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/SetVector.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/CFG.h"
28 #include "llvm/Analysis/ConstantFolding.h"
29 #include "llvm/Analysis/Dominators.h"
30 #include "llvm/Analysis/InstructionSimplify.h"
31 #include "llvm/Analysis/Loads.h"
32 #include "llvm/Analysis/MemoryBuiltins.h"
33 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
34 #include "llvm/Analysis/PHITransAddr.h"
35 #include "llvm/Analysis/ValueTracking.h"
36 #include "llvm/Assembly/Writer.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/GlobalVariable.h"
39 #include "llvm/IR/IRBuilder.h"
40 #include "llvm/IR/IntrinsicInst.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/IR/Metadata.h"
43 #include "llvm/Support/Allocator.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/PatternMatch.h"
47 #include "llvm/Target/TargetLibraryInfo.h"
48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
49 #include "llvm/Transforms/Utils/SSAUpdater.h"
52 using namespace PatternMatch;
54 STATISTIC(NumGVNInstr, "Number of instructions deleted");
55 STATISTIC(NumGVNLoad, "Number of loads deleted");
56 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
57 STATISTIC(NumGVNBlocks, "Number of blocks merged");
58 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
59 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
60 STATISTIC(NumPRELoad, "Number of loads PRE'd");
62 static cl::opt<bool> EnablePRE("enable-pre",
63 cl::init(true), cl::Hidden);
64 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
66 // Maximum allowed recursion depth.
67 static cl::opt<uint32_t>
68 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
69 cl::desc("Max recurse depth (default = 1000)"));
71 //===----------------------------------------------------------------------===//
73 //===----------------------------------------------------------------------===//
75 /// This class holds the mapping between values and value numbers. It is used
76 /// as an efficient mechanism to determine the expression-wise equivalence of
82 SmallVector<uint32_t, 4> varargs;
84 Expression(uint32_t o = ~2U) : opcode(o) { }
86 bool operator==(const Expression &other) const {
87 if (opcode != other.opcode)
89 if (opcode == ~0U || opcode == ~1U)
91 if (type != other.type)
93 if (varargs != other.varargs)
98 friend hash_code hash_value(const Expression &Value) {
99 return hash_combine(Value.opcode, Value.type,
100 hash_combine_range(Value.varargs.begin(),
101 Value.varargs.end()));
106 DenseMap<Value*, uint32_t> valueNumbering;
107 DenseMap<Expression, uint32_t> expressionNumbering;
109 MemoryDependenceAnalysis *MD;
112 uint32_t nextValueNumber;
114 Expression create_expression(Instruction* I);
115 Expression create_cmp_expression(unsigned Opcode,
116 CmpInst::Predicate Predicate,
117 Value *LHS, Value *RHS);
118 Expression create_extractvalue_expression(ExtractValueInst* EI);
119 uint32_t lookup_or_add_call(CallInst* C);
121 ValueTable() : nextValueNumber(1) { }
122 uint32_t lookup_or_add(Value *V);
123 uint32_t lookup(Value *V) const;
124 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
125 Value *LHS, Value *RHS);
126 void add(Value *V, uint32_t num);
128 void erase(Value *v);
129 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
130 AliasAnalysis *getAliasAnalysis() const { return AA; }
131 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
132 void setDomTree(DominatorTree* D) { DT = D; }
133 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
134 void verifyRemoved(const Value *) const;
139 template <> struct DenseMapInfo<Expression> {
140 static inline Expression getEmptyKey() {
144 static inline Expression getTombstoneKey() {
148 static unsigned getHashValue(const Expression e) {
149 using llvm::hash_value;
150 return static_cast<unsigned>(hash_value(e));
152 static bool isEqual(const Expression &LHS, const Expression &RHS) {
159 //===----------------------------------------------------------------------===//
160 // ValueTable Internal Functions
161 //===----------------------------------------------------------------------===//
163 Expression ValueTable::create_expression(Instruction *I) {
165 e.type = I->getType();
166 e.opcode = I->getOpcode();
167 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
169 e.varargs.push_back(lookup_or_add(*OI));
170 if (I->isCommutative()) {
171 // Ensure that commutative instructions that only differ by a permutation
172 // of their operands get the same value number by sorting the operand value
173 // numbers. Since all commutative instructions have two operands it is more
174 // efficient to sort by hand rather than using, say, std::sort.
175 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
176 if (e.varargs[0] > e.varargs[1])
177 std::swap(e.varargs[0], e.varargs[1]);
180 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
181 // Sort the operand value numbers so x<y and y>x get the same value number.
182 CmpInst::Predicate Predicate = C->getPredicate();
183 if (e.varargs[0] > e.varargs[1]) {
184 std::swap(e.varargs[0], e.varargs[1]);
185 Predicate = CmpInst::getSwappedPredicate(Predicate);
187 e.opcode = (C->getOpcode() << 8) | Predicate;
188 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
189 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
191 e.varargs.push_back(*II);
197 Expression ValueTable::create_cmp_expression(unsigned Opcode,
198 CmpInst::Predicate Predicate,
199 Value *LHS, Value *RHS) {
200 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
201 "Not a comparison!");
203 e.type = CmpInst::makeCmpResultType(LHS->getType());
204 e.varargs.push_back(lookup_or_add(LHS));
205 e.varargs.push_back(lookup_or_add(RHS));
207 // Sort the operand value numbers so x<y and y>x get the same value number.
208 if (e.varargs[0] > e.varargs[1]) {
209 std::swap(e.varargs[0], e.varargs[1]);
210 Predicate = CmpInst::getSwappedPredicate(Predicate);
212 e.opcode = (Opcode << 8) | Predicate;
216 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
217 assert(EI != 0 && "Not an ExtractValueInst?");
219 e.type = EI->getType();
222 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
223 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
224 // EI might be an extract from one of our recognised intrinsics. If it
225 // is we'll synthesize a semantically equivalent expression instead on
226 // an extract value expression.
227 switch (I->getIntrinsicID()) {
228 case Intrinsic::sadd_with_overflow:
229 case Intrinsic::uadd_with_overflow:
230 e.opcode = Instruction::Add;
232 case Intrinsic::ssub_with_overflow:
233 case Intrinsic::usub_with_overflow:
234 e.opcode = Instruction::Sub;
236 case Intrinsic::smul_with_overflow:
237 case Intrinsic::umul_with_overflow:
238 e.opcode = Instruction::Mul;
245 // Intrinsic recognized. Grab its args to finish building the expression.
246 assert(I->getNumArgOperands() == 2 &&
247 "Expect two args for recognised intrinsics.");
248 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
249 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
254 // Not a recognised intrinsic. Fall back to producing an extract value
256 e.opcode = EI->getOpcode();
257 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
259 e.varargs.push_back(lookup_or_add(*OI));
261 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
263 e.varargs.push_back(*II);
268 //===----------------------------------------------------------------------===//
269 // ValueTable External Functions
270 //===----------------------------------------------------------------------===//
272 /// add - Insert a value into the table with a specified value number.
273 void ValueTable::add(Value *V, uint32_t num) {
274 valueNumbering.insert(std::make_pair(V, num));
277 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
278 if (AA->doesNotAccessMemory(C)) {
279 Expression exp = create_expression(C);
280 uint32_t &e = expressionNumbering[exp];
281 if (!e) e = nextValueNumber++;
282 valueNumbering[C] = e;
284 } else if (AA->onlyReadsMemory(C)) {
285 Expression exp = create_expression(C);
286 uint32_t &e = expressionNumbering[exp];
288 e = nextValueNumber++;
289 valueNumbering[C] = e;
293 e = nextValueNumber++;
294 valueNumbering[C] = e;
298 MemDepResult local_dep = MD->getDependency(C);
300 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
301 valueNumbering[C] = nextValueNumber;
302 return nextValueNumber++;
305 if (local_dep.isDef()) {
306 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
308 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
309 valueNumbering[C] = nextValueNumber;
310 return nextValueNumber++;
313 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
314 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
315 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
317 valueNumbering[C] = nextValueNumber;
318 return nextValueNumber++;
322 uint32_t v = lookup_or_add(local_cdep);
323 valueNumbering[C] = v;
328 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
329 MD->getNonLocalCallDependency(CallSite(C));
330 // FIXME: Move the checking logic to MemDep!
333 // Check to see if we have a single dominating call instruction that is
335 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
336 const NonLocalDepEntry *I = &deps[i];
337 if (I->getResult().isNonLocal())
340 // We don't handle non-definitions. If we already have a call, reject
341 // instruction dependencies.
342 if (!I->getResult().isDef() || cdep != 0) {
347 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
348 // FIXME: All duplicated with non-local case.
349 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
350 cdep = NonLocalDepCall;
359 valueNumbering[C] = nextValueNumber;
360 return nextValueNumber++;
363 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
364 valueNumbering[C] = nextValueNumber;
365 return nextValueNumber++;
367 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
368 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
369 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
371 valueNumbering[C] = nextValueNumber;
372 return nextValueNumber++;
376 uint32_t v = lookup_or_add(cdep);
377 valueNumbering[C] = v;
381 valueNumbering[C] = nextValueNumber;
382 return nextValueNumber++;
386 /// lookup_or_add - Returns the value number for the specified value, assigning
387 /// it a new number if it did not have one before.
