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 #include "llvm/Transforms/Scalar.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/Hashing.h"
22 #include "llvm/ADT/MapVector.h"
23 #include "llvm/ADT/PostOrderIterator.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/ADT/SmallPtrSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AssumptionCache.h"
29 #include "llvm/Analysis/CFG.h"
30 #include "llvm/Analysis/ConstantFolding.h"
31 #include "llvm/Analysis/GlobalsModRef.h"
32 #include "llvm/Analysis/InstructionSimplify.h"
33 #include "llvm/Analysis/Loads.h"
34 #include "llvm/Analysis/MemoryBuiltins.h"
35 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
36 #include "llvm/Analysis/PHITransAddr.h"
37 #include "llvm/Analysis/TargetLibraryInfo.h"
38 #include "llvm/Analysis/ValueTracking.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/IntrinsicInst.h"
44 #include "llvm/IR/LLVMContext.h"
45 #include "llvm/IR/Metadata.h"
46 #include "llvm/IR/PatternMatch.h"
47 #include "llvm/Support/Allocator.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/raw_ostream.h"
51 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
52 #include "llvm/Transforms/Utils/Local.h"
53 #include "llvm/Transforms/Utils/SSAUpdater.h"
56 using namespace PatternMatch;
58 #define DEBUG_TYPE "gvn"
60 STATISTIC(NumGVNInstr, "Number of instructions deleted");
61 STATISTIC(NumGVNLoad, "Number of loads deleted");
62 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
63 STATISTIC(NumGVNBlocks, "Number of blocks merged");
64 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
65 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
66 STATISTIC(NumPRELoad, "Number of loads PRE'd");
68 static cl::opt<bool> EnablePRE("enable-pre",
69 cl::init(true), cl::Hidden);
70 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
72 // Maximum allowed recursion depth.
73 static cl::opt<uint32_t>
74 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
75 cl::desc("Max recurse depth (default = 1000)"));
77 //===----------------------------------------------------------------------===//
79 //===----------------------------------------------------------------------===//
81 /// This class holds the mapping between values and value numbers. It is used
82 /// as an efficient mechanism to determine the expression-wise equivalence of
88 SmallVector<uint32_t, 4> varargs;
90 Expression(uint32_t o = ~2U) : opcode(o) { }
92 bool operator==(const Expression &other) const {
93 if (opcode != other.opcode)
95 if (opcode == ~0U || opcode == ~1U)
97 if (type != other.type)
99 if (varargs != other.varargs)
104 friend hash_code hash_value(const Expression &Value) {
105 return hash_combine(Value.opcode, Value.type,
106 hash_combine_range(Value.varargs.begin(),
107 Value.varargs.end()));
112 DenseMap<Value*, uint32_t> valueNumbering;
113 DenseMap<Expression, uint32_t> expressionNumbering;
115 MemoryDependenceAnalysis *MD;
118 uint32_t nextValueNumber;
120 Expression create_expression(Instruction* I);
121 Expression create_cmp_expression(unsigned Opcode,
122 CmpInst::Predicate Predicate,
123 Value *LHS, Value *RHS);
124 Expression create_extractvalue_expression(ExtractValueInst* EI);
125 uint32_t lookup_or_add_call(CallInst* C);
127 ValueTable() : nextValueNumber(1) { }
128 uint32_t lookup_or_add(Value *V);
129 uint32_t lookup(Value *V) const;
130 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
131 Value *LHS, Value *RHS);
132 void add(Value *V, uint32_t num);
134 void erase(Value *v);
135 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
136 AliasAnalysis *getAliasAnalysis() const { return AA; }
137 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
138 void setDomTree(DominatorTree* D) { DT = D; }
139 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
140 void verifyRemoved(const Value *) const;
145 template <> struct DenseMapInfo<Expression> {
146 static inline Expression getEmptyKey() {
150 static inline Expression getTombstoneKey() {
154 static unsigned getHashValue(const Expression e) {
155 using llvm::hash_value;
156 return static_cast<unsigned>(hash_value(e));
158 static bool isEqual(const Expression &LHS, const Expression &RHS) {
165 //===----------------------------------------------------------------------===//
166 // ValueTable Internal Functions
167 //===----------------------------------------------------------------------===//
169 Expression ValueTable::create_expression(Instruction *I) {
171 e.type = I->getType();
172 e.opcode = I->getOpcode();
173 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
175 e.varargs.push_back(lookup_or_add(*OI));
176 if (I->isCommutative()) {
177 // Ensure that commutative instructions that only differ by a permutation
178 // of their operands get the same value number by sorting the operand value
179 // numbers. Since all commutative instructions have two operands it is more
180 // efficient to sort by hand rather than using, say, std::sort.
