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/SetVector.h"
24 #include "llvm/ADT/SmallPtrSet.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/InstructionSimplify.h"
30 #include "llvm/Analysis/Loads.h"
31 #include "llvm/Analysis/MemoryBuiltins.h"
32 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
33 #include "llvm/Analysis/PHITransAddr.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/Dominators.h"
37 #include "llvm/IR/GlobalVariable.h"
38 #include "llvm/IR/IRBuilder.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/LLVMContext.h"
41 #include "llvm/IR/Metadata.h"
42 #include "llvm/Support/Allocator.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/PatternMatch.h"
46 #include "llvm/Target/TargetLibraryInfo.h"
47 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
48 #include "llvm/Transforms/Utils/SSAUpdater.h"
51 using namespace PatternMatch;
53 STATISTIC(NumGVNInstr, "Number of instructions deleted");
54 STATISTIC(NumGVNLoad, "Number of loads deleted");
55 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
56 STATISTIC(NumGVNBlocks, "Number of blocks merged");
57 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
58 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
59 STATISTIC(NumPRELoad, "Number of loads PRE'd");
61 static cl::opt<bool> EnablePRE("enable-pre",
62 cl::init(true), cl::Hidden);
63 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
65 // Maximum allowed recursion depth.
66 static cl::opt<uint32_t>
67 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
68 cl::desc("Max recurse depth (default = 1000)"));
70 //===----------------------------------------------------------------------===//
72 //===----------------------------------------------------------------------===//
74 /// This class holds the mapping between values and value numbers. It is used
75 /// as an efficient mechanism to determine the expression-wise equivalence of
81 SmallVector<uint32_t, 4> varargs;
83 Expression(uint32_t o = ~2U) : opcode(o) { }
85 bool operator==(const Expression &other) const {
86 if (opcode != other.opcode)
88 if (opcode == ~0U || opcode == ~1U)
90 if (type != other.type)
92 if (varargs != other.varargs)
97 friend hash_code hash_value(const Expression &Value) {
98 return hash_combine(Value.opcode, Value.type,
99 hash_combine_range(Value.varargs.begin(),
100 Value.varargs.end()));
105 DenseMap<Value*, uint32_t> valueNumbering;
106 DenseMap<Expression, uint32_t> expressionNumbering;
108 MemoryDependenceAnalysis *MD;
111 uint32_t nextValueNumber;
113 Expression create_expression(Instruction* I);
114 Expression create_cmp_expression(unsigned Opcode,
115 CmpInst::Predicate Predicate,
116 Value *LHS, Value *RHS);
117 Expression create_extractvalue_expression(ExtractValueInst* EI);
118 uint32_t lookup_or_add_call(CallInst* C);
120 ValueTable() : nextValueNumber(1) { }
121 uint32_t lookup_or_add(Value *V);
122 uint32_t lookup(Value *V) const;
123 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
124 Value *LHS, Value *RHS);
125 void add(Value *V, uint32_t num);
127 void erase(Value *v);
128 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
129 AliasAnalysis *getAliasAnalysis() const { return AA; }
130 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
131 void setDomTree(DominatorTree* D) { DT = D; }
132 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
133 void verifyRemoved(const Value *) const;
138 template <> struct DenseMapInfo<Expression> {
139 static inline Expression getEmptyKey() {
143 static inline Expression getTombstoneKey() {
147 static unsigned getHashValue(const Expression e) {
148 using llvm::hash_value;
149 return static_cast<unsigned>(hash_value(e));
151 static bool isEqual(const Expression &LHS, const Expression &RHS) {
158 //===----------------------------------------------------------------------===//
159 // ValueTable Internal Functions
160 //===----------------------------------------------------------------------===//
162 Expression ValueTable::create_expression(Instruction *I) {
164 e.type = I->getType();
165 e.opcode = I->getOpcode();
166 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
168 e.varargs.push_back(lookup_or_add(*OI));
169 if (I->isCommutative()) {
170 // Ensure that commutative instructions that only differ by a permutation
171 // of their operands get the same value number by sorting the operand value
172 // numbers. Since all commutative instructions have two operands it is more
173 // efficient to sort by hand rather than using, say, std::sort.
174 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
175 if (e.varargs[0] > e.varargs[1])
176 std::swap(e.varargs[0], e.varargs[1]);
179 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
180 // Sort the operand value numbers so x<y and y>x get the same value number.
181 CmpInst::Predicate Predicate = C->getPredicate();
182 if (e.varargs[0] > e.varargs[1]) {
183 std::swap(e.varargs[0], e.varargs[1]);
184 Predicate = CmpInst::getSwappedPredicate(Predicate);
186 e.opcode = (C->getOpcode() << 8) | Predicate;
187 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
188 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
190 e.varargs.push_back(*II);
196 Expression ValueTable::create_cmp_expression(unsigned Opcode,
197 CmpInst::Predicate Predicate,
198 Value *LHS, Value *RHS) {
199 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
200 "Not a comparison!");
202 e.type = CmpInst::makeCmpResultType(LHS->getType());
203 e.varargs.push_back(lookup_or_add(LHS));
204 e.varargs.push_back(lookup_or_add(RHS));
206 // Sort the operand value numbers so x<y and y>x get the same value number.
207 if (e.varargs[0] > e.varargs[1]) {
208 std::swap(e.varargs[0], e.varargs[1]);
209 Predicate = CmpInst::getSwappedPredicate(Predicate);
211 e.opcode = (Opcode << 8) | Predicate;
215 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
216 assert(EI != 0 && "Not an ExtractValueInst?");
218 e.type = EI->getType();
221 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
222 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
223 // EI might be an extract from one of our recognised intrinsics. If it
224 // is we'll synthesize a semantically equivalent expression instead on
225 // an extract value expression.
226 switch (I->getIntrinsicID()) {
227 case Intrinsic::sadd_with_overflow:
228 case Intrinsic::uadd_with_overflow:
229 e.opcode = Instruction::Add;
231 case Intrinsic::ssub_with_overflow:
232 case Intrinsic::usub_with_overflow:
233 e.opcode = Instruction::Sub;
235 case Intrinsic::smul_with_overflow:
236 case Intrinsic::umul_with_overflow:
237 e.opcode = Instruction::Mul;
244 // Intrinsic recognized. Grab its args to finish building the expression.
245 assert(I->getNumArgOperands() == 2 &&
246 "Expect two args for recognised intrinsics.");
247 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
248 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
253 // Not a recognised intrinsic. Fall back to producing an extract value
255 e.opcode = EI->getOpcode();
256 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
258 e.varargs.push_back(lookup_or_add(*OI));
260 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
262 e.varargs.push_back(*II);
267 //===----------------------------------------------------------------------===//
268 // ValueTable External Functions
269 //===----------------------------------------------------------------------===//
271 /// add - Insert a value into the table with a specified value number.
272 void ValueTable::add(Value *V, uint32_t num) {
273 valueNumbering.insert(std::make_pair(V, num));
276 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
277 if (AA->doesNotAccessMemory(C)) {
278 Expression exp = create_expression(C);
279 uint32_t &e = expressionNumbering[exp];
280 if (!e) e = nextValueNumber++;
281 valueNumbering[C] = e;
283 } else if (AA->onlyReadsMemory(C)) {
284 Expression exp = create_expression(C);
285 uint32_t &e = expressionNumbering[exp];
287 e = nextValueNumber++;
288 valueNumbering[C] = e;
292 e = nextValueNumber++;
293 valueNumbering[C] = e;
297 MemDepResult local_dep = MD->getDependency(C);
299 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
300 valueNumbering[C] = nextValueNumber;
301 return nextValueNumber++;
304 if (local_dep.isDef()) {
305 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
307 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
308 valueNumbering[C] = nextValueNumber;
309 return nextValueNumber++;
312 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
313 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
314 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
316 valueNumbering[C] = nextValueNumber;
317 return nextValueNumber++;
321 uint32_t v = lookup_or_add(local_cdep);
322 valueNumbering[C] = v;
327 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
328 MD->getNonLocalCallDependency(CallSite(C));
329 // FIXME: Move the checking logic to MemDep!
332 // Check to see if we have a single dominating call instruction that is
334 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
335 const NonLocalDepEntry *I = &deps[i];
336 if (I->getResult().isNonLocal())
339 // We don't handle non-definitions. If we already have a call, reject
340 // instruction dependencies.
341 if (!I->getResult().isDef() || cdep != 0) {
346 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
347 // FIXME: All duplicated with non-local case.