388 uint32_t ValueTable::lookup_or_add(Value *V) {
389 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
390 if (VI != valueNumbering.end())
393 if (!isa<Instruction>(V)) {
394 valueNumbering[V] = nextValueNumber;
395 return nextValueNumber++;
398 Instruction* I = cast<Instruction>(V);
400 switch (I->getOpcode()) {
401 case Instruction::Call:
402 return lookup_or_add_call(cast<CallInst>(I));
403 case Instruction::Add:
404 case Instruction::FAdd:
405 case Instruction::Sub:
406 case Instruction::FSub:
407 case Instruction::Mul:
408 case Instruction::FMul:
409 case Instruction::UDiv:
410 case Instruction::SDiv:
411 case Instruction::FDiv:
412 case Instruction::URem:
413 case Instruction::SRem:
414 case Instruction::FRem:
415 case Instruction::Shl:
416 case Instruction::LShr:
417 case Instruction::AShr:
418 case Instruction::And:
419 case Instruction::Or:
420 case Instruction::Xor:
421 case Instruction::ICmp:
422 case Instruction::FCmp:
423 case Instruction::Trunc:
424 case Instruction::ZExt:
425 case Instruction::SExt:
426 case Instruction::FPToUI:
427 case Instruction::FPToSI:
428 case Instruction::UIToFP:
429 case Instruction::SIToFP:
430 case Instruction::FPTrunc:
431 case Instruction::FPExt:
432 case Instruction::PtrToInt:
433 case Instruction::IntToPtr:
434 case Instruction::BitCast:
435 case Instruction::Select:
436 case Instruction::ExtractElement:
437 case Instruction::InsertElement:
438 case Instruction::ShuffleVector:
439 case Instruction::InsertValue:
440 case Instruction::GetElementPtr:
441 exp = create_expression(I);
443 case Instruction::ExtractValue:
444 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
447 valueNumbering[V] = nextValueNumber;
448 return nextValueNumber++;
451 uint32_t& e = expressionNumbering[exp];
452 if (!e) e = nextValueNumber++;
453 valueNumbering[V] = e;
457 /// lookup - Returns the value number of the specified value. Fails if
458 /// the value has not yet been numbered.
459 uint32_t ValueTable::lookup(Value *V) const {
460 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
461 assert(VI != valueNumbering.end() && "Value not numbered?");
465 /// lookup_or_add_cmp - Returns the value number of the given comparison,
466 /// assigning it a new number if it did not have one before. Useful when
467 /// we deduced the result of a comparison, but don't immediately have an
468 /// instruction realizing that comparison to hand.
469 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
470 CmpInst::Predicate Predicate,
471 Value *LHS, Value *RHS) {
472 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
473 uint32_t& e = expressionNumbering[exp];
474 if (!e) e = nextValueNumber++;
478 /// clear - Remove all entries from the ValueTable.
479 void ValueTable::clear() {
480 valueNumbering.clear();
481 expressionNumbering.clear();
485 /// erase - Remove a value from the value numbering.
486 void ValueTable::erase(Value *V) {
487 valueNumbering.erase(V);
490 /// verifyRemoved - Verify that the value is removed from all internal data
492 void ValueTable::verifyRemoved(const Value *V) const {
493 for (DenseMap<Value*, uint32_t>::const_iterator
494 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
495 assert(I->first != V && "Inst still occurs in value numbering map!");
499 //===----------------------------------------------------------------------===//
501 //===----------------------------------------------------------------------===//
505 struct AvailableValueInBlock {
506 /// BB - The basic block in question.
509 SimpleVal, // A simple offsetted value that is accessed.
510 LoadVal, // A value produced by a load.
511 MemIntrin, // A memory intrinsic which is loaded from.
512 UndefVal // A UndefValue representing a value from dead block (which
513 // is not yet physically removed from the CFG).
516 /// V - The value that is live out of the block.
517 PointerIntPair<Value *, 2, ValType> Val;
519 /// Offset - The byte offset in Val that is interesting for the load query.
522 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
523 unsigned Offset = 0) {
524 AvailableValueInBlock Res;
526 Res.Val.setPointer(V);
527 Res.Val.setInt(SimpleVal);
532 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
533 unsigned Offset = 0) {
534 AvailableValueInBlock Res;
536 Res.Val.setPointer(MI);
537 Res.Val.setInt(MemIntrin);
542 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
543 unsigned Offset = 0) {
544 AvailableValueInBlock Res;
546 Res.Val.setPointer(LI);
547 Res.Val.setInt(LoadVal);
552 static AvailableValueInBlock getUndef(BasicBlock *BB) {
553 AvailableValueInBlock Res;
555 Res.Val.setPointer(0);
556 Res.Val.setInt(UndefVal);
561 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
562 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
563 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
564 bool isUndefValue() const { return Val.getInt() == UndefVal; }
566 Value *getSimpleValue() const {
567 assert(isSimpleValue() && "Wrong accessor");
568 return Val.getPointer();
571 LoadInst *getCoercedLoadValue() const {
572 assert(isCoercedLoadValue() && "Wrong accessor");
573 return cast<LoadInst>(Val.getPointer());
576 MemIntrinsic *getMemIntrinValue() const {
577 assert(isMemIntrinValue() && "Wrong accessor");
578 return cast<MemIntrinsic>(Val.getPointer());
581 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
582 /// defined here to the specified type. This handles various coercion cases.
583 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const;
586 class GVN : public FunctionPass {
588 MemoryDependenceAnalysis *MD;
590 const DataLayout *TD;
591 const TargetLibraryInfo *TLI;
592 SetVector<BasicBlock *> DeadBlocks;
596 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
597 /// have that value number. Use findLeader to query it.
598 struct LeaderTableEntry {
600 const BasicBlock *BB;
601 LeaderTableEntry *Next;
603 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
604 BumpPtrAllocator TableAllocator;
606 SmallVector<Instruction*, 8> InstrsToErase;
608 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
609 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
610 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
613 static char ID; // Pass identification, replacement for typeid
614 explicit GVN(bool noloads = false)
615 : FunctionPass(ID), NoLoads(noloads), MD(0) {
616 initializeGVNPass(*PassRegistry::getPassRegistry());
619 bool runOnFunction(Function &F);
621 /// markInstructionForDeletion - This removes the specified instruction from
622 /// our various maps and marks it for deletion.
623 void markInstructionForDeletion(Instruction *I) {
625 InstrsToErase.push_back(I);
628 const DataLayout *getDataLayout() const { return TD; }
629 DominatorTree &getDominatorTree() const { return *DT; }
630 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
631 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
633 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
634 /// its value number.
635 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
636 LeaderTableEntry &Curr = LeaderTable[N];
643 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
646 Node->Next = Curr.Next;
650 /// removeFromLeaderTable - Scan the list of values corresponding to a given
651 /// value number, and remove the given instruction if encountered.
652 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
653 LeaderTableEntry* Prev = 0;
654 LeaderTableEntry* Curr = &LeaderTable[N];
656 while (Curr->Val != I || Curr->BB != BB) {
662 Prev->Next = Curr->Next;
668 LeaderTableEntry* Next = Curr->Next;
669 Curr->Val = Next->Val;
671 Curr->Next = Next->Next;
676 // List of critical edges to be split between iterations.
677 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
679 // This transformation requires dominator postdominator info
680 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
681 AU.addRequired<DominatorTree>();
682 AU.addRequired<TargetLibraryInfo>();
684 AU.addRequired<MemoryDependenceAnalysis>();
685 AU.addRequired<AliasAnalysis>();
687 AU.addPreserved<DominatorTree>();
688 AU.addPreserved<AliasAnalysis>();
692 // Helper fuctions of redundant load elimination
693 bool processLoad(LoadInst *L);
694 bool processNonLocalLoad(LoadInst *L);
695 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
696 AvailValInBlkVect &ValuesPerBlock,
697 UnavailBlkVect &UnavailableBlocks);
698 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
699 UnavailBlkVect &UnavailableBlocks);
701 // Other helper routines
702 bool processInstruction(Instruction *I);
703 bool processBlock(BasicBlock *BB);
704 void dump(DenseMap<uint32_t, Value*> &d);
705 bool iterateOnFunction(Function &F);
706 bool performPRE(Function &F);
707 Value *findLeader(const BasicBlock *BB, uint32_t num);
708 void cleanupGlobalSets();
709 void verifyRemoved(const Instruction *I) const;
710 bool splitCriticalEdges();
711 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
712 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
713 const BasicBlockEdge &Root);
714 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
715 bool processFoldableCondBr(BranchInst *BI);
716 void addDeadBlock(BasicBlock *BB);
717 void assignValNumForDeadCode();
723 // createGVNPass - The public interface to this file...