181 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
182 if (e.varargs[0] > e.varargs[1])
183 std::swap(e.varargs[0], e.varargs[1]);
186 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
187 // Sort the operand value numbers so x<y and y>x get the same value number.
188 CmpInst::Predicate Predicate = C->getPredicate();
189 if (e.varargs[0] > e.varargs[1]) {
190 std::swap(e.varargs[0], e.varargs[1]);
191 Predicate = CmpInst::getSwappedPredicate(Predicate);
193 e.opcode = (C->getOpcode() << 8) | Predicate;
194 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
195 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
197 e.varargs.push_back(*II);
203 Expression ValueTable::create_cmp_expression(unsigned Opcode,
204 CmpInst::Predicate Predicate,
205 Value *LHS, Value *RHS) {
206 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
207 "Not a comparison!");
209 e.type = CmpInst::makeCmpResultType(LHS->getType());
210 e.varargs.push_back(lookup_or_add(LHS));
211 e.varargs.push_back(lookup_or_add(RHS));
213 // Sort the operand value numbers so x<y and y>x get the same value number.
214 if (e.varargs[0] > e.varargs[1]) {
215 std::swap(e.varargs[0], e.varargs[1]);
216 Predicate = CmpInst::getSwappedPredicate(Predicate);
218 e.opcode = (Opcode << 8) | Predicate;
222 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
223 assert(EI && "Not an ExtractValueInst?");
225 e.type = EI->getType();
228 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
229 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
230 // EI might be an extract from one of our recognised intrinsics. If it
231 // is we'll synthesize a semantically equivalent expression instead on
232 // an extract value expression.
233 switch (I->getIntrinsicID()) {
234 case Intrinsic::sadd_with_overflow:
235 case Intrinsic::uadd_with_overflow:
236 e.opcode = Instruction::Add;
238 case Intrinsic::ssub_with_overflow:
239 case Intrinsic::usub_with_overflow:
240 e.opcode = Instruction::Sub;
242 case Intrinsic::smul_with_overflow:
243 case Intrinsic::umul_with_overflow:
244 e.opcode = Instruction::Mul;
251 // Intrinsic recognized. Grab its args to finish building the expression.
252 assert(I->getNumArgOperands() == 2 &&
253 "Expect two args for recognised intrinsics.");
254 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
255 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
260 // Not a recognised intrinsic. Fall back to producing an extract value
262 e.opcode = EI->getOpcode();
263 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
265 e.varargs.push_back(lookup_or_add(*OI));
267 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
269 e.varargs.push_back(*II);
274 //===----------------------------------------------------------------------===//
275 // ValueTable External Functions
276 //===----------------------------------------------------------------------===//
278 /// add - Insert a value into the table with a specified value number.
279 void ValueTable::add(Value *V, uint32_t num) {
280 valueNumbering.insert(std::make_pair(V, num));
283 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
284 if (AA->doesNotAccessMemory(C)) {
285 Expression exp = create_expression(C);
286 uint32_t &e = expressionNumbering[exp];
287 if (!e) e = nextValueNumber++;
288 valueNumbering[C] = e;
290 } else if (AA->onlyReadsMemory(C)) {
291 Expression exp = create_expression(C);
292 uint32_t &e = expressionNumbering[exp];
294 e = nextValueNumber++;
295 valueNumbering[C] = e;
299 e = nextValueNumber++;
300 valueNumbering[C] = e;
304 MemDepResult local_dep = MD->getDependency(C);
306 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
307 valueNumbering[C] = nextValueNumber;
308 return nextValueNumber++;
311 if (local_dep.isDef()) {
312 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
314 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
315 valueNumbering[C] = nextValueNumber;
316 return nextValueNumber++;
319 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
320 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
321 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
323 valueNumbering[C] = nextValueNumber;
324 return nextValueNumber++;
328 uint32_t v = lookup_or_add(local_cdep);
329 valueNumbering[C] = v;
334 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
335 MD->getNonLocalCallDependency(CallSite(C));
336 // FIXME: Move the checking logic to MemDep!