348 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
349 cdep = NonLocalDepCall;
358 valueNumbering[C] = nextValueNumber;
359 return nextValueNumber++;
362 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
363 valueNumbering[C] = nextValueNumber;
364 return nextValueNumber++;
366 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
367 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
368 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
370 valueNumbering[C] = nextValueNumber;
371 return nextValueNumber++;
375 uint32_t v = lookup_or_add(cdep);
376 valueNumbering[C] = v;
380 valueNumbering[C] = nextValueNumber;
381 return nextValueNumber++;
385 /// lookup_or_add - Returns the value number for the specified value, assigning
386 /// it a new number if it did not have one before.
387 uint32_t ValueTable::lookup_or_add(Value *V) {
388 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
389 if (VI != valueNumbering.end())
392 if (!isa<Instruction>(V)) {
393 valueNumbering[V] = nextValueNumber;
394 return nextValueNumber++;
397 Instruction* I = cast<Instruction>(V);
399 switch (I->getOpcode()) {
400 case Instruction::Call:
401 return lookup_or_add_call(cast<CallInst>(I));
402 case Instruction::Add:
403 case Instruction::FAdd:
404 case Instruction::Sub:
405 case Instruction::FSub:
406 case Instruction::Mul:
407 case Instruction::FMul:
408 case Instruction::UDiv:
409 case Instruction::SDiv:
410 case Instruction::FDiv:
411 case Instruction::URem:
412 case Instruction::SRem:
413 case Instruction::FRem:
414 case Instruction::Shl:
415 case Instruction::LShr:
416 case Instruction::AShr:
417 case Instruction::And:
418 case Instruction::Or:
419 case Instruction::Xor:
420 case Instruction::ICmp:
421 case Instruction::FCmp:
422 case Instruction::Trunc:
423 case Instruction::ZExt:
424 case Instruction::SExt:
425 case Instruction::FPToUI:
426 case Instruction::FPToSI:
427 case Instruction::UIToFP:
428 case Instruction::SIToFP:
429 case Instruction::FPTrunc:
430 case Instruction::FPExt:
431 case Instruction::PtrToInt:
432 case Instruction::IntToPtr:
433 case Instruction::BitCast:
434 case Instruction::Select:
435 case Instruction::ExtractElement:
436 case Instruction::InsertElement:
437 case Instruction::ShuffleVector:
438 case Instruction::InsertValue:
439 case Instruction::GetElementPtr:
440 exp = create_expression(I);
442 case Instruction::ExtractValue:
443 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
446 valueNumbering[V] = nextValueNumber;
447 return nextValueNumber++;
450 uint32_t& e = expressionNumbering[exp];
451 if (!e) e = nextValueNumber++;
452 valueNumbering[V] = e;
456 /// lookup - Returns the value number of the specified value. Fails if
457 /// the value has not yet been numbered.
458 uint32_t ValueTable::lookup(Value *V) const {
459 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
460 assert(VI != valueNumbering.end() && "Value not numbered?");
464 /// lookup_or_add_cmp - Returns the value number of the given comparison,
465 /// assigning it a new number if it did not have one before. Useful when
466 /// we deduced the result of a comparison, but don't immediately have an
467 /// instruction realizing that comparison to hand.
468 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
469 CmpInst::Predicate Predicate,
470 Value *LHS, Value *RHS) {
471 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
472 uint32_t& e = expressionNumbering[exp];
473 if (!e) e = nextValueNumber++;
477 /// clear - Remove all entries from the ValueTable.
478 void ValueTable::clear() {
479 valueNumbering.clear();
480 expressionNumbering.clear();
484 /// erase - Remove a value from the value numbering.
485 void ValueTable::erase(Value *V) {
486 valueNumbering.erase(V);
489 /// verifyRemoved - Verify that the value is removed from all internal data
491 void ValueTable::verifyRemoved(const Value *V) const {
492 for (DenseMap<Value*, uint32_t>::const_iterator
493 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
494 assert(I->first != V && "Inst still occurs in value numbering map!");
498 //===----------------------------------------------------------------------===//
500 //===----------------------------------------------------------------------===//
504 struct AvailableValueInBlock {
505 /// BB - The basic block in question.
508 SimpleVal, // A simple offsetted value that is accessed.
509 LoadVal, // A value produced by a load.
510 MemIntrin, // A memory intrinsic which is loaded from.
511 UndefVal // A UndefValue representing a value from dead block (which
512 // is not yet physically removed from the CFG).
515 /// V - The value that is live out of the block.
516 PointerIntPair<Value *, 2, ValType> Val;
518 /// Offset - The byte offset in Val that is interesting for the load query.
521 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
522 unsigned Offset = 0) {
523 AvailableValueInBlock Res;
525 Res.Val.setPointer(V);
526 Res.Val.setInt(SimpleVal);
531 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
532 unsigned Offset = 0) {
533 AvailableValueInBlock Res;
535 Res.Val.setPointer(MI);
536 Res.Val.setInt(MemIntrin);
541 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
542 unsigned Offset = 0) {
543 AvailableValueInBlock Res;
545 Res.Val.setPointer(LI);
546 Res.Val.setInt(LoadVal);
551 static AvailableValueInBlock getUndef(BasicBlock *BB) {
552 AvailableValueInBlock Res;
554 Res.Val.setPointer(0);
555 Res.Val.setInt(UndefVal);
560 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
561 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
562 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
563 bool isUndefValue() const { return Val.getInt() == UndefVal; }
565 Value *getSimpleValue() const {
566 assert(isSimpleValue() && "Wrong accessor");
567 return Val.getPointer();
570 LoadInst *getCoercedLoadValue() const {
571 assert(isCoercedLoadValue() && "Wrong accessor");
572 return cast<LoadInst>(Val.getPointer());
575 MemIntrinsic *getMemIntrinValue() const {
576 assert(isMemIntrinValue() && "Wrong accessor");
577 return cast<MemIntrinsic>(Val.getPointer());
580 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
581 /// defined here to the specified type. This handles various coercion cases.
582 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const;
585 class GVN : public FunctionPass {
587 MemoryDependenceAnalysis *MD;
589 const DataLayout *TD;
590 const TargetLibraryInfo *TLI;
591 SetVector<BasicBlock *> DeadBlocks;
595 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
596 /// have that value number. Use findLeader to query it.
597 struct LeaderTableEntry {
599 const BasicBlock *BB;
600 LeaderTableEntry *Next;
602 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
603 BumpPtrAllocator TableAllocator;
605 SmallVector<Instruction*, 8> InstrsToErase;
607 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
608 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
609 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
612 static char ID; // Pass identification, replacement for typeid
613 explicit GVN(bool noloads = false)
614 : FunctionPass(ID), NoLoads(noloads), MD(0) {
615 initializeGVNPass(*PassRegistry::getPassRegistry());
618 bool runOnFunction(Function &F);
620 /// markInstructionForDeletion - This removes the specified instruction from
621 /// our various maps and marks it for deletion.
622 void markInstructionForDeletion(Instruction *I) {
624 InstrsToErase.push_back(I);
627 const DataLayout *getDataLayout() const { return TD; }
628 DominatorTree &getDominatorTree() const { return *DT; }
629 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
630 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
632 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
633 /// its value number.
634 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
635 LeaderTableEntry &Curr = LeaderTable[N];
642 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
645 Node->Next = Curr.Next;
649 /// removeFromLeaderTable - Scan the list of values corresponding to a given
650 /// value number, and remove the given instruction if encountered.
651 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
652 LeaderTableEntry* Prev = 0;
653 LeaderTableEntry* Curr = &LeaderTable[N];
655 while (Curr->Val != I || Curr->BB != BB) {
661 Prev->Next = Curr->Next;
667 LeaderTableEntry* Next = Curr->Next;
668 Curr->Val = Next->Val;
670 Curr->Next = Next->Next;
675 // List of critical edges to be split between iterations.
676 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
678 // This transformation requires dominator postdominator info
679 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
680 AU.addRequired<DominatorTreeWrapperPass>();
681 AU.addRequired<TargetLibraryInfo>();
683 AU.addRequired<MemoryDependenceAnalysis>();
684 AU.addRequired<AliasAnalysis>();
686 AU.addPreserved<DominatorTreeWrapperPass>();
687 AU.addPreserved<AliasAnalysis>();
691 // Helper fuctions of redundant load elimination
692 bool processLoad(LoadInst *L);
693 bool processNonLocalLoad(LoadInst *L);
694 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
695 AvailValInBlkVect &ValuesPerBlock,
696 UnavailBlkVect &UnavailableBlocks);
697 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
698 UnavailBlkVect &UnavailableBlocks);
700 // Other helper routines
701 bool processInstruction(Instruction *I);
702 bool processBlock(BasicBlock *BB);
703 void dump(DenseMap<uint32_t, Value*> &d);
704 bool iterateOnFunction(Function &F);
705 bool performPRE(Function &F);
706 Value *findLeader(const BasicBlock *BB, uint32_t num);
707 void cleanupGlobalSets();
708 void verifyRemoved(const Instruction *I) const;
709 bool splitCriticalEdges();
710 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
711 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
712 const BasicBlockEdge &Root);
713 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
714 bool processFoldableCondBr(BranchInst *BI);
715 void addDeadBlock(BasicBlock *BB);
716 void assignValNumForDeadCode();
722 // createGVNPass - The public interface to this file...