724 FunctionPass *llvm::createGVNPass(bool NoLoads) {
725 return new GVN(NoLoads);
728 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
729 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
730 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
731 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
732 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
733 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
735 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
736 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
738 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
739 E = d.end(); I != E; ++I) {
740 errs() << I->first << "\n";
747 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
748 /// we're analyzing is fully available in the specified block. As we go, keep
749 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
750 /// map is actually a tri-state map with the following values:
751 /// 0) we know the block *is not* fully available.
752 /// 1) we know the block *is* fully available.
753 /// 2) we do not know whether the block is fully available or not, but we are
754 /// currently speculating that it will be.
755 /// 3) we are speculating for this block and have used that to speculate for
757 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
758 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
759 uint32_t RecurseDepth) {
760 if (RecurseDepth > MaxRecurseDepth)
763 // Optimistically assume that the block is fully available and check to see
764 // if we already know about this block in one lookup.
765 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
766 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
768 // If the entry already existed for this block, return the precomputed value.
770 // If this is a speculative "available" value, mark it as being used for
771 // speculation of other blocks.
772 if (IV.first->second == 2)
773 IV.first->second = 3;
774 return IV.first->second != 0;
777 // Otherwise, see if it is fully available in all predecessors.
778 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
780 // If this block has no predecessors, it isn't live-in here.
782 goto SpeculationFailure;
784 for (; PI != PE; ++PI)
785 // If the value isn't fully available in one of our predecessors, then it
786 // isn't fully available in this block either. Undo our previous
787 // optimistic assumption and bail out.
788 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
789 goto SpeculationFailure;
793 // SpeculationFailure - If we get here, we found out that this is not, after
794 // all, a fully-available block. We have a problem if we speculated on this and
795 // used the speculation to mark other blocks as available.
797 char &BBVal = FullyAvailableBlocks[BB];
799 // If we didn't speculate on this, just return with it set to false.
805 // If we did speculate on this value, we could have blocks set to 1 that are
806 // incorrect. Walk the (transitive) successors of this block and mark them as
808 SmallVector<BasicBlock*, 32> BBWorklist;
809 BBWorklist.push_back(BB);
812 BasicBlock *Entry = BBWorklist.pop_back_val();
813 // Note that this sets blocks to 0 (unavailable) if they happen to not
814 // already be in FullyAvailableBlocks. This is safe.
815 char &EntryVal = FullyAvailableBlocks[Entry];
816 if (EntryVal == 0) continue; // Already unavailable.
818 // Mark as unavailable.
821 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
822 BBWorklist.push_back(*I);
823 } while (!BBWorklist.empty());
829 /// CanCoerceMustAliasedValueToLoad - Return true if
830 /// CoerceAvailableValueToLoadType will succeed.
831 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
833 const DataLayout &TD) {
834 // If the loaded or stored value is an first class array or struct, don't try
835 // to transform them. We need to be able to bitcast to integer.
836 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
837 StoredVal->getType()->isStructTy() ||
838 StoredVal->getType()->isArrayTy())
841 // The store has to be at least as big as the load.
842 if (TD.getTypeSizeInBits(StoredVal->getType()) <
843 TD.getTypeSizeInBits(LoadTy))
849 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
850 /// then a load from a must-aliased pointer of a different type, try to coerce
851 /// the stored value. LoadedTy is the type of the load we want to replace and
852 /// InsertPt is the place to insert new instructions.
854 /// If we can't do it, return null.
855 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
857 Instruction *InsertPt,
858 const DataLayout &TD) {
859 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
862 // If this is already the right type, just return it.
863 Type *StoredValTy = StoredVal->getType();
865 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
866 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
868 // If the store and reload are the same size, we can always reuse it.
869 if (StoreSize == LoadSize) {
870 // Pointer to Pointer -> use bitcast.
871 if (StoredValTy->getScalarType()->isPointerTy() &&
872 LoadedTy->getScalarType()->isPointerTy())
873 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
875 // Convert source pointers to integers, which can be bitcast.
876 if (StoredValTy->getScalarType()->isPointerTy()) {
877 StoredValTy = TD.getIntPtrType(StoredValTy);
878 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
881 Type *TypeToCastTo = LoadedTy;
882 if (TypeToCastTo->getScalarType()->isPointerTy())
883 TypeToCastTo = TD.getIntPtrType(TypeToCastTo);
885 if (StoredValTy != TypeToCastTo)
886 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
888 // Cast to pointer if the load needs a pointer type.
889 if (LoadedTy->getScalarType()->isPointerTy())
890 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
895 // If the loaded value is smaller than the available value, then we can
896 // extract out a piece from it. If the available value is too small, then we
897 // can't do anything.
898 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
900 // Convert source pointers to integers, which can be manipulated.
901 if (StoredValTy->getScalarType()->isPointerTy()) {
902 StoredValTy = TD.getIntPtrType(StoredValTy);
903 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
906 // Convert vectors and fp to integer, which can be manipulated.
907 if (!StoredValTy->isIntegerTy()) {
908 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
909 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
912 // If this is a big-endian system, we need to shift the value down to the low
913 // bits so that a truncate will work.
914 if (TD.isBigEndian()) {
915 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
916 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
919 // Truncate the integer to the right size now.
920 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
921 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
923 if (LoadedTy == NewIntTy)
926 // If the result is a pointer, inttoptr.
927 if (LoadedTy->getScalarType()->isPointerTy())
928 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
930 // Otherwise, bitcast.
931 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
934 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
935 /// memdep query of a load that ends up being a clobbering memory write (store,
936 /// memset, memcpy, memmove). This means that the write *may* provide bits used
937 /// by the load but we can't be sure because the pointers don't mustalias.
939 /// Check this case to see if there is anything more we can do before we give
940 /// up. This returns -1 if we have to give up, or a byte number in the stored
941 /// value of the piece that feeds the load.
942 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
944 uint64_t WriteSizeInBits,
945 const DataLayout &TD) {
946 // If the loaded or stored value is a first class array or struct, don't try
947 // to transform them. We need to be able to bitcast to integer.
948 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
951 int64_t StoreOffset = 0, LoadOffset = 0;
952 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&TD);
953 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &TD);
954 if (StoreBase != LoadBase)
957 // If the load and store are to the exact same address, they should have been
958 // a must alias. AA must have gotten confused.
959 // FIXME: Study to see if/when this happens. One case is forwarding a memset
960 // to a load from the base of the memset.
962 if (LoadOffset == StoreOffset) {
963 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
964 << "Base = " << *StoreBase << "\n"
965 << "Store Ptr = " << *WritePtr << "\n"
966 << "Store Offs = " << StoreOffset << "\n"
967 << "Load Ptr = " << *LoadPtr << "\n";
972 // If the load and store don't overlap at all, the store doesn't provide
973 // anything to the load. In this case, they really don't alias at all, AA
974 // must have gotten confused.
975 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
977 if ((WriteSizeInBits & 7) | (LoadSize & 7))
979 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
983 bool isAAFailure = false;
984 if (StoreOffset < LoadOffset)
985 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
987 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
991 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
992 << "Base = " << *StoreBase << "\n"
993 << "Store Ptr = " << *WritePtr << "\n"
994 << "Store Offs = " << StoreOffset << "\n"
995 << "Load Ptr = " << *LoadPtr << "\n";
1001 // If the Load isn't completely contained within the stored bits, we don't
1002 // have all the bits to feed it. We could do something crazy in the future
1003 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1005 if (StoreOffset > LoadOffset ||
1006 StoreOffset+StoreSize < LoadOffset+LoadSize)
1009 // Okay, we can do this transformation. Return the number of bytes into the
1010 // store that the load is.
1011 return LoadOffset-StoreOffset;
1014 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1015 /// memdep query of a load that ends up being a clobbering store.
1016 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1018 const DataLayout &TD) {
1019 // Cannot handle reading from store of first-class aggregate yet.
1020 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1021 DepSI->getValueOperand()->getType()->isArrayTy())
1024 Value *StorePtr = DepSI->getPointerOperand();
1025 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1026 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1027 StorePtr, StoreSize, TD);
1030 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
1031 /// memdep query of a load that ends up being clobbered by another load. See if
1032 /// the other load can feed into the second load.
1033 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1034 LoadInst *DepLI, const DataLayout &TD){
1035 // Cannot handle reading from store of first-class aggregate yet.
1036 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1039 Value *DepPtr = DepLI->getPointerOperand();
1040 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType());
1041 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD);
1042 if (R != -1) return R;
1044 // If we have a load/load clobber an DepLI can be widened to cover this load,
1045 // then we should widen it!
1046 int64_t LoadOffs = 0;
1047 const Value *LoadBase =
1048 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &TD);
1049 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
1051 unsigned Size = MemoryDependenceAnalysis::
1052 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD);
1053 if (Size == 0) return -1;
1055 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD);
1060 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1062 const DataLayout &TD) {
1063 // If the mem operation is a non-constant size, we can't handle it.
1064 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1065 if (SizeCst == 0) return -1;
1066 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1068 // If this is memset, we just need to see if the offset is valid in the size
1070 if (MI->getIntrinsicID() == Intrinsic::memset)
1071 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1074 // If we have a memcpy/memmove, the only case we can handle is if this is a
1075 // copy from constant memory. In that case, we can read directly from the
1077 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1079 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1080 if (Src == 0) return -1;
1082 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
1083 if (GV == 0 || !GV->isConstant()) return -1;
1085 // See if the access is within the bounds of the transfer.