337 CallInst* cdep = nullptr;
339 // Check to see if we have a single dominating call instruction that is
341 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
342 const NonLocalDepEntry *I = &deps[i];
343 if (I->getResult().isNonLocal())
346 // We don't handle non-definitions. If we already have a call, reject
347 // instruction dependencies.
348 if (!I->getResult().isDef() || cdep != nullptr) {
353 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
354 // FIXME: All duplicated with non-local case.
355 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
356 cdep = NonLocalDepCall;
365 valueNumbering[C] = nextValueNumber;
366 return nextValueNumber++;
369 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
370 valueNumbering[C] = nextValueNumber;
371 return nextValueNumber++;
373 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
374 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
375 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
377 valueNumbering[C] = nextValueNumber;
378 return nextValueNumber++;
382 uint32_t v = lookup_or_add(cdep);
383 valueNumbering[C] = v;
387 valueNumbering[C] = nextValueNumber;
388 return nextValueNumber++;
392 /// lookup_or_add - Returns the value number for the specified value, assigning
393 /// it a new number if it did not have one before.
394 uint32_t ValueTable::lookup_or_add(Value *V) {
395 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
396 if (VI != valueNumbering.end())
399 if (!isa<Instruction>(V)) {
400 valueNumbering[V] = nextValueNumber;
401 return nextValueNumber++;
404 Instruction* I = cast<Instruction>(V);
406 switch (I->getOpcode()) {
407 case Instruction::Call:
408 return lookup_or_add_call(cast<CallInst>(I));
409 case Instruction::Add:
410 case Instruction::FAdd:
411 case Instruction::Sub:
412 case Instruction::FSub:
413 case Instruction::Mul:
414 case Instruction::FMul:
415 case Instruction::UDiv:
416 case Instruction::SDiv:
417 case Instruction::FDiv:
418 case Instruction::URem:
419 case Instruction::SRem:
420 case Instruction::FRem:
421 case Instruction::Shl:
422 case Instruction::LShr:
423 case Instruction::AShr:
424 case Instruction::And:
425 case Instruction::Or:
426 case Instruction::Xor:
427 case Instruction::ICmp:
428 case Instruction::FCmp:
429 case Instruction::Trunc:
430 case Instruction::ZExt:
431 case Instruction::SExt:
432 case Instruction::FPToUI:
433 case Instruction::FPToSI:
434 case Instruction::UIToFP:
435 case Instruction::SIToFP:
436 case Instruction::FPTrunc:
437 case Instruction::FPExt:
438 case Instruction::PtrToInt:
439 case Instruction::IntToPtr:
440 case Instruction::BitCast:
441 case Instruction::Select:
442 case Instruction::ExtractElement:
443 case Instruction::InsertElement:
444 case Instruction::ShuffleVector:
445 case Instruction::InsertValue:
446 case Instruction::GetElementPtr:
447 exp = create_expression(I);
449 case Instruction::ExtractValue:
450 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
453 valueNumbering[V] = nextValueNumber;
454 return nextValueNumber++;
457 uint32_t& e = expressionNumbering[exp];
458 if (!e) e = nextValueNumber++;
459 valueNumbering[V] = e;
463 /// Returns the value number of the specified value. Fails if
464 /// the value has not yet been numbered.
465 uint32_t ValueTable::lookup(Value *V) const {
466 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
467 assert(VI != valueNumbering.end() && "Value not numbered?");
471 /// Returns the value number of the given comparison,
472 /// assigning it a new number if it did not have one before. Useful when
473 /// we deduced the result of a comparison, but don't immediately have an
474 /// instruction realizing that comparison to hand.
475 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
476 CmpInst::Predicate Predicate,
477 Value *LHS, Value *RHS) {
478 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
479 uint32_t& e = expressionNumbering[exp];
480 if (!e) e = nextValueNumber++;
484 /// Remove all entries from the ValueTable.