723 FunctionPass *llvm::createGVNPass(bool NoLoads) {
724 return new GVN(NoLoads);
727 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
728 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
729 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
730 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
731 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
732 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
734 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
735 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
737 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
738 E = d.end(); I != E; ++I) {
739 errs() << I->first << "\n";
746 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
747 /// we're analyzing is fully available in the specified block. As we go, keep
748 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
749 /// map is actually a tri-state map with the following values:
750 /// 0) we know the block *is not* fully available.
751 /// 1) we know the block *is* fully available.
752 /// 2) we do not know whether the block is fully available or not, but we are
753 /// currently speculating that it will be.
754 /// 3) we are speculating for this block and have used that to speculate for
756 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
757 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
758 uint32_t RecurseDepth) {
759 if (RecurseDepth > MaxRecurseDepth)
762 // Optimistically assume that the block is fully available and check to see
763 // if we already know about this block in one lookup.
764 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
765 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
767 // If the entry already existed for this block, return the precomputed value.
769 // If this is a speculative "available" value, mark it as being used for
770 // speculation of other blocks.
771 if (IV.first->second == 2)
772 IV.first->second = 3;
773 return IV.first->second != 0;
776 // Otherwise, see if it is fully available in all predecessors.
777 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
779 // If this block has no predecessors, it isn't live-in here.
781 goto SpeculationFailure;
783 for (; PI != PE; ++PI)
784 // If the value isn't fully available in one of our predecessors, then it
785 // isn't fully available in this block either. Undo our previous
786 // optimistic assumption and bail out.
787 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
788 goto SpeculationFailure;
792 // SpeculationFailure - If we get here, we found out that this is not, after
793 // all, a fully-available block. We have a problem if we speculated on this and
794 // used the speculation to mark other blocks as available.
796 char &BBVal = FullyAvailableBlocks[BB];
798 // If we didn't speculate on this, just return with it set to false.
804 // If we did speculate on this value, we could have blocks set to 1 that are
805 // incorrect. Walk the (transitive) successors of this block and mark them as
807 SmallVector<BasicBlock*, 32> BBWorklist;
808 BBWorklist.push_back(BB);
811 BasicBlock *Entry = BBWorklist.pop_back_val();
812 // Note that this sets blocks to 0 (unavailable) if they happen to not
813 // already be in FullyAvailableBlocks. This is safe.
814 char &EntryVal = FullyAvailableBlocks[Entry];
815 if (EntryVal == 0) continue; // Already unavailable.
817 // Mark as unavailable.
820 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
821 BBWorklist.push_back(*I);
822 } while (!BBWorklist.empty());
828 /// CanCoerceMustAliasedValueToLoad - Return true if
829 /// CoerceAvailableValueToLoadType will succeed.
830 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
832 const DataLayout &TD) {
833 // If the loaded or stored value is an first class array or struct, don't try
834 // to transform them. We need to be able to bitcast to integer.
835 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
836 StoredVal->getType()->isStructTy() ||
837 StoredVal->getType()->isArrayTy())
840 // The store has to be at least as big as the load.
841 if (TD.getTypeSizeInBits(StoredVal->getType()) <
842 TD.getTypeSizeInBits(LoadTy))
848 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
849 /// then a load from a must-aliased pointer of a different type, try to coerce
850 /// the stored value. LoadedTy is the type of the load we want to replace and
851 /// InsertPt is the place to insert new instructions.
853 /// If we can't do it, return null.
854 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
856 Instruction *InsertPt,
857 const DataLayout &TD) {
858 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
861 // If this is already the right type, just return it.
862 Type *StoredValTy = StoredVal->getType();
864 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
865 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
867 // If the store and reload are the same size, we can always reuse it.
868 if (StoreSize == LoadSize) {
869 // Pointer to Pointer -> use bitcast.
870 if (StoredValTy->getScalarType()->isPointerTy() &&
871 LoadedTy->getScalarType()->isPointerTy())
872 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
874 // Convert source pointers to integers, which can be bitcast.
875 if (StoredValTy->getScalarType()->isPointerTy()) {
876 StoredValTy = TD.getIntPtrType(StoredValTy);
877 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
880 Type *TypeToCastTo = LoadedTy;
881 if (TypeToCastTo->getScalarType()->isPointerTy())
882 TypeToCastTo = TD.getIntPtrType(TypeToCastTo);
884 if (StoredValTy != TypeToCastTo)
885 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
887 // Cast to pointer if the load needs a pointer type.
888 if (LoadedTy->getScalarType()->isPointerTy())
889 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
894 // If the loaded value is smaller than the available value, then we can
895 // extract out a piece from it. If the available value is too small, then we
896 // can't do anything.
897 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
899 // Convert source pointers to integers, which can be manipulated.
900 if (StoredValTy->getScalarType()->isPointerTy()) {
901 StoredValTy = TD.getIntPtrType(StoredValTy);
902 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
905 // Convert vectors and fp to integer, which can be manipulated.
906 if (!StoredValTy->isIntegerTy()) {
907 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
908 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
911 // If this is a big-endian system, we need to shift the value down to the low
912 // bits so that a truncate will work.
913 if (TD.isBigEndian()) {
914 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
915 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
918 // Truncate the integer to the right size now.
919 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
920 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
922 if (LoadedTy == NewIntTy)
925 // If the result is a pointer, inttoptr.
926 if (LoadedTy->getScalarType()->isPointerTy())
927 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
929 // Otherwise, bitcast.
930 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
933 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
934 /// memdep query of a load that ends up being a clobbering memory write (store,
935 /// memset, memcpy, memmove). This means that the write *may* provide bits used
936 /// by the load but we can't be sure because the pointers don't mustalias.
938 /// Check this case to see if there is anything more we can do before we give
939 /// up. This returns -1 if we have to give up, or a byte number in the stored
940 /// value of the piece that feeds the load.
941 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
943 uint64_t WriteSizeInBits,
944 const DataLayout &TD) {
945 // If the loaded or stored value is a first class array or struct, don't try
946 // to transform them. We need to be able to bitcast to integer.
947 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
950 int64_t StoreOffset = 0, LoadOffset = 0;
951 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&TD);
952 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &TD);
953 if (StoreBase != LoadBase)
956 // If the load and store are to the exact same address, they should have been
957 // a must alias. AA must have gotten confused.
958 // FIXME: Study to see if/when this happens. One case is forwarding a memset
959 // to a load from the base of the memset.
961 if (LoadOffset == StoreOffset) {
962 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
963 << "Base = " << *StoreBase << "\n"
964 << "Store Ptr = " << *WritePtr << "\n"
965 << "Store Offs = " << StoreOffset << "\n"
966 << "Load Ptr = " << *LoadPtr << "\n";
971 // If the load and store don't overlap at all, the store doesn't provide
972 // anything to the load. In this case, they really don't alias at all, AA
973 // must have gotten confused.
974 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
976 if ((WriteSizeInBits & 7) | (LoadSize & 7))
978 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
982 bool isAAFailure = false;
983 if (StoreOffset < LoadOffset)
984 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
986 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
990 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
991 << "Base = " << *StoreBase << "\n"
992 << "Store Ptr = " << *WritePtr << "\n"
993 << "Store Offs = " << StoreOffset << "\n"
994 << "Load Ptr = " << *LoadPtr << "\n";
1000 // If the Load isn't completely contained within the stored bits, we don't
1001 // have all the bits to feed it. We could do something crazy in the future
1002 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1004 if (StoreOffset > LoadOffset ||
1005 StoreOffset+StoreSize < LoadOffset+LoadSize)
1008 // Okay, we can do this transformation. Return the number of bytes into the
1009 // store that the load is.
1010 return LoadOffset-StoreOffset;
1013 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1014 /// memdep query of a load that ends up being a clobbering store.