1086 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1087 MI->getDest(), MemSizeInBits, TD);
1091 // Otherwise, see if we can constant fold a load from the constant with the
1092 // offset applied as appropriate.
1093 Src = ConstantExpr::getBitCast(Src,
1094 llvm::Type::getInt8PtrTy(Src->getContext()));
1095 Constant *OffsetCst =
1096 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1097 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1098 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1099 if (ConstantFoldLoadFromConstPtr(Src, &TD))
1105 /// GetStoreValueForLoad - This function is called when we have a
1106 /// memdep query of a load that ends up being a clobbering store. This means
1107 /// that the store provides bits used by the load but we the pointers don't
1108 /// mustalias. Check this case to see if there is anything more we can do
1109 /// before we give up.
1110 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1112 Instruction *InsertPt, const DataLayout &TD){
1113 LLVMContext &Ctx = SrcVal->getType()->getContext();
1115 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1116 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
1118 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1120 // Compute which bits of the stored value are being used by the load. Convert
1121 // to an integer type to start with.
1122 if (SrcVal->getType()->getScalarType()->isPointerTy())
1123 SrcVal = Builder.CreatePtrToInt(SrcVal,
1124 TD.getIntPtrType(SrcVal->getType()));
1125 if (!SrcVal->getType()->isIntegerTy())
1126 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1128 // Shift the bits to the least significant depending on endianness.
1130 if (TD.isLittleEndian())
1131 ShiftAmt = Offset*8;
1133 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1136 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1138 if (LoadSize != StoreSize)
1139 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1141 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1144 /// GetLoadValueForLoad - This function is called when we have a
1145 /// memdep query of a load that ends up being a clobbering load. This means
1146 /// that the load *may* provide bits used by the load but we can't be sure
1147 /// because the pointers don't mustalias. Check this case to see if there is
1148 /// anything more we can do before we give up.
1149 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1150 Type *LoadTy, Instruction *InsertPt,
1152 const DataLayout &TD = *gvn.getDataLayout();
1153 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1154 // widen SrcVal out to a larger load.
1155 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType());
1156 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
1157 if (Offset+LoadSize > SrcValSize) {
1158 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1159 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1160 // If we have a load/load clobber an DepLI can be widened to cover this
1161 // load, then we should widen it to the next power of 2 size big enough!
1162 unsigned NewLoadSize = Offset+LoadSize;
1163 if (!isPowerOf2_32(NewLoadSize))
1164 NewLoadSize = NextPowerOf2(NewLoadSize);
1166 Value *PtrVal = SrcVal->getPointerOperand();
1168 // Insert the new load after the old load. This ensures that subsequent
1169 // memdep queries will find the new load. We can't easily remove the old
1170 // load completely because it is already in the value numbering table.
1171 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1173 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1174 DestPTy = PointerType::get(DestPTy,
1175 cast<PointerType>(PtrVal->getType())->getAddressSpace());
1176 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1177 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1178 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1179 NewLoad->takeName(SrcVal);
1180 NewLoad->setAlignment(SrcVal->getAlignment());
1182 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1183 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1185 // Replace uses of the original load with the wider load. On a big endian
1186 // system, we need to shift down to get the relevant bits.
1187 Value *RV = NewLoad;
1188 if (TD.isBigEndian())
1189 RV = Builder.CreateLShr(RV,
1190 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1191 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1192 SrcVal->replaceAllUsesWith(RV);
1194 // We would like to use gvn.markInstructionForDeletion here, but we can't
1195 // because the load is already memoized into the leader map table that GVN
1196 // tracks. It is potentially possible to remove the load from the table,
1197 // but then there all of the operations based on it would need to be
1198 // rehashed. Just leave the dead load around.
1199 gvn.getMemDep().removeInstruction(SrcVal);
1203 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD);
1207 /// GetMemInstValueForLoad - This function is called when we have a
1208 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1209 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1210 Type *LoadTy, Instruction *InsertPt,
1211 const DataLayout &TD){
1212 LLVMContext &Ctx = LoadTy->getContext();
1213 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1215 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1217 // We know that this method is only called when the mem transfer fully
1218 // provides the bits for the load.
1219 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1220 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1221 // independently of what the offset is.
1222 Value *Val = MSI->getValue();
1224 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1226 Value *OneElt = Val;
1228 // Splat the value out to the right number of bits.
1229 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1230 // If we can double the number of bytes set, do it.
1231 if (NumBytesSet*2 <= LoadSize) {
1232 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1233 Val = Builder.CreateOr(Val, ShVal);
1238 // Otherwise insert one byte at a time.
1239 Value *ShVal = Builder.CreateShl(Val, 1*8);
1240 Val = Builder.CreateOr(OneElt, ShVal);
1244 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1247 // Otherwise, this is a memcpy/memmove from a constant global.
1248 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1249 Constant *Src = cast<Constant>(MTI->getSource());
1251 // Otherwise, see if we can constant fold a load from the constant with the
1252 // offset applied as appropriate.
1253 Src = ConstantExpr::getBitCast(Src,
1254 llvm::Type::getInt8PtrTy(Src->getContext()));
1255 Constant *OffsetCst =
1256 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1257 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1258 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1259 return ConstantFoldLoadFromConstPtr(Src, &TD);
1263 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1264 /// construct SSA form, allowing us to eliminate LI. This returns the value
1265 /// that should be used at LI's definition site.
1266 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1267 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1269 // Check for the fully redundant, dominating load case. In this case, we can
1270 // just use the dominating value directly.
1271 if (ValuesPerBlock.size() == 1 &&
1272 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1274 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1275 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1278 // Otherwise, we have to construct SSA form.
1279 SmallVector<PHINode*, 8> NewPHIs;
1280 SSAUpdater SSAUpdate(&NewPHIs);
1281 SSAUpdate.Initialize(LI->getType(), LI->getName());
1283 Type *LoadTy = LI->getType();
1285 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1286 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1287 BasicBlock *BB = AV.BB;
1289 if (SSAUpdate.HasValueForBlock(BB))
1292 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1295 // Perform PHI construction.
1296 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1298 // If new PHI nodes were created, notify alias analysis.
1299 if (V->getType()->getScalarType()->isPointerTy()) {
1300 AliasAnalysis *AA = gvn.getAliasAnalysis();
1302 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1303 AA->copyValue(LI, NewPHIs[i]);
1305 // Now that we've copied information to the new PHIs, scan through
1306 // them again and inform alias analysis that we've added potentially
1307 // escaping uses to any values that are operands to these PHIs.
1308 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1309 PHINode *P = NewPHIs[i];
1310 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1311 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1312 AA->addEscapingUse(P->getOperandUse(jj));
1320 Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1322 if (isSimpleValue()) {
1323 Res = getSimpleValue();
1324 if (Res->getType() != LoadTy) {
1325 const DataLayout *TD = gvn.getDataLayout();
1326 assert(TD && "Need target data to handle type mismatch case");
1327 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1330 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1331 << *getSimpleValue() << '\n'
1332 << *Res << '\n' << "\n\n\n");
1334 } else if (isCoercedLoadValue()) {
1335 LoadInst *Load = getCoercedLoadValue();
1336 if (Load->getType() == LoadTy && Offset == 0) {
1339 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1342 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1343 << *getCoercedLoadValue() << '\n'
1344 << *Res << '\n' << "\n\n\n");
1346 } else if (isMemIntrinValue()) {
1347 const DataLayout *TD = gvn.getDataLayout();
1348 assert(TD && "Need target data to handle type mismatch case");
1349 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1350 LoadTy, BB->getTerminator(), *TD);
1351 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1352 << " " << *getMemIntrinValue() << '\n'
1353 << *Res << '\n' << "\n\n\n");
1355 assert(isUndefValue() && "Should be UndefVal");
1356 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1357 return UndefValue::get(LoadTy);
1362 static bool isLifetimeStart(const Instruction *Inst) {
1363 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1364 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1368 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1369 AvailValInBlkVect &ValuesPerBlock,
1370 UnavailBlkVect &UnavailableBlocks) {
1372 // Filter out useless results (non-locals, etc). Keep track of the blocks
1373 // where we have a value available in repl, also keep track of whether we see
1374 // dependencies that produce an unknown value for the load (such as a call
1375 // that could potentially clobber the load).
1376 unsigned NumDeps = Deps.size();
1377 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1378 BasicBlock *DepBB = Deps[i].getBB();
1379 MemDepResult DepInfo = Deps[i].getResult();
1381 if (DeadBlocks.count(DepBB)) {
1382 // Dead dependent mem-op disguise as a load evaluating the same value
1383 // as the load in question.
1384 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1388 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1389 UnavailableBlocks.push_back(DepBB);
1393 if (DepInfo.isClobber()) {
1394 // The address being loaded in this non-local block may not be the same as
1395 // the pointer operand of the load if PHI translation occurs. Make sure
1396 // to consider the right address.
