verify that the specified instruction does not occur in our
2607 /// internal data structures.
2608 void GVN::verifyRemoved(const Instruction *Inst) const {
2609 VN.verifyRemoved(Inst);
2611 // Walk through the value number scope to make sure the instruction isn't
2612 // ferreted away in it.
2613 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2614 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2615 const LeaderTableEntry *Node = &I->second;
2616 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2618 while (Node->Next) {
2620 assert(Node->Val != Inst && "Inst still in value numbering scope!");