1015 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1017 const DataLayout &TD) {
1018 // Cannot handle reading from store of first-class aggregate yet.
1019 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1020 DepSI->getValueOperand()->getType()->isArrayTy())
1023 Value *StorePtr = DepSI->getPointerOperand();
1024 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1025 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1026 StorePtr, StoreSize, TD);
1029 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
1030 /// memdep query of a load that ends up being clobbered by another load. See if
1031 /// the other load can feed into the second load.
1032 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1033 LoadInst *DepLI, const DataLayout &TD){
1034 // Cannot handle reading from store of first-class aggregate yet.
1035 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1038 Value *DepPtr = DepLI->getPointerOperand();
1039 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType());
1040 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD);
1041 if (R != -1) return R;
1043 // If we have a load/load clobber an DepLI can be widened to cover this load,
1044 // then we should widen it!
1045 int64_t LoadOffs = 0;
1046 const Value *LoadBase =
1047 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &TD);
1048 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
1050 unsigned Size = MemoryDependenceAnalysis::
1051 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD);
1052 if (Size == 0) return -1;
1054 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD);
1059 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1061 const DataLayout &TD) {
1062 // If the mem operation is a non-constant size, we can't handle it.
1063 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1064 if (SizeCst == 0) return -1;
1065 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1067 // If this is memset, we just need to see if the offset is valid in the size
1069 if (MI->getIntrinsicID() == Intrinsic::memset)
1070 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1073 // If we have a memcpy/memmove, the only case we can handle is if this is a
1074 // copy from constant memory. In that case, we can read directly from the
1076 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1078 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1079 if (Src == 0) return -1;
1081 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
1082 if (GV == 0 || !GV->isConstant()) return -1;
1084 // See if the access is within the bounds of the transfer.
1085 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1086 MI->getDest(), MemSizeInBits, TD);
1090 unsigned AS = Src->getType()->getPointerAddressSpace();
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 Type::getInt8PtrTy(Src->getContext(), AS));
1095 Constant *OffsetCst =
1096 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1097 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1098 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
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 PtrVal->getType()->getPointerAddressSpace());
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());
1250 unsigned AS = Src->getType()->getPointerAddressSpace();
1252 // Otherwise, see if we can constant fold a load from the constant with the
1253 // offset applied as appropriate.
1254 Src = ConstantExpr::getBitCast(Src,
1255 Type::getInt8PtrTy(Src->getContext(), AS));
1256 Constant *OffsetCst =
1257 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1258 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1259 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1260 return ConstantFoldLoadFromConstPtr(Src, &TD);
1264 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1265 /// construct SSA form, allowing us to eliminate LI. This returns the value
1266 /// that should be used at LI's definition site.
1267 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1268 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1270 // Check for the fully redundant, dominating load case. In this case, we can
1271 // just use the dominating value directly.
1272 if (ValuesPerBlock.size() == 1 &&
1273 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1275 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1276 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1279 // Otherwise, we have to construct SSA form.
1280 SmallVector<PHINode*, 8> NewPHIs;
1281 SSAUpdater SSAUpdate(&NewPHIs);
1282 SSAUpdate.Initialize(LI->getType(), LI->getName());
1284 Type *LoadTy = LI->getType();
1286 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1287 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1288 BasicBlock *BB = AV.BB;
1290 if (SSAUpdate.HasValueForBlock(BB))
1293 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1296 // Perform PHI construction.
1297 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1299 // If new PHI nodes were created, notify alias analysis.
1300 if (V->getType()->getScalarType()->isPointerTy()) {
1301 AliasAnalysis *AA = gvn.getAliasAnalysis();
1303 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1304 AA->copyValue(LI, NewPHIs[i]);
1306 // Now that we've copied information to the new PHIs, scan through
1307 // them again and inform alias analysis that we've added potentially
1308 // escaping uses to any values that are operands to these PHIs.
1309 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1310 PHINode *P = NewPHIs[i];
1311 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1312 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1313 AA->addEscapingUse(P->getOperandUse(jj));
1321 Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1323 if (isSimpleValue()) {
1324 Res = getSimpleValue();
1325 if (Res->getType() != LoadTy) {
1326 const DataLayout *TD = gvn.getDataLayout();
1327 assert(TD && "Need target data to handle type mismatch case");
1328 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1331 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1332 << *getSimpleValue() << '\n'
1333 << *Res << '\n' << "\n\n\n");
1335 } else if (isCoercedLoadValue()) {
1336 LoadInst *Load = getCoercedLoadValue();
1337 if (Load->getType() == LoadTy && Offset == 0) {
1340 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1343 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1344 << *getCoercedLoadValue() << '\n'
1345 << *Res << '\n' << "\n\n\n");
1347 } else if (isMemIntrinValue()) {
1348 const DataLayout *TD = gvn.getDataLayout();
1349 assert(TD && "Need target data to handle type mismatch case");
1350 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1351 LoadTy, BB->getTerminator(), *TD);
1352 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1353 << " " << *getMemIntrinValue() << '\n'
1354 << *Res << '\n' << "\n\n\n");
1356 assert(isUndefValue() && "Should be UndefVal");
1357 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1358 return UndefValue::get(LoadTy);
1363 static bool isLifetimeStart(const Instruction *Inst) {
1364 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1365 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1369 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1370 AvailValInBlkVect &ValuesPerBlock,
1371 UnavailBlkVect &UnavailableBlocks) {
1373 // Filter out useless results (non-locals, etc). Keep track of the blocks
1374 // where we have a value available in repl, also keep track of whether we see
1375 // dependencies that produce an unknown value for the load (such as a call
1376 // that could potentially clobber the load).
1377 unsigned NumDeps = Deps.size();
1378 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1379 BasicBlock *DepBB = Deps[i].getBB();
1380 MemDepResult DepInfo = Deps[i].getResult();
1382 if (DeadBlocks.count(DepBB)) {
1383 // Dead dependent mem-op disguise as a load evaluating the same value
1384 // as the load in question.
1385 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1389 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1390 UnavailableBlocks.push_back(DepBB);
1394 if (DepInfo.isClobber()) {
1395 // The address being loaded in this non-local block may not be the same as
1396 // the pointer operand of the load if PHI translation occurs. Make sure
1397 // to consider the right address.
1398 Value *Address = Deps[i].getAddress();
1400 // If the dependence is to a store that writes to a superset of the bits
1401 // read by the load, we can extract the bits we need for the load from the
1403 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1404 if (TD && Address) {
1405 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1408 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1409 DepSI->getValueOperand(),
1416 // Check to see if we have something like this:
1419 // if we have this, replace the later with an extraction from the former.
1420 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1421 // If this is a clobber and L is the first instruction in its block, then
1422 // we have the first instruction in the entry block.
1423 if (DepLI != LI && Address && TD) {
1424 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
1425 LI->getPointerOperand(),
1429 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1436 // If the clobbering value is a memset/memcpy/memmove, see if we can
1437 // forward a value on from it.
1438 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1439 if (TD && Address) {
1440 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1443 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1450 UnavailableBlocks.push_back(DepBB);
1454 // DepInfo.isDef() here
1456 Instruction *DepInst = DepInfo.getInst();
1458 // Loading the allocation -> undef.
1459 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1460 // Loading immediately after lifetime begin -> undef.
1461 isLifetimeStart(DepInst)) {
1462 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1463 UndefValue::get(LI->getType())));
1467 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1468 // Reject loads and stores that are to the same address but are of
1469 // different types if we have to.
1470 if (S->getValueOperand()->getType() != LI->getType()) {
1471 // If the stored value is larger or equal to the loaded value, we can
1473 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1474 LI->getType(), *TD)) {
1475 UnavailableBlocks.push_back(DepBB);
1480 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1481 S->getValueOperand()));
1485 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1486 // If the types mismatch and we can't handle it, reject reuse of the load.
1487 if (LD->getType() != LI->getType()) {
1488 // If the stored value is larger or equal to the loaded value, we can
1490 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1491 UnavailableBlocks.push_back(DepBB);
1495 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1499 UnavailableBlocks.push_back(DepBB);
1503 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1504 UnavailBlkVect &UnavailableBlocks) {
1505 // Okay, we have *some* definitions of the value. This means that the value
1506 // is available in some of our (transitive) predecessors. Lets think about
1507 // doing PRE of this load. This will involve inserting a new load into the
1508 // predecessor when it's not available. We could do this in general, but
1509 // prefer to not increase code size. As such, we only do this when we know
1510 // that we only have to insert *one* load (which means we're basically moving
1511 // the load, not inserting a new one).