1397 Value *Address = Deps[i].getAddress();
1399 // If the dependence is to a store that writes to a superset of the bits
1400 // read by the load, we can extract the bits we need for the load from the
1402 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1403 if (TD && Address) {
1404 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1407 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1408 DepSI->getValueOperand(),
1415 // Check to see if we have something like this:
1418 // if we have this, replace the later with an extraction from the former.
1419 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1420 // If this is a clobber and L is the first instruction in its block, then
1421 // we have the first instruction in the entry block.
1422 if (DepLI != LI && Address && TD) {
1423 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
1424 LI->getPointerOperand(),
1428 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1435 // If the clobbering value is a memset/memcpy/memmove, see if we can
1436 // forward a value on from it.
1437 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1438 if (TD && Address) {
1439 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1442 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1449 UnavailableBlocks.push_back(DepBB);
1453 // DepInfo.isDef() here
1455 Instruction *DepInst = DepInfo.getInst();
1457 // Loading the allocation -> undef.
1458 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1459 // Loading immediately after lifetime begin -> undef.
1460 isLifetimeStart(DepInst)) {
1461 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1462 UndefValue::get(LI->getType())));
1466 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1467 // Reject loads and stores that are to the same address but are of
1468 // different types if we have to.
1469 if (S->getValueOperand()->getType() != LI->getType()) {
1470 // If the stored value is larger or equal to the loaded value, we can
1472 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1473 LI->getType(), *TD)) {
1474 UnavailableBlocks.push_back(DepBB);
1479 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1480 S->getValueOperand()));
1484 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1485 // If the types mismatch and we can't handle it, reject reuse of the load.
1486 if (LD->getType() != LI->getType()) {
1487 // If the stored value is larger or equal to the loaded value, we can
1489 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1490 UnavailableBlocks.push_back(DepBB);
1494 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1498 UnavailableBlocks.push_back(DepBB);
1502 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1503 UnavailBlkVect &UnavailableBlocks) {
1504 // Okay, we have *some* definitions of the value. This means that the value
1505 // is available in some of our (transitive) predecessors. Lets think about
1506 // doing PRE of this load. This will involve inserting a new load into the
1507 // predecessor when it's not available. We could do this in general, but
1508 // prefer to not increase code size. As such, we only do this when we know
1509 // that we only have to insert *one* load (which means we're basically moving
1510 // the load, not inserting a new one).
1512 SmallPtrSet<BasicBlock *, 4> Blockers;
1513 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1514 Blockers.insert(UnavailableBlocks[i]);
1516 // Let's find the first basic block with more than one predecessor. Walk
1517 // backwards through predecessors if needed.
1518 BasicBlock *LoadBB = LI->getParent();
1519 BasicBlock *TmpBB = LoadBB;
1521 while (TmpBB->getSinglePredecessor()) {
1522 TmpBB = TmpBB->getSinglePredecessor();
1523 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1525 if (Blockers.count(TmpBB))
1528 // If any of these blocks has more than one successor (i.e. if the edge we
1529 // just traversed was critical), then there are other paths through this
1530 // block along which the load may not be anticipated. Hoisting the load
1531 // above this block would be adding the load to execution paths along
1532 // which it was not previously executed.
1533 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1540 // Check to see how many predecessors have the loaded value fully
1542 DenseMap<BasicBlock*, Value*> PredLoads;
1543 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1544 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1545 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1546 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1547 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1549 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1550 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1552 BasicBlock *Pred = *PI;
1553 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1556 PredLoads[Pred] = 0;
1558 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1559 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1560 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1561 << Pred->getName() << "': " << *LI << '\n');
1565 if (LoadBB->isLandingPad()) {
1567 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1568 << Pred->getName() << "': " << *LI << '\n');
1572 CriticalEdgePred.push_back(Pred);
1576 // Decide whether PRE is profitable for this load.
1577 unsigned NumUnavailablePreds = PredLoads.size();
1578 assert(NumUnavailablePreds != 0 &&
1579 "Fully available value should already be eliminated!");
1581 // If this load is unavailable in multiple predecessors, reject it.
1582 // FIXME: If we could restructure the CFG, we could make a common pred with
1583 // all the preds that don't have an available LI and insert a new load into
1585 if (NumUnavailablePreds != 1)
1588 // Split critical edges, and update the unavailable predecessors accordingly.
1589 for (SmallVectorImpl<BasicBlock *>::iterator I = CriticalEdgePred.begin(),
1590 E = CriticalEdgePred.end(); I != E; I++) {
1591 BasicBlock *OrigPred = *I;
1592 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1593 PredLoads.erase(OrigPred);
1594 PredLoads[NewPred] = 0;
1595 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1596 << LoadBB->getName() << '\n');
1599 // Check if the load can safely be moved to all the unavailable predecessors.
1600 bool CanDoPRE = true;
1601 SmallVector<Instruction*, 8> NewInsts;
1602 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1603 E = PredLoads.end(); I != E; ++I) {
1604 BasicBlock *UnavailablePred = I->first;
1606 // Do PHI translation to get its value in the predecessor if necessary. The
1607 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1609 // If all preds have a single successor, then we know it is safe to insert
1610 // the load on the pred (?!?), so we can insert code to materialize the
1611 // pointer if it is not available.
1612 PHITransAddr Address(LI->getPointerOperand(), TD);
1614 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1617 // If we couldn't find or insert a computation of this phi translated value,
1620 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1621 << *LI->getPointerOperand() << "\n");
1626 I->second = LoadPtr;
1630 while (!NewInsts.empty()) {
1631 Instruction *I = NewInsts.pop_back_val();
1632 if (MD) MD->removeInstruction(I);
1633 I->eraseFromParent();
1635 // HINT:Don't revert the edge-splitting as following transformation may
1636 // also need to split these critial edges.
1637 return !CriticalEdgePred.empty();
1640 // Okay, we can eliminate this load by inserting a reload in the predecessor
1641 // and using PHI construction to get the value in the other predecessors, do
1643 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1644 DEBUG(if (!NewInsts.empty())
1645 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1646 << *NewInsts.back() << '\n');
1648 // Assign value numbers to the new instructions.
1649 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1650 // FIXME: We really _ought_ to insert these value numbers into their
1651 // parent's availability map. However, in doing so, we risk getting into
1652 // ordering issues. If a block hasn't been processed yet, we would be
1653 // marking a value as AVAIL-IN, which isn't what we intend.
1654 VN.lookup_or_add(NewInsts[i]);
1657 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1658 E = PredLoads.end(); I != E; ++I) {
1659 BasicBlock *UnavailablePred = I->first;
1660 Value *LoadPtr = I->second;
1662 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1664 UnavailablePred->getTerminator());
1666 // Transfer the old load's TBAA tag to the new load.
1667 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1668 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1670 // Transfer DebugLoc.
1671 NewLoad->setDebugLoc(LI->getDebugLoc());
1673 // Add the newly created load.
1674 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1676 MD->invalidateCachedPointerInfo(LoadPtr);
1677 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1680 // Perform PHI construction.
1681 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1682 LI->replaceAllUsesWith(V);
1683 if (isa<PHINode>(V))
1685 if (V->getType()->getScalarType()->isPointerTy())
1686 MD->invalidateCachedPointerInfo(V);
1687 markInstructionForDeletion(LI);
1692 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1693 /// non-local by performing PHI construction.
1694 bool GVN::processNonLocalLoad(LoadInst *LI) {
1695 // Step 1: Find the non-local dependencies of the load.
1697 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1698 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1700 // If we had to process more than one hundred blocks to find the
1701 // dependencies, this load isn't worth worrying about. Optimizing
1702 // it will be too expensive.
1703 unsigned NumDeps = Deps.size();
1707 // If we had a phi translation failure, we'll have a single entry which is a
1708 // clobber in the current block. Reject this early.
1710 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1712 dbgs() << "GVN: non-local load ";
1713 WriteAsOperand(dbgs(), LI);
1714 dbgs() << " has unknown dependencies\n";
1719 // Step 2: Analyze the availability of the load
1720 AvailValInBlkVect ValuesPerBlock;
1721 UnavailBlkVect UnavailableBlocks;
1722 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1724 // If we have no predecessors that produce a known value for this load, exit
1726 if (ValuesPerBlock.empty())
1729 // Step 3: Eliminate fully redundancy.
1731 // If all of the instructions we depend on produce a known value for this
1732 // load, then it is fully redundant and we can use PHI insertion to compute
1733 // its value. Insert PHIs and remove the fully redundant value now.
1734 if (UnavailableBlocks.empty()) {
1735 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1737 // Perform PHI construction.
1738 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1739 LI->replaceAllUsesWith(V);
1741 if (isa<PHINode>(V))
1743 if (V->getType()->getScalarType()->isPointerTy())
1744 MD->invalidateCachedPointerInfo(V);
1745 markInstructionForDeletion(LI);
1750 // Step 4: Eliminate partial redundancy.