485 void ValueTable::clear() {
486 valueNumbering.clear();
487 expressionNumbering.clear();
491 /// Remove a value from the value numbering.
492 void ValueTable::erase(Value *V) {
493 valueNumbering.erase(V);
496 /// verifyRemoved - Verify that the value is removed from all internal data
498 void ValueTable::verifyRemoved(const Value *V) const {
499 for (DenseMap<Value*, uint32_t>::const_iterator
500 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
501 assert(I->first != V && "Inst still occurs in value numbering map!");
505 //===----------------------------------------------------------------------===//
507 //===----------------------------------------------------------------------===//
511 struct AvailableValueInBlock {
512 /// BB - The basic block in question.
515 SimpleVal, // A simple offsetted value that is accessed.
516 LoadVal, // A value produced by a load.
517 MemIntrin, // A memory intrinsic which is loaded from.
518 UndefVal // A UndefValue representing a value from dead block (which
519 // is not yet physically removed from the CFG).
522 /// V - The value that is live out of the block.
523 PointerIntPair<Value *, 2, ValType> Val;
525 /// Offset - The byte offset in Val that is interesting for the load query.
528 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
529 unsigned Offset = 0) {
530 AvailableValueInBlock Res;
532 Res.Val.setPointer(V);
533 Res.Val.setInt(SimpleVal);
538 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
539 unsigned Offset = 0) {
540 AvailableValueInBlock Res;
542 Res.Val.setPointer(MI);
543 Res.Val.setInt(MemIntrin);
548 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
549 unsigned Offset = 0) {
550 AvailableValueInBlock Res;
552 Res.Val.setPointer(LI);
553 Res.Val.setInt(LoadVal);
558 static AvailableValueInBlock getUndef(BasicBlock *BB) {
559 AvailableValueInBlock Res;
561 Res.Val.setPointer(nullptr);
562 Res.Val.setInt(UndefVal);
567 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
568 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
569 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
570 bool isUndefValue() const { return Val.getInt() == UndefVal; }
572 Value *getSimpleValue() const {
573 assert(isSimpleValue() && "Wrong accessor");
574 return Val.getPointer();
577 LoadInst *getCoercedLoadValue() const {
578 assert(isCoercedLoadValue() && "Wrong accessor");
579 return cast<LoadInst>(Val.getPointer());
582 MemIntrinsic *getMemIntrinValue() const {
583 assert(isMemIntrinValue() && "Wrong accessor");
584 return cast<MemIntrinsic>(Val.getPointer());
587 /// Emit code into this block to adjust the value defined here to the
588 /// specified type. This handles various coercion cases.
589 Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const;
592 class GVN : public FunctionPass {
594 MemoryDependenceAnalysis *MD;
596 const TargetLibraryInfo *TLI;
598 SetVector<BasicBlock *> DeadBlocks;
602 /// A mapping from value numbers to lists of Value*'s that
603 /// have that value number. Use findLeader to query it.
604 struct LeaderTableEntry {
606 const BasicBlock *BB;
607 LeaderTableEntry *Next;
609 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
610 BumpPtrAllocator TableAllocator;
612 // Block-local map of equivalent values to their leader, does not
613 // propagate to any successors. Entries added mid-block are applied
614 // to the remaining instructions in the block.
615 SmallMapVector<llvm::Value *, llvm::Constant *, 4> ReplaceWithConstMap;
616 SmallVector<Instruction*, 8> InstrsToErase;
618 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
619 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
620 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
623 static char ID; // Pass identification, replacement for typeid
624 explicit GVN(bool noloads = false)
625 : FunctionPass(ID), NoLoads(noloads), MD(nullptr) {
626 initializeGVNPass(*PassRegistry::getPassRegistry());
629 bool runOnFunction(Function &F) override;
631 /// This removes the specified instruction from
632 /// our various maps and marks it for deletion.
633 void markInstructionForDeletion(Instruction *I) {
635 InstrsToErase.push_back(I);
638 DominatorTree &getDominatorTree() const { return *DT; }
639 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
640 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
642 /// Push a new Value to the LeaderTable onto the list for its value number.
643 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
644 LeaderTableEntry &Curr = LeaderTable[N];
651 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
654 Node->Next = Curr.Next;
658 /// Scan the list of values corresponding to a given
659 /// value number, and remove the given instruction if encountered.