1513 SmallPtrSet<BasicBlock *, 4> Blockers;
1514 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1515 Blockers.insert(UnavailableBlocks[i]);
1517 // Let's find the first basic block with more than one predecessor. Walk
1518 // backwards through predecessors if needed.
1519 BasicBlock *LoadBB = LI->getParent();
1520 BasicBlock *TmpBB = LoadBB;
1522 while (TmpBB->getSinglePredecessor()) {
1523 TmpBB = TmpBB->getSinglePredecessor();
1524 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1526 if (Blockers.count(TmpBB))
1529 // If any of these blocks has more than one successor (i.e. if the edge we
1530 // just traversed was critical), then there are other paths through this
1531 // block along which the load may not be anticipated. Hoisting the load
1532 // above this block would be adding the load to execution paths along
1533 // which it was not previously executed.
1534 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1541 // Check to see how many predecessors have the loaded value fully
1543 DenseMap<BasicBlock*, Value*> PredLoads;
1544 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1545 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1546 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1547 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1548 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1550 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1551 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1553 BasicBlock *Pred = *PI;
1554 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1557 PredLoads[Pred] = 0;
1559 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1560 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1561 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1562 << Pred->getName() << "': " << *LI << '\n');
1566 if (LoadBB->isLandingPad()) {
1568 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1569 << Pred->getName() << "': " << *LI << '\n');
1573 CriticalEdgePred.push_back(Pred);
1577 // Decide whether PRE is profitable for this load.
1578 unsigned NumUnavailablePreds = PredLoads.size();
1579 assert(NumUnavailablePreds != 0 &&
1580 "Fully available value should already be eliminated!");
1582 // If this load is unavailable in multiple predecessors, reject it.
1583 // FIXME: If we could restructure the CFG, we could make a common pred with
1584 // all the preds that don't have an available LI and insert a new load into
1586 if (NumUnavailablePreds != 1)
1589 // Split critical edges, and update the unavailable predecessors accordingly.
1590 for (SmallVectorImpl<BasicBlock *>::iterator I = CriticalEdgePred.begin(),
1591 E = CriticalEdgePred.end(); I != E; I++) {
1592 BasicBlock *OrigPred = *I;
1593 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1594 PredLoads.erase(OrigPred);
1595 PredLoads[NewPred] = 0;
1596 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1597 << LoadBB->getName() << '\n');
1600 // Check if the load can safely be moved to all the unavailable predecessors.
1601 bool CanDoPRE = true;
1602 SmallVector<Instruction*, 8> NewInsts;
1603 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1604 E = PredLoads.end(); I != E; ++I) {
1605 BasicBlock *UnavailablePred = I->first;
1607 // Do PHI translation to get its value in the predecessor if necessary. The
1608 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1610 // If all preds have a single successor, then we know it is safe to insert
1611 // the load on the pred (?!?), so we can insert code to materialize the
1612 // pointer if it is not available.
1613 PHITransAddr Address(LI->getPointerOperand(), TD);
1615 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1618 // If we couldn't find or insert a computation of this phi translated value,
1621 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1622 << *LI->getPointerOperand() << "\n");
1627 I->second = LoadPtr;
1631 while (!NewInsts.empty()) {
1632 Instruction *I = NewInsts.pop_back_val();
1633 if (MD) MD->removeInstruction(I);
1634 I->eraseFromParent();
1636 // HINT:Don't revert the edge-splitting as following transformation may
1637 // also need to split these critial edges.
1638 return !CriticalEdgePred.empty();
1641 // Okay, we can eliminate this load by inserting a reload in the predecessor
1642 // and using PHI construction to get the value in the other predecessors, do
1644 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1645 DEBUG(if (!NewInsts.empty())
1646 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1647 << *NewInsts.back() << '\n');
1649 // Assign value numbers to the new instructions.
1650 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1651 // FIXME: We really _ought_ to insert these value numbers into their
1652 // parent's availability map. However, in doing so, we risk getting into
1653 // ordering issues. If a block hasn't been processed yet, we would be
1654 // marking a value as AVAIL-IN, which isn't what we intend.
1655 VN.lookup_or_add(NewInsts[i]);
1658 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1659 E = PredLoads.end(); I != E; ++I) {
1660 BasicBlock *UnavailablePred = I->first;
1661 Value *LoadPtr = I->second;
1663 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1665 UnavailablePred->getTerminator());
1667 // Transfer the old load's TBAA tag to the new load.
1668 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1669 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1671 // Transfer DebugLoc.
1672 NewLoad->setDebugLoc(LI->getDebugLoc());
1674 // Add the newly created load.
1675 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1677 MD->invalidateCachedPointerInfo(LoadPtr);
1678 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1681 // Perform PHI construction.
1682 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1683 LI->replaceAllUsesWith(V);
1684 if (isa<PHINode>(V))
1686 if (V->getType()->getScalarType()->isPointerTy())
1687 MD->invalidateCachedPointerInfo(V);
1688 markInstructionForDeletion(LI);
1693 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1694 /// non-local by performing PHI construction.
1695 bool GVN::processNonLocalLoad(LoadInst *LI) {
1696 // Step 1: Find the non-local dependencies of the load.
1698 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1699 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1701 // If we had to process more than one hundred blocks to find the
1702 // dependencies, this load isn't worth worrying about. Optimizing
1703 // it will be too expensive.
1704 unsigned NumDeps = Deps.size();
1708 // If we had a phi translation failure, we'll have a single entry which is a
1709 // clobber in the current block. Reject this early.
1711 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1713 dbgs() << "GVN: non-local load ";
1714 LI->printAsOperand(dbgs());
1715 dbgs() << " has unknown dependencies\n";
1720 // Step 2: Analyze the availability of the load
1721 AvailValInBlkVect ValuesPerBlock;
1722 UnavailBlkVect UnavailableBlocks;
1723 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1725 // If we have no predecessors that produce a known value for this load, exit
1727 if (ValuesPerBlock.empty())
1730 // Step 3: Eliminate fully redundancy.
1732 // If all of the instructions we depend on produce a known value for this
1733 // load, then it is fully redundant and we can use PHI insertion to compute
1734 // its value. Insert PHIs and remove the fully redundant value now.
1735 if (UnavailableBlocks.empty()) {
1736 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1738 // Perform PHI construction.
1739 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1740 LI->replaceAllUsesWith(V);
1742 if (isa<PHINode>(V))
1744 if (V->getType()->getScalarType()->isPointerTy())
1745 MD->invalidateCachedPointerInfo(V);
1746 markInstructionForDeletion(LI);
1751 // Step 4: Eliminate partial redundancy.
1752 if (!EnablePRE || !EnableLoadPRE)
1755 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1759 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1760 // Patch the replacement so that it is not more restrictive than the value
1762 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1763 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1764 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) &&
1765 isa<OverflowingBinaryOperator>(ReplOp)) {
1766 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap())
1767 ReplOp->setHasNoSignedWrap(false);
1768 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap())
1769 ReplOp->setHasNoUnsignedWrap(false);
1771 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1772 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
1773 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata);
1774 for (int i = 0, n = Metadata.size(); i < n; ++i) {
1775 unsigned Kind = Metadata[i].first;
1776 MDNode *IMD = I->getMetadata(Kind);
1777 MDNode *ReplMD = Metadata[i].second;
1780 ReplInst->setMetadata(Kind, NULL); // Remove unknown metadata
1782 case LLVMContext::MD_dbg:
1783 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1784 case LLVMContext::MD_tbaa:
1785 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD));
1787 case LLVMContext::MD_range:
1788 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD));
1790 case LLVMContext::MD_prof:
1791 llvm_unreachable("MD_prof in a non-terminator instruction");
1793 case LLVMContext::MD_fpmath:
1794 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD));
1801 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1802 patchReplacementInstruction(I, Repl);
1803 I->replaceAllUsesWith(Repl);
1806 /// processLoad - Attempt to eliminate a load, first by eliminating it
1807 /// locally, and then attempting non-local elimination if that fails.
1808 bool GVN::processLoad(LoadInst *L) {
1815 if (L->use_empty()) {
1816 markInstructionForDeletion(L);
1820 // ... to a pointer that has been loaded from before...
1821 MemDepResult Dep = MD->getDependency(L);
1823 // If we have a clobber and target data is around, see if this is a clobber
1824 // that we can fix up through code synthesis.