1751 if (!EnablePRE || !EnableLoadPRE)
1754 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1758 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1759 // Patch the replacement so that it is not more restrictive than the value
1761 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1762 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1763 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) &&
1764 isa<OverflowingBinaryOperator>(ReplOp)) {
1765 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap())
1766 ReplOp->setHasNoSignedWrap(false);
1767 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap())
1768 ReplOp->setHasNoUnsignedWrap(false);
1770 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1771 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
1772 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata);
1773 for (int i = 0, n = Metadata.size(); i < n; ++i) {
1774 unsigned Kind = Metadata[i].first;
1775 MDNode *IMD = I->getMetadata(Kind);
1776 MDNode *ReplMD = Metadata[i].second;
1779 ReplInst->setMetadata(Kind, NULL); // Remove unknown metadata
1781 case LLVMContext::MD_dbg:
1782 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1783 case LLVMContext::MD_tbaa:
1784 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD));
1786 case LLVMContext::MD_range:
1787 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD));
1789 case LLVMContext::MD_prof:
1790 llvm_unreachable("MD_prof in a non terminator instruction");
1792 case LLVMContext::MD_fpmath:
1793 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD));
1800 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1801 patchReplacementInstruction(I, Repl);
1802 I->replaceAllUsesWith(Repl);
1805 /// processLoad - Attempt to eliminate a load, first by eliminating it
1806 /// locally, and then attempting non-local elimination if that fails.
1807 bool GVN::processLoad(LoadInst *L) {
1814 if (L->use_empty()) {
1815 markInstructionForDeletion(L);
1819 // ... to a pointer that has been loaded from before...
1820 MemDepResult Dep = MD->getDependency(L);
1822 // If we have a clobber and target data is around, see if this is a clobber
1823 // that we can fix up through code synthesis.
1824 if (Dep.isClobber() && TD) {
1825 // Check to see if we have something like this:
1826 // store i32 123, i32* %P
1827 // %A = bitcast i32* %P to i8*
1828 // %B = gep i8* %A, i32 1
1831 // We could do that by recognizing if the clobber instructions are obviously
1832 // a common base + constant offset, and if the previous store (or memset)
1833 // completely covers this load. This sort of thing can happen in bitfield
1835 Value *AvailVal = 0;
1836 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1837 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1838 L->getPointerOperand(),
1841 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1842 L->getType(), L, *TD);
1845 // Check to see if we have something like this:
1848 // if we have this, replace the later with an extraction from the former.
1849 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1850 // If this is a clobber and L is the first instruction in its block, then
1851 // we have the first instruction in the entry block.
1855 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1856 L->getPointerOperand(),
1859 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1862 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1863 // a value on from it.
1864 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1865 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1866 L->getPointerOperand(),
1869 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
1873 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1874 << *AvailVal << '\n' << *L << "\n\n\n");
1876 // Replace the load!
1877 L->replaceAllUsesWith(AvailVal);
1878 if (AvailVal->getType()->getScalarType()->isPointerTy())
1879 MD->invalidateCachedPointerInfo(AvailVal);
1880 markInstructionForDeletion(L);
1886 // If the value isn't available, don't do anything!
1887 if (Dep.isClobber()) {
1889 // fast print dep, using operator<< on instruction is too slow.
1890 dbgs() << "GVN: load ";
1891 WriteAsOperand(dbgs(), L);
1892 Instruction *I = Dep.getInst();
1893 dbgs() << " is clobbered by " << *I << '\n';
1898 // If it is defined in another block, try harder.
1899 if (Dep.isNonLocal())
1900 return processNonLocalLoad(L);
1904 // fast print dep, using operator<< on instruction is too slow.
1905 dbgs() << "GVN: load ";
1906 WriteAsOperand(dbgs(), L);
1907 dbgs() << " has unknown dependence\n";
1912 Instruction *DepInst = Dep.getInst();
1913 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1914 Value *StoredVal = DepSI->getValueOperand();
1916 // The store and load are to a must-aliased pointer, but they may not
1917 // actually have the same type. See if we know how to reuse the stored
1918 // value (depending on its type).
1919 if (StoredVal->getType() != L->getType()) {
1921 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1926 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1927 << '\n' << *L << "\n\n\n");
1934 L->replaceAllUsesWith(StoredVal);
1935 if (StoredVal->getType()->getScalarType()->isPointerTy())
1936 MD->invalidateCachedPointerInfo(StoredVal);
1937 markInstructionForDeletion(L);
1942 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1943 Value *AvailableVal = DepLI;
1945 // The loads are of a must-aliased pointer, but they may not actually have
1946 // the same type. See if we know how to reuse the previously loaded value
1947 // (depending on its type).
1948 if (DepLI->getType() != L->getType()) {
1950 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1952 if (AvailableVal == 0)
1955 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1956 << "\n" << *L << "\n\n\n");
1963 patchAndReplaceAllUsesWith(L, AvailableVal);
1964 if (DepLI->getType()->getScalarType()->isPointerTy())
1965 MD->invalidateCachedPointerInfo(DepLI);
1966 markInstructionForDeletion(L);
1971 // If this load really doesn't depend on anything, then we must be loading an
1972 // undef value. This can happen when loading for a fresh allocation with no
1973 // intervening stores, for example.
1974 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1975 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1976 markInstructionForDeletion(L);
1981 // If this load occurs either right after a lifetime begin,
1982 // then the loaded value is undefined.
1983 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1984 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1985 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1986 markInstructionForDeletion(L);
1995 // findLeader - In order to find a leader for a given value number at a
1996 // specific basic block, we first obtain the list of all Values for that number,
1997 // and then scan the list to find one whose block dominates the block in
1998 // question. This is fast because dominator tree queries consist of only
1999 // a few comparisons of DFS numbers.
2000 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
2001 LeaderTableEntry Vals = LeaderTable[num];
2002 if (!Vals.Val) return 0;
2005 if (DT->dominates(Vals.BB, BB)) {
2007 if (isa<Constant>(Val)) return Val;
2010 LeaderTableEntry* Next = Vals.Next;
2012 if (DT->dominates(Next->BB, BB)) {
2013 if (isa<Constant>(Next->Val)) return Next->Val;
2014 if (!Val) Val = Next->Val;
2023 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
2024 /// use is dominated by the given basic block. Returns the number of uses that
2026 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
2027 const BasicBlockEdge &Root) {
2029 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2031 Use &U = (UI++).getUse();
2033 if (DT->dominates(Root, U)) {
2041 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
2042 /// true if every path from the entry block to 'Dst' passes via this edge. In
2043 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2044 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2045 DominatorTree *DT) {
2046 // While in theory it is interesting to consider the case in which Dst has
2047 // more than one predecessor, because Dst might be part of a loop which is
2048 // only reachable from Src, in practice it is pointless since at the time
2049 // GVN runs all such loops have preheaders, which means that Dst will have
2050 // been changed to have only one predecessor, namely Src.
2051 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2052 const BasicBlock *Src = E.getStart();
2053 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2058 /// propagateEquality - The given values are known to be equal in every block
2059 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2060 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2061 bool GVN::propagateEquality(Value *LHS, Value *RHS,
2062 const BasicBlockEdge &Root) {
2063 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2064 Worklist.push_back(std::make_pair(LHS, RHS));
2065 bool Changed = false;
2066 // For speed, compute a conservative fast approximation to
2067 // DT->dominates(Root, Root.getEnd());
2068 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2070 while (!Worklist.empty()) {
2071 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2072 LHS = Item.first; RHS = Item.second;
2074 if (LHS == RHS) continue;
2075 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2077 // Don't try to propagate equalities between constants.
2078 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2080 // Prefer a constant on the right-hand side, or an Argument if no constants.
2081 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2082 std::swap(LHS, RHS);
2083 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2085 // If there is no obvious reason to prefer the left-hand side over the right-
2086 // hand side, ensure the longest lived term is on the right-hand side, so the
2087 // shortest lived term will be replaced by the longest lived. This tends to
2088 // expose more simplifications.
2089 uint32_t LVN = VN.lookup_or_add(LHS);
2090 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2091 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2092 // Move the 'oldest' value to the right-hand side, using the value number as
2094 uint32_t RVN = VN.lookup_or_add(RHS);
2096 std::swap(LHS, RHS);
2101 // If value numbering later sees that an instruction in the scope is equal
2102 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2103 // the invariant that instructions only occur in the leader table for their
2104 // own value number (this is used by removeFromLeaderTable), do not do this
2105 // if RHS is an instruction (if an instruction in the scope is morphed into
2106 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2107 // using the leader table is about compiling faster, not optimizing better).
2108 // The leader table only tracks basic blocks, not edges. Only add to if we
2109 // have the simple case where the edge dominates the end.
2110 if (RootDominatesEnd && !isa<Instruction>(RHS))
2111 addToLeaderTable(LVN, RHS, Root.getEnd());
2113 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2114 // LHS always has at least one use that is not dominated by Root, this will
2115 // never do anything if LHS has only one use.
2116 if (!LHS->hasOneUse()) {
2117 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2118 Changed |= NumReplacements > 0;
2119 NumGVNEqProp += NumReplacements;
2122 // Now try to deduce additional equalities from this one. For example, if the
2123 // known equality was "(A != B)" == "false" then it follows that A and B are
2124 // equal in the scope. Only boolean equalities with an explicit true or false
2125 // RHS are currently supported.
2126 if (!RHS->getType()->isIntegerTy(1))
2127 // Not a boolean equality - bail out.
2129 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2131 // RHS neither 'true' nor 'false' - bail out.
2133 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2134 bool isKnownTrue = CI->isAllOnesValue();
2135 bool isKnownFalse = !isKnownTrue;
2137 // If "A && B" is known true then both A and B are known true. If "A || B"
2138 // is known false then both A and B are known false.
2140 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2141 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2142 Worklist.push_back(std::make_pair(A, RHS));
2143 Worklist.push_back(std::make_pair(B, RHS));
2147 // If we are propagating an equality like "(A == B)" == "true" then also
2148 // propagate the equality A == B. When propagating a comparison such as
2149 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2150 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
2151 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2153 // If "A == B" is known true, or "A != B" is known false, then replace
2154 // A with B everywhere in the scope.
2155 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2156 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2157 Worklist.push_back(std::make_pair(Op0, Op1));
2159 // If "A >= B" is known true, replace "A < B" with false everywhere.
2160 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2161 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2162 // Since we don't have the instruction "A < B" immediately to hand, work out
2163 // the value number that it would have and use that to find an appropriate
2164 // instruction (if any).
2165 uint32_t NextNum = VN.getNextUnusedValueNumber();
2166 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2167 // If the number we were assigned was brand new then there is no point in
2168 // looking for an instruction realizing it: there cannot be one!
2169 if (Num < NextNum) {
2170 Value *NotCmp = findLeader(Root.getEnd(), Num);
2171 if (NotCmp && isa<Instruction>(NotCmp)) {
2172 unsigned NumReplacements =
2173 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2174 Changed |= NumReplacements > 0;
2175 NumGVNEqProp += NumReplacements;
2178 // Ensure that any instruction in scope that gets the "A < B" value number
2179 // is replaced with false.
2180 // The leader table only tracks basic blocks, not edges. Only add to if we
2181 // have the simple case where the edge dominates the end.
2182 if (RootDominatesEnd)
2183 addToLeaderTable(Num, NotVal, Root.getEnd());
2192 /// processInstruction - When calculating availability, handle an instruction
2193 /// by inserting it into the appropriate sets
2194 bool GVN::processInstruction(Instruction *I) {
2195 // Ignore dbg info intrinsics.
2196 if (isa<DbgInfoIntrinsic>(I))
2199 // If the instruction can be easily simplified then do so now in preference
2200 // to value numbering it. Value numbering often exposes redundancies, for
2201 // example if it determines that %y is equal to %x then the instruction
2202 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2203 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) {
2204 I->replaceAllUsesWith(V);
2205 if (MD && V->getType()->getScalarType()->isPointerTy())
2206 MD->invalidateCachedPointerInfo(V);
2207 markInstructionForDeletion(I);
2212 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2213 if (processLoad(LI))
2216 unsigned Num = VN.lookup_or_add(LI);
2217 addToLeaderTable(Num, LI, LI->getParent());
2221 // For conditional branches, we can perform simple conditional propagation on
2222 // the condition value itself.
2223 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2224 if (!BI->isConditional())
2227 if (isa<Constant>(BI->getCondition()))
2228 return processFoldableCondBr(BI);
2230 Value *BranchCond = BI->getCondition();
2231 BasicBlock *TrueSucc = BI->getSuccessor(0);
2232 BasicBlock *FalseSucc = BI->getSuccessor(1);
2233 // Avoid multiple edges early.
2234 if (TrueSucc == FalseSucc)
2237 BasicBlock *Parent = BI->getParent();
2238 bool Changed = false;
2240 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2241 BasicBlockEdge TrueE(Parent, TrueSucc);
2242 Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
2244 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2245 BasicBlockEdge FalseE(Parent, FalseSucc);
2246 Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
2251 // For switches, propagate the case values into the case destinations.
2252 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2253 Value *SwitchCond = SI->getCondition();
2254 BasicBlock *Parent = SI->getParent();
2255 bool Changed = false;
2257 // Remember how many outgoing edges there are to every successor.
2258 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2259 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2260 ++SwitchEdges[SI->getSuccessor(i)];
2262 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2264 BasicBlock *Dst = i.getCaseSuccessor();
2265 // If there is only a single edge, propagate the case value into it.
2266 if (SwitchEdges.lookup(Dst) == 1) {
2267 BasicBlockEdge E(Parent, Dst);
2268 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
2274 // Instructions with void type don't return a value, so there's
2275 // no point in trying to find redundancies in them.
2276 if (I->getType()->isVoidTy()) return false;
2278 uint32_t NextNum = VN.getNextUnusedValueNumber();
2279 unsigned Num = VN.lookup_or_add(I);
2281 // Allocations are always uniquely numbered, so we can save time and memory
2282 // by fast failing them.
2283 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2284 addToLeaderTable(Num, I, I->getParent());
2288 // If the number we were assigned was a brand new VN, then we don't
2289 // need to do a lookup to see if the number already exists
2290 // somewhere in the domtree: it can't!
2291 if (Num >= NextNum) {
2292 addToLeaderTable(Num, I, I->getParent());
2296 // Perform fast-path value-number based elimination of values inherited from
2298 Value *repl = findLeader(I->getParent(), Num);
2300 // Failure, just remember this instance for future use.
2301 addToLeaderTable(Num, I, I->getParent());
2306 patchAndReplaceAllUsesWith(I, repl);
2307 if (MD && repl->getType()->getScalarType()->isPointerTy())
2308 MD->invalidateCachedPointerInfo(repl);
2309 markInstructionForDeletion(I);
2313 /// runOnFunction - This is the main transformation entry point for a function.
2314 bool GVN::runOnFunction(Function& F) {
2316 MD = &getAnalysis<MemoryDependenceAnalysis>();
2317 DT = &getAnalysis<DominatorTree>();
2318 TD = getAnalysisIfAvailable<DataLayout>();
2319 TLI = &getAnalysis<TargetLibraryInfo>();
2320 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2324 bool Changed = false;
2325 bool ShouldContinue = true;
2327 // Merge unconditional branches, allowing PRE to catch more
2328 // optimization opportunities.
2329 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2330 BasicBlock *BB = FI++;
2332 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2333 if (removedBlock) ++NumGVNBlocks;
2335 Changed |= removedBlock;
2338 unsigned Iteration = 0;
2339 while (ShouldContinue) {
2340 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2341 ShouldContinue = iterateOnFunction(F);
2342 Changed |= ShouldContinue;
2347 // Fabricate val-num for dead-code in order to suppress assertion in
2349 assignValNumForDeadCode();
2350 bool PREChanged = true;
2351 while (PREChanged) {
2352 PREChanged = performPRE(F);
2353 Changed |= PREChanged;
2357 // FIXME: Should perform GVN again after PRE does something. PRE can move
2358 // computations into blocks where they become fully redundant. Note that
2359 // we can't do this until PRE's critical edge splitting updates memdep.
2360 // Actually, when this happens, we should just fully integrate PRE into GVN.
2362 cleanupGlobalSets();
2363 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2371 bool GVN::processBlock(BasicBlock *BB) {
2372 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2373 // (and incrementing BI before processing an instruction).
2374 assert(InstrsToErase.empty() &&
2375 "We expect InstrsToErase to be empty across iterations");
2376 if (DeadBlocks.count(BB))
2379 bool ChangedFunction = false;
2381 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2383 ChangedFunction |= processInstruction(BI);
2384 if (InstrsToErase.empty()) {
2389 // If we need some instructions deleted, do it now.
2390 NumGVNInstr += InstrsToErase.size();
2392 // Avoid iterator invalidation.
2393 bool AtStart = BI == BB->begin();
2397 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2398 E = InstrsToErase.end(); I != E; ++I) {
2399 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2400 if (MD) MD->removeInstruction(*I);
2401 DEBUG(verifyRemoved(*I));
2402 (*I)->eraseFromParent();
2404 InstrsToErase.clear();
2412 return ChangedFunction;
2415 /// performPRE - Perform a purely local form of PRE that looks for diamond
2416 /// control flow patterns and attempts to perform simple PRE at the join point.
2417 bool GVN::performPRE(Function &F) {
2418 bool Changed = false;
2419 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2420 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2421 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2422 BasicBlock *CurrentBlock = *DI;
2424 // Nothing to PRE in the entry block.
2425 if (CurrentBlock == &F.getEntryBlock()) continue;
2427 // Don't perform PRE on a landing pad.
2428 if (CurrentBlock->isLandingPad()) continue;
2430 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2431 BE = CurrentBlock->end(); BI != BE; ) {
2432 Instruction *CurInst = BI++;
2434 if (isa<AllocaInst>(CurInst) ||
2435 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2436 CurInst->getType()->isVoidTy() ||
2437 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2438 isa<DbgInfoIntrinsic>(CurInst))
2441 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2442 // sinking the compare again, and it would force the code generator to
2443 // move the i1 from processor flags or predicate registers into a general
2444 // purpose register.
2445 if (isa<CmpInst>(CurInst))
2448 // We don't currently value number ANY inline asm calls.