1825 if (Dep.isClobber() && TD) {
1826 // Check to see if we have something like this:
1827 // store i32 123, i32* %P
1828 // %A = bitcast i32* %P to i8*
1829 // %B = gep i8* %A, i32 1
1832 // We could do that by recognizing if the clobber instructions are obviously
1833 // a common base + constant offset, and if the previous store (or memset)
1834 // completely covers this load. This sort of thing can happen in bitfield
1836 Value *AvailVal = 0;
1837 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1838 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1839 L->getPointerOperand(),
1842 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1843 L->getType(), L, *TD);
1846 // Check to see if we have something like this:
1849 // if we have this, replace the later with an extraction from the former.
1850 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1851 // If this is a clobber and L is the first instruction in its block, then
1852 // we have the first instruction in the entry block.
1856 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1857 L->getPointerOperand(),
1860 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1863 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1864 // a value on from it.
1865 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1866 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1867 L->getPointerOperand(),
1870 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
1874 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1875 << *AvailVal << '\n' << *L << "\n\n\n");
1877 // Replace the load!
1878 L->replaceAllUsesWith(AvailVal);
1879 if (AvailVal->getType()->getScalarType()->isPointerTy())
1880 MD->invalidateCachedPointerInfo(AvailVal);
1881 markInstructionForDeletion(L);
1887 // If the value isn't available, don't do anything!
1888 if (Dep.isClobber()) {
1890 // fast print dep, using operator<< on instruction is too slow.
1891 dbgs() << "GVN: load ";
1892 L->printAsOperand(dbgs());
1893 Instruction *I = Dep.getInst();
1894 dbgs() << " is clobbered by " << *I << '\n';
1899 // If it is defined in another block, try harder.
1900 if (Dep.isNonLocal())
1901 return processNonLocalLoad(L);
1905 // fast print dep, using operator<< on instruction is too slow.
1906 dbgs() << "GVN: load ";
1907 L->printAsOperand(dbgs());
1908 dbgs() << " has unknown dependence\n";
1913 Instruction *DepInst = Dep.getInst();
1914 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1915 Value *StoredVal = DepSI->getValueOperand();
1917 // The store and load are to a must-aliased pointer, but they may not
1918 // actually have the same type. See if we know how to reuse the stored
1919 // value (depending on its type).
1920 if (StoredVal->getType() != L->getType()) {
1922 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1927 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1928 << '\n' << *L << "\n\n\n");
1935 L->replaceAllUsesWith(StoredVal);
1936 if (StoredVal->getType()->getScalarType()->isPointerTy())
1937 MD->invalidateCachedPointerInfo(StoredVal);
1938 markInstructionForDeletion(L);
1943 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1944 Value *AvailableVal = DepLI;
1946 // The loads are of a must-aliased pointer, but they may not actually have
1947 // the same type. See if we know how to reuse the previously loaded value
1948 // (depending on its type).
1949 if (DepLI->getType() != L->getType()) {
1951 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1953 if (AvailableVal == 0)
1956 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1957 << "\n" << *L << "\n\n\n");
1964 patchAndReplaceAllUsesWith(L, AvailableVal);
1965 if (DepLI->getType()->getScalarType()->isPointerTy())
1966 MD->invalidateCachedPointerInfo(DepLI);
1967 markInstructionForDeletion(L);
1972 // If this load really doesn't depend on anything, then we must be loading an
1973 // undef value. This can happen when loading for a fresh allocation with no
1974 // intervening stores, for example.
1975 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1976 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1977 markInstructionForDeletion(L);
1982 // If this load occurs either right after a lifetime begin,
1983 // then the loaded value is undefined.
1984 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1985 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1986 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1987 markInstructionForDeletion(L);
1996 // findLeader - In order to find a leader for a given value number at a
1997 // specific basic block, we first obtain the list of all Values for that number,
1998 // and then scan the list to find one whose block dominates the block in
1999 // question. This is fast because dominator tree queries consist of only
2000 // a few comparisons of DFS numbers.
2001 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
2002 LeaderTableEntry Vals = LeaderTable[num];
2003 if (!Vals.Val) return 0;
2006 if (DT->dominates(Vals.BB, BB)) {
2008 if (isa<Constant>(Val)) return Val;
2011 LeaderTableEntry* Next = Vals.Next;
2013 if (DT->dominates(Next->BB, BB)) {
2014 if (isa<Constant>(Next->Val)) return Next->Val;
2015 if (!Val) Val = Next->Val;
2024 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
2025 /// use is dominated by the given basic block. Returns the number of uses that
2027 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
2028 const BasicBlockEdge &Root) {
2030 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2032 Use &U = (UI++).getUse();
2034 if (DT->dominates(Root, U)) {
2042 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
2043 /// true if every path from the entry block to 'Dst' passes via this edge. In
2044 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2045 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2046 DominatorTree *DT) {
2047 // While in theory it is interesting to consider the case in which Dst has
2048 // more than one predecessor, because Dst might be part of a loop which is
2049 // only reachable from Src, in practice it is pointless since at the time
2050 // GVN runs all such loops have preheaders, which means that Dst will have
2051 // been changed to have only one predecessor, namely Src.
2052 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2053 const BasicBlock *Src = E.getStart();
2054 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2059 /// propagateEquality - The given values are known to be equal in every block
2060 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2061 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2062 bool GVN::propagateEquality(Value *LHS, Value *RHS,
2063 const BasicBlockEdge &Root) {
2064 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2065 Worklist.push_back(std::make_pair(LHS, RHS));
2066 bool Changed = false;
2067 // For speed, compute a conservative fast approximation to
2068 // DT->dominates(Root, Root.getEnd());
2069 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2071 while (!Worklist.empty()) {
2072 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2073 LHS = Item.first; RHS = Item.second;
2075 if (LHS == RHS) continue;
2076 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2078 // Don't try to propagate equalities between constants.
2079 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2081 // Prefer a constant on the right-hand side, or an Argument if no constants.
2082 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2083 std::swap(LHS, RHS);
2084 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2086 // If there is no obvious reason to prefer the left-hand side over the right-
2087 // hand side, ensure the longest lived term is on the right-hand side, so the
2088 // shortest lived term will be replaced by the longest lived. This tends to
2089 // expose more simplifications.
2090 uint32_t LVN = VN.lookup_or_add(LHS);
2091 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2092 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2093 // Move the 'oldest' value to the right-hand side, using the value number as
2095 uint32_t RVN = VN.lookup_or_add(RHS);
2097 std::swap(LHS, RHS);
2102 // If value numbering later sees that an instruction in the scope is equal
2103 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2104 // the invariant that instructions only occur in the leader table for their
2105 // own value number (this is used by removeFromLeaderTable), do not do this
2106 // if RHS is an instruction (if an instruction in the scope is morphed into
2107 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2108 // using the leader table is about compiling faster, not optimizing better).
2109 // The leader table only tracks basic blocks, not edges. Only add to if we
2110 // have the simple case where the edge dominates the end.
2111 if (RootDominatesEnd && !isa<Instruction>(RHS))
2112 addToLeaderTable(LVN, RHS, Root.getEnd());
2114 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2115 // LHS always has at least one use that is not dominated by Root, this will
2116 // never do anything if LHS has only one use.
2117 if (!LHS->hasOneUse()) {
2118 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2119 Changed |= NumReplacements > 0;
2120 NumGVNEqProp += NumReplacements;
2123 // Now try to deduce additional equalities from this one. For example, if the
2124 // known equality was "(A != B)" == "false" then it follows that A and B are
2125 // equal in the scope. Only boolean equalities with an explicit true or false
2126 // RHS are currently supported.
2127 if (!RHS->getType()->isIntegerTy(1))
2128 // Not a boolean equality - bail out.
2130 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2132 // RHS neither 'true' nor 'false' - bail out.
2134 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2135 bool isKnownTrue = CI->isAllOnesValue();
2136 bool isKnownFalse = !isKnownTrue;
2138 // If "A && B" is known true then both A and B are known true. If "A || B"
2139 // is known false then both A and B are known false.
2141 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2142 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2143 Worklist.push_back(std::make_pair(A, RHS));
2144 Worklist.push_back(std::make_pair(B, RHS));
2148 // If we are propagating an equality like "(A == B)" == "true" then also
2149 // propagate the equality A == B. When propagating a comparison such as
2150 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2151 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
2152 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2154 // If "A == B" is known true, or "A != B" is known false, then replace
2155 // A with B everywhere in the scope.
2156 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2157 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2158 Worklist.push_back(std::make_pair(Op0, Op1));
2160 // If "A >= B" is known true, replace "A < B" with false everywhere.
2161 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2162 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2163 // Since we don't have the instruction "A < B" immediately to hand, work out
2164 // the value number that it would have and use that to find an appropriate
2165 // instruction (if any).
2166 uint32_t NextNum = VN.getNextUnusedValueNumber();
2167 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2168 // If the number we were assigned was brand new then there is no point in
2169 // looking for an instruction realizing it: there cannot be one!
2170 if (Num < NextNum) {
2171 Value *NotCmp = findLeader(Root.getEnd(), Num);
2172 if (NotCmp && isa<Instruction>(NotCmp)) {
2173 unsigned NumReplacements =
2174 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2175 Changed |= NumReplacements > 0;
2176 NumGVNEqProp += NumReplacements;
2179 // Ensure that any instruction in scope that gets the "A < B" value number
2180 // is replaced with false.
2181 // The leader table only tracks basic blocks, not edges. Only add to if we
2182 // have the simple case where the edge dominates the end.
2183 if (RootDominatesEnd)
2184 addToLeaderTable(Num, NotVal, Root.getEnd());
2193 /// processInstruction - When calculating availability, handle an instruction
2194 /// by inserting it into the appropriate sets
2195 bool GVN::processInstruction(Instruction *I) {
2196 // Ignore dbg info intrinsics.
2197 if (isa<DbgInfoIntrinsic>(I))
2200 // If the instruction can be easily simplified then do so now in preference
2201 // to value numbering it. Value numbering often exposes redundancies, for
2202 // example if it determines that %y is equal to %x then the instruction
2203 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2204 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) {
2205 I->replaceAllUsesWith(V);
2206 if (MD && V->getType()->getScalarType()->isPointerTy())
2207 MD->invalidateCachedPointerInfo(V);
2208 markInstructionForDeletion(I);
2213 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2214 if (processLoad(LI))
2217 unsigned Num = VN.lookup_or_add(LI);
2218 addToLeaderTable(Num, LI, LI->getParent());
2222 // For conditional branches, we can perform simple conditional propagation on
2223 // the condition value itself.
2224 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2225 if (!BI->isConditional())
2228 if (isa<Constant>(BI->getCondition()))
2229 return processFoldableCondBr(BI);
2231 Value *BranchCond = BI->getCondition();
2232 BasicBlock *TrueSucc = BI->getSuccessor(0);
2233 BasicBlock *FalseSucc = BI->getSuccessor(1);
2234 // Avoid multiple edges early.
2235 if (TrueSucc == FalseSucc)
2238 BasicBlock *Parent = BI->getParent();
2239 bool Changed = false;
2241 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2242 BasicBlockEdge TrueE(Parent, TrueSucc);
2243 Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
2245 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2246 BasicBlockEdge FalseE(Parent, FalseSucc);
2247 Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
2252 // For switches, propagate the case values into the case destinations.
2253 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2254 Value *SwitchCond = SI->getCondition();
2255 BasicBlock *Parent = SI->getParent();
2256 bool Changed = false;
2258 // Remember how many outgoing edges there are to every successor.
2259 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2260 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2261 ++SwitchEdges[SI->getSuccessor(i)];
2263 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2265 BasicBlock *Dst = i.getCaseSuccessor();
2266 // If there is only a single edge, propagate the case value into it.
2267 if (SwitchEdges.lookup(Dst) == 1) {
2268 BasicBlockEdge E(Parent, Dst);
2269 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
2275 // Instructions with void type don't return a value, so there's
2276 // no point in trying to find redundancies in them.
2277 if (I->getType()->isVoidTy()) return false;
2279 uint32_t NextNum = VN.getNextUnusedValueNumber();
2280 unsigned Num = VN.lookup_or_add(I);
2282 // Allocations are always uniquely numbered, so we can save time and memory
2283 // by fast failing them.
2284 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2285 addToLeaderTable(Num, I, I->getParent());
2289 // If the number we were assigned was a brand new VN, then we don't
2290 // need to do a lookup to see if the number already exists
2291 // somewhere in the domtree: it can't!
2292 if (Num >= NextNum) {
2293 addToLeaderTable(Num, I, I->getParent());
2297 // Perform fast-path value-number based elimination of values inherited from
2299 Value *repl = findLeader(I->getParent(), Num);
2301 // Failure, just remember this instance for future use.
2302 addToLeaderTable(Num, I, I->getParent());
2307 patchAndReplaceAllUsesWith(I, repl);
2308 if (MD && repl->getType()->getScalarType()->isPointerTy())
2309 MD->invalidateCachedPointerInfo(repl);
2310 markInstructionForDeletion(I);
2314 /// runOnFunction - This is the main transformation entry point for a function.
2315 bool GVN::runOnFunction(Function& F) {
2317 MD = &getAnalysis<MemoryDependenceAnalysis>();
2318 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2319 TD = getAnalysisIfAvailable<DataLayout>();
2320 TLI = &getAnalysis<TargetLibraryInfo>();
2321 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2325 bool Changed = false;
2326 bool ShouldContinue = true;
2328 // Merge unconditional branches, allowing PRE to catch more
2329 // optimization opportunities.
2330 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2331 BasicBlock *BB = FI++;
2333 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2334 if (removedBlock) ++NumGVNBlocks;
2336 Changed |= removedBlock;
2339 unsigned Iteration = 0;
2340 while (ShouldContinue) {
2341 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2342 ShouldContinue = iterateOnFunction(F);
2343 Changed |= ShouldContinue;
2348 // Fabricate val-num for dead-code in order to suppress assertion in
2350 assignValNumForDeadCode();
2351 bool PREChanged = true;
2352 while (PREChanged) {
2353 PREChanged = performPRE(F);
2354 Changed |= PREChanged;
2358 // FIXME: Should perform GVN again after PRE does something. PRE can move
2359 // computations into blocks where they become fully redundant. Note that
2360 // we can't do this until PRE's critical edge splitting updates memdep.
2361 // Actually, when this happens, we should just fully integrate PRE into GVN.
2363 cleanupGlobalSets();
2364 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2372 bool GVN::processBlock(BasicBlock *BB) {
2373 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2374 // (and incrementing BI before processing an instruction).
2375 assert(InstrsToErase.empty() &&
2376 "We expect InstrsToErase to be empty across iterations");
2377 if (DeadBlocks.count(BB))
2380 bool ChangedFunction = false;
2382 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2384 ChangedFunction |= processInstruction(BI);
2385 if (InstrsToErase.empty()) {
2390 // If we need some instructions deleted, do it now.
2391 NumGVNInstr += InstrsToErase.size();
2393 // Avoid iterator invalidation.
2394 bool AtStart = BI == BB->begin();
2398 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2399 E = InstrsToErase.end(); I != E; ++I) {
2400 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2401 if (MD) MD->removeInstruction(*I);
2402 DEBUG(verifyRemoved(*I));
2403 (*I)->eraseFromParent();
2405 InstrsToErase.clear();
2413 return ChangedFunction;
2416 /// performPRE - Perform a purely local form of PRE that looks for diamond
2417 /// control flow patterns and attempts to perform simple PRE at the join point.
2418 bool GVN::performPRE(Function &F) {
2419 bool Changed = false;
2420 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2421 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2422 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2423 BasicBlock *CurrentBlock = *DI;
2425 // Nothing to PRE in the entry block.
2426 if (CurrentBlock == &F.getEntryBlock()) continue;
2428 // Don't perform PRE on a landing pad.
2429 if (CurrentBlock->isLandingPad()) continue;
2431 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2432 BE = CurrentBlock->end(); BI != BE; ) {
2433 Instruction *CurInst = BI++;
2435 if (isa<AllocaInst>(CurInst) ||
2436 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2437 CurInst->getType()->isVoidTy() ||
2438 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2439 isa<DbgInfoIntrinsic>(CurInst))
2442 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2443 // sinking the compare again, and it would force the code generator to
2444 // move the i1 from processor flags or predicate registers into a general
2445 // purpose register.
2446 if (isa<CmpInst>(CurInst))
2449 // We don't currently value number ANY inline asm calls.
2450 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2451 if (CallI->isInlineAsm())
2454 uint32_t ValNo = VN.lookup(CurInst);
2456 // Look for the predecessors for PRE opportunities. We're
2457 // only trying to solve the basic diamond case, where
2458 // a value is computed in the successor and one predecessor,
2459 // but not the other. We also explicitly disallow cases
2460 // where the successor is its own predecessor, because they're
2461 // more complicated to get right.
2462 unsigned NumWith = 0;
2463 unsigned NumWithout = 0;
2464 BasicBlock *PREPred = 0;
2467 for (pred_iterator PI = pred_begin(CurrentBlock),
2468 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2469 BasicBlock *P = *PI;
2470 // We're not interested in PRE where the block is its
2471 // own predecessor, or in blocks with predecessors
2472 // that are not reachable.