2449 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2450 if (CallI->isInlineAsm())
2453 uint32_t ValNo = VN.lookup(CurInst);
2455 // Look for the predecessors for PRE opportunities. We're
2456 // only trying to solve the basic diamond case, where
2457 // a value is computed in the successor and one predecessor,
2458 // but not the other. We also explicitly disallow cases
2459 // where the successor is its own predecessor, because they're
2460 // more complicated to get right.
2461 unsigned NumWith = 0;
2462 unsigned NumWithout = 0;
2463 BasicBlock *PREPred = 0;
2466 for (pred_iterator PI = pred_begin(CurrentBlock),
2467 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2468 BasicBlock *P = *PI;
2469 // We're not interested in PRE where the block is its
2470 // own predecessor, or in blocks with predecessors
2471 // that are not reachable.
2472 if (P == CurrentBlock) {
2475 } else if (!DT->isReachableFromEntry(P)) {
2480 Value* predV = findLeader(P, ValNo);
2482 predMap.push_back(std::make_pair(static_cast<Value *>(0), P));
2485 } else if (predV == CurInst) {
2486 /* CurInst dominates this predecessor. */
2490 predMap.push_back(std::make_pair(predV, P));
2495 // Don't do PRE when it might increase code size, i.e. when
2496 // we would need to insert instructions in more than one pred.
2497 if (NumWithout != 1 || NumWith == 0)
2500 // Don't do PRE across indirect branch.
2501 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2504 // We can't do PRE safely on a critical edge, so instead we schedule
2505 // the edge to be split and perform the PRE the next time we iterate
2507 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2508 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2509 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2513 // Instantiate the expression in the predecessor that lacked it.
2514 // Because we are going top-down through the block, all value numbers
2515 // will be available in the predecessor by the time we need them. Any
2516 // that weren't originally present will have been instantiated earlier
2518 Instruction *PREInstr = CurInst->clone();
2519 bool success = true;
2520 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2521 Value *Op = PREInstr->getOperand(i);
2522 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2525 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2526 PREInstr->setOperand(i, V);
2533 // Fail out if we encounter an operand that is not available in
2534 // the PRE predecessor. This is typically because of loads which
2535 // are not value numbered precisely.
2537 DEBUG(verifyRemoved(PREInstr));
2542 PREInstr->insertBefore(PREPred->getTerminator());
2543 PREInstr->setName(CurInst->getName() + ".pre");
2544 PREInstr->setDebugLoc(CurInst->getDebugLoc());
2545 VN.add(PREInstr, ValNo);
2548 // Update the availability map to include the new instruction.
2549 addToLeaderTable(ValNo, PREInstr, PREPred);
2551 // Create a PHI to make the value available in this block.
2552 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(),
2553 CurInst->getName() + ".pre-phi",
2554 CurrentBlock->begin());
2555 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2556 if (Value *V = predMap[i].first)
2557 Phi->addIncoming(V, predMap[i].second);
2559 Phi->addIncoming(PREInstr, PREPred);
2563 addToLeaderTable(ValNo, Phi, CurrentBlock);
2564 Phi->setDebugLoc(CurInst->getDebugLoc());
2565 CurInst->replaceAllUsesWith(Phi);
2566 if (Phi->getType()->getScalarType()->isPointerTy()) {
2567 // Because we have added a PHI-use of the pointer value, it has now
2568 // "escaped" from alias analysis' perspective. We need to inform
2570 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
2572 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2573 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2577 MD->invalidateCachedPointerInfo(Phi);
2580 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2582 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2583 if (MD) MD->removeInstruction(CurInst);
2584 DEBUG(verifyRemoved(CurInst));
2585 CurInst->eraseFromParent();
2590 if (splitCriticalEdges())
2596 /// Split the critical edge connecting the given two blocks, and return
2597 /// the block inserted to the critical edge.
2598 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2599 BasicBlock *BB = SplitCriticalEdge(Pred, Succ, this);
2601 MD->invalidateCachedPredecessors();
2605 /// splitCriticalEdges - Split critical edges found during the previous
2606 /// iteration that may enable further optimization.
2607 bool GVN::splitCriticalEdges() {
2608 if (toSplit.empty())
2611 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2612 SplitCriticalEdge(Edge.first, Edge.second, this);
2613 } while (!toSplit.empty());
2614 if (MD) MD->invalidateCachedPredecessors();
2618 /// iterateOnFunction - Executes one iteration of GVN
2619 bool GVN::iterateOnFunction(Function &F) {
2620 cleanupGlobalSets();
2622 // Top-down walk of the dominator tree
2623 bool Changed = false;
2625 // Needed for value numbering with phi construction to work.
2626 ReversePostOrderTraversal<Function*> RPOT(&F);
2627 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2628 RE = RPOT.end(); RI != RE; ++RI)
2629 Changed |= processBlock(*RI);
2631 // Save the blocks this function have before transformation begins. GVN may
2632 // split critical edge, and hence may invalidate the RPO/DT iterator.
2634 std::vector<BasicBlock *> BBVect;
2635 BBVect.reserve(256);
2636 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2637 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2638 BBVect.push_back(DI->getBlock());
2640 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2642 Changed |= processBlock(*I);
2648 void GVN::cleanupGlobalSets() {
2650 LeaderTable.clear();
2651 TableAllocator.Reset();
2654 /// verifyRemoved - Verify that the specified instruction does not occur in our
2655 /// internal data structures.
2656 void GVN::verifyRemoved(const Instruction *Inst) const {
2657 VN.verifyRemoved(Inst);
2659 // Walk through the value number scope to make sure the instruction isn't
2660 // ferreted away in it.
2661 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2662 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2663 const LeaderTableEntry *Node = &I->second;
2664 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2666 while (Node->Next) {
2668 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2673 // BB is declared dead, which implied other blocks become dead as well. This
2674 // function is to add all these blocks to "DeadBlocks". For the dead blocks'
2675 // live successors, update their phi nodes by replacing the operands
2676 // corresponding to dead blocks with UndefVal.
2678 void GVN::addDeadBlock(BasicBlock *BB) {
2679 SmallVector<BasicBlock *, 4> NewDead;
2680 SmallSetVector<BasicBlock *, 4> DF;
2682 NewDead.push_back(BB);
2683 while (!NewDead.empty()) {
2684 BasicBlock *D = NewDead.pop_back_val();
2685 if (DeadBlocks.count(D))
2688 // All blocks dominated by D are dead.
2689 SmallVector<BasicBlock *, 8> Dom;
2690 DT->getDescendants(D, Dom);
2691 DeadBlocks.insert(Dom.begin(), Dom.end());
2693 // Figure out the dominance-frontier(D).
2694 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
2695 E = Dom.end(); I != E; I++) {
2697 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
2698 BasicBlock *S = *SI;
2699 if (DeadBlocks.count(S))
2702 bool AllPredDead = true;
2703 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
2704 if (!DeadBlocks.count(*PI)) {
2705 AllPredDead = false;
2710 // S could be proved dead later on. That is why we don't update phi
2711 // operands at this moment.
2714 // While S is not dominated by D, it is dead by now. This could take
2715 // place if S already have a dead predecessor before D is declared
2717 NewDead.push_back(S);
2723 // For the dead blocks' live successors, update their phi nodes by replacing
2724 // the operands corresponding to dead blocks with UndefVal.
2725 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2728 if (DeadBlocks.count(B))
2731 for (pred_iterator PI = pred_begin(B), PE = pred_end(B); PI != PE; PI++) {
2732 BasicBlock *P = *PI;
2733 if (!DeadBlocks.count(P))
2735 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2736 PHINode &Phi = cast<PHINode>(*II);
2737 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2738 UndefValue::get(Phi.getType()));
2744 // If the given branch is recognized as a foldable branch (i.e. conditional
2745 // branch with constant condition), it will perform following analyses and
2747 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2748 // R be the target of the dead out-coming edge.
2749 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2750 // edge. The result of this step will be {X| X is dominated by R}
2751 // 2) Identify those blocks which haves at least one dead prodecessor. The
2752 // result of this step will be dominance-frontier(R).
2753 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2754 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2756 // Return true iff *NEW* dead code are found.
2757 bool GVN::processFoldableCondBr(BranchInst *BI) {
2758 if (!BI || BI->isUnconditional())
2761 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2765 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2766 BI->getSuccessor(1) : BI->getSuccessor(0);
2767 if (DeadBlocks.count(DeadRoot))
2770 if (!DeadRoot->getSinglePredecessor())
2771 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2773 addDeadBlock(DeadRoot);
2777 // performPRE() will trigger assert if it come across an instruciton without
2778 // associated val-num. As it normally has far more live instructions than dead
2779 // instructions, it makes more sense just to "fabricate" a val-number for the
2780 // dead code than checking if instruction involved is dead or not.
2781 void GVN::assignValNumForDeadCode() {
2782 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
2783 E = DeadBlocks.end(); I != E; I++) {
2784 BasicBlock *BB = *I;
2785 for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
2787 Instruction *Inst = &*II;
2788 unsigned ValNum = VN.lookup_or_add(Inst);
2789 addToLeaderTable(ValNum, Inst, BB);