2473 if (P == CurrentBlock) {
2476 } else if (!DT->isReachableFromEntry(P)) {
2481 Value* predV = findLeader(P, ValNo);
2483 predMap.push_back(std::make_pair(static_cast<Value *>(0), P));
2486 } else if (predV == CurInst) {
2487 /* CurInst dominates this predecessor. */
2491 predMap.push_back(std::make_pair(predV, P));
2496 // Don't do PRE when it might increase code size, i.e. when
2497 // we would need to insert instructions in more than one pred.
2498 if (NumWithout != 1 || NumWith == 0)
2501 // Don't do PRE across indirect branch.
2502 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2505 // We can't do PRE safely on a critical edge, so instead we schedule
2506 // the edge to be split and perform the PRE the next time we iterate
2508 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2509 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2510 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2514 // Instantiate the expression in the predecessor that lacked it.
2515 // Because we are going top-down through the block, all value numbers
2516 // will be available in the predecessor by the time we need them. Any
2517 // that weren't originally present will have been instantiated earlier
2519 Instruction *PREInstr = CurInst->clone();
2520 bool success = true;
2521 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2522 Value *Op = PREInstr->getOperand(i);
2523 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2526 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2527 PREInstr->setOperand(i, V);
2534 // Fail out if we encounter an operand that is not available in
2535 // the PRE predecessor. This is typically because of loads which
2536 // are not value numbered precisely.
2538 DEBUG(verifyRemoved(PREInstr));
2543 PREInstr->insertBefore(PREPred->getTerminator());
2544 PREInstr->setName(CurInst->getName() + ".pre");
2545 PREInstr->setDebugLoc(CurInst->getDebugLoc());
2546 VN.add(PREInstr, ValNo);
2549 // Update the availability map to include the new instruction.
2550 addToLeaderTable(ValNo, PREInstr, PREPred);
2552 // Create a PHI to make the value available in this block.
2553 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(),
2554 CurInst->getName() + ".pre-phi",
2555 CurrentBlock->begin());
2556 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2557 if (Value *V = predMap[i].first)
2558 Phi->addIncoming(V, predMap[i].second);
2560 Phi->addIncoming(PREInstr, PREPred);
2564 addToLeaderTable(ValNo, Phi, CurrentBlock);
2565 Phi->setDebugLoc(CurInst->getDebugLoc());
2566 CurInst->replaceAllUsesWith(Phi);
2567 if (Phi->getType()->getScalarType()->isPointerTy()) {
2568 // Because we have added a PHI-use of the pointer value, it has now
2569 // "escaped" from alias analysis' perspective. We need to inform
2571 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
2573 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2574 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2578 MD->invalidateCachedPointerInfo(Phi);
2581 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2583 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2584 if (MD) MD->removeInstruction(CurInst);
2585 DEBUG(verifyRemoved(CurInst));
2586 CurInst->eraseFromParent();
2591 if (splitCriticalEdges())
2597 /// Split the critical edge connecting the given two blocks, and return
2598 /// the block inserted to the critical edge.
2599 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2600 BasicBlock *BB = SplitCriticalEdge(Pred, Succ, this);
2602 MD->invalidateCachedPredecessors();
2606 /// splitCriticalEdges - Split critical edges found during the previous
2607 /// iteration that may enable further optimization.
2608 bool GVN::splitCriticalEdges() {
2609 if (toSplit.empty())
2612 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2613 SplitCriticalEdge(Edge.first, Edge.second, this);
2614 } while (!toSplit.empty());
2615 if (MD) MD->invalidateCachedPredecessors();
2619 /// iterateOnFunction - Executes one iteration of GVN
2620 bool GVN::iterateOnFunction(Function &F) {
2621 cleanupGlobalSets();
2623 // Top-down walk of the dominator tree
2624 bool Changed = false;
2626 // Needed for value numbering with phi construction to work.
2627 ReversePostOrderTraversal<Function*> RPOT(&F);
2628 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2629 RE = RPOT.end(); RI != RE; ++RI)
2630 Changed |= processBlock(*RI);
2632 // Save the blocks this function have before transformation begins. GVN may
2633 // split critical edge, and hence may invalidate the RPO/DT iterator.
2635 std::vector<BasicBlock *> BBVect;
2636 BBVect.reserve(256);
2637 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2638 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2639 BBVect.push_back(DI->getBlock());
2641 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2643 Changed |= processBlock(*I);
2649 void GVN::cleanupGlobalSets() {
2651 LeaderTable.clear();
2652 TableAllocator.Reset();
2655 /// verifyRemoved - Verify that the specified instruction does not occur in our
2656 /// internal data structures.
2657 void GVN::verifyRemoved(const Instruction *Inst) const {
2658 VN.verifyRemoved(Inst);
2660 // Walk through the value number scope to make sure the instruction isn't
2661 // ferreted away in it.
2662 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2663 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2664 const LeaderTableEntry *Node = &I->second;
2665 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2667 while (Node->Next) {
2669 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2674 // BB is declared dead, which implied other blocks become dead as well. This
2675 // function is to add all these blocks to "DeadBlocks". For the dead blocks'
2676 // live successors, update their phi nodes by replacing the operands
2677 // corresponding to dead blocks with UndefVal.
2679 void GVN::addDeadBlock(BasicBlock *BB) {
2680 SmallVector<BasicBlock *, 4> NewDead;
2681 SmallSetVector<BasicBlock *, 4> DF;
2683 NewDead.push_back(BB);
2684 while (!NewDead.empty()) {
2685 BasicBlock *D = NewDead.pop_back_val();
2686 if (DeadBlocks.count(D))
2689 // All blocks dominated by D are dead.
2690 SmallVector<BasicBlock *, 8> Dom;
2691 DT->getDescendants(D, Dom);
2692 DeadBlocks.insert(Dom.begin(), Dom.end());
2694 // Figure out the dominance-frontier(D).
2695 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
2696 E = Dom.end(); I != E; I++) {
2698 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
2699 BasicBlock *S = *SI;
2700 if (DeadBlocks.count(S))
2703 bool AllPredDead = true;
2704 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
2705 if (!DeadBlocks.count(*PI)) {
2706 AllPredDead = false;
2711 // S could be proved dead later on. That is why we don't update phi
2712 // operands at this moment.
2715 // While S is not dominated by D, it is dead by now. This could take
2716 // place if S already have a dead predecessor before D is declared
2718 NewDead.push_back(S);
2724 // For the dead blocks' live successors, update their phi nodes by replacing
2725 // the operands corresponding to dead blocks with UndefVal.
2726 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2729 if (DeadBlocks.count(B))
2732 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2733 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
2734 PE = Preds.end(); PI != PE; PI++) {
2735 BasicBlock *P = *PI;
2737 if (!DeadBlocks.count(P))
2740 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2741 if (BasicBlock *S = splitCriticalEdges(P, B))
2742 DeadBlocks.insert(P = S);
2745 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2746 PHINode &Phi = cast<PHINode>(*II);
2747 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2748 UndefValue::get(Phi.getType()));
2754 // If the given branch is recognized as a foldable branch (i.e. conditional
2755 // branch with constant condition), it will perform following analyses and
2757 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2758 // R be the target of the dead out-coming edge.
2759 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2760 // edge. The result of this step will be {X| X is dominated by R}
2761 // 2) Identify those blocks which haves at least one dead prodecessor. The
2762 // result of this step will be dominance-frontier(R).
2763 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2764 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2766 // Return true iff *NEW* dead code are found.
2767 bool GVN::processFoldableCondBr(BranchInst *BI) {
2768 if (!BI || BI->isUnconditional())
2771 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2775 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2776 BI->getSuccessor(1) : BI->getSuccessor(0);
2777 if (DeadBlocks.count(DeadRoot))
2780 if (!DeadRoot->getSinglePredecessor())
2781 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2783 addDeadBlock(DeadRoot);
2787 // performPRE() will trigger assert if it come across an instruciton without
2788 // associated val-num. As it normally has far more live instructions than dead
2789 // instructions, it makes more sense just to "fabricate" a val-number for the
2790 // dead code than checking if instruction involved is dead or not.
2791 void GVN::assignValNumForDeadCode() {
2792 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
2793 E = DeadBlocks.end(); I != E; I++) {
2794 BasicBlock *BB = *I;
2795 for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
2797 Instruction *Inst = &*II;
2798 unsigned ValNum = VN.lookup_or_add(Inst);
2799 addToLeaderTable(ValNum, Inst, BB);