660 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
661 LeaderTableEntry* Prev = nullptr;
662 LeaderTableEntry* Curr = &LeaderTable[N];
664 while (Curr && (Curr->Val != I || Curr->BB != BB)) {
673 Prev->Next = Curr->Next;
679 LeaderTableEntry* Next = Curr->Next;
680 Curr->Val = Next->Val;
682 Curr->Next = Next->Next;
687 // List of critical edges to be split between iterations.
688 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
690 // This transformation requires dominator postdominator info
691 void getAnalysisUsage(AnalysisUsage &AU) const override {
692 AU.addRequired<AssumptionCacheTracker>();
693 AU.addRequired<DominatorTreeWrapperPass>();
694 AU.addRequired<TargetLibraryInfoWrapperPass>();
696 AU.addRequired<MemoryDependenceAnalysis>();
697 AU.addRequired<AAResultsWrapperPass>();
699 AU.addPreserved<DominatorTreeWrapperPass>();
700 AU.addPreserved<GlobalsAAWrapperPass>();
704 // Helper functions of redundant load elimination
705 bool processLoad(LoadInst *L);
706 bool processNonLocalLoad(LoadInst *L);
707 bool processAssumeIntrinsic(IntrinsicInst *II);
708 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
709 AvailValInBlkVect &ValuesPerBlock,
710 UnavailBlkVect &UnavailableBlocks);
711 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
712 UnavailBlkVect &UnavailableBlocks);
714 // Other helper routines
715 bool processInstruction(Instruction *I);
716 bool processBlock(BasicBlock *BB);
717 void dump(DenseMap<uint32_t, Value*> &d);
718 bool iterateOnFunction(Function &F);
719 bool performPRE(Function &F);
720 bool performScalarPRE(Instruction *I);
721 bool performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
723 Value *findLeader(const BasicBlock *BB, uint32_t num);
724 void cleanupGlobalSets();
725 void verifyRemoved(const Instruction *I) const;
726 bool splitCriticalEdges();
727 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
728 bool replaceOperandsWithConsts(Instruction *I) const;
729 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
730 bool DominatesByEdge);
731 bool processFoldableCondBr(BranchInst *BI);
732 void addDeadBlock(BasicBlock *BB);
733 void assignValNumForDeadCode();
739 // The public interface to this file...
740 FunctionPass *llvm::createGVNPass(bool NoLoads) {
741 return new GVN(NoLoads);
744 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
745 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
746 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
747 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
748 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
749 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
750 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
751 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
753 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
754 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
756 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
757 E = d.end(); I != E; ++I) {
758 errs() << I->first << "\n";
765 /// Return true if we can prove that the value
766 /// we're analyzing is fully available in the specified block. As we go, keep
767 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
768 /// map is actually a tri-state map with the following values:
769 /// 0) we know the block *is not* fully available.
770 /// 1) we know the block *is* fully available.
771 /// 2) we do not know whether the block is fully available or not, but we are
772 /// currently speculating that it will be.
773 /// 3) we are speculating for this block and have used that to speculate for
775 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
776 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
777 uint32_t RecurseDepth) {
778 if (RecurseDepth > MaxRecurseDepth)
781 // Optimistically assume that the block is fully available and check to see
782 // if we already know about this block in one lookup.
783 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
784 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
786 // If the entry already existed for this block, return the precomputed value.
788 // If this is a speculative "available" value, mark it as being used for
789 // speculation of other blocks.
790 if (IV.first->second == 2)
791 IV.first->second = 3;
792 return IV.first->second != 0;
795 // Otherwise, see if it is fully available in all predecessors.
796 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
798 // If this block has no predecessors, it isn't live-in here.
800 goto SpeculationFailure;
802 for (; PI != PE; ++PI)
803 // If the value isn't fully available in one of our predecessors, then it
804 // isn't fully available in this block either. Undo our previous
805 // optimistic assumption and bail out.
806 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
807 goto SpeculationFailure;
811 // If we get here, we found out that this is not, after
812 // all, a fully-available block. We have a problem if we speculated on this and
813 // used the speculation to mark other blocks as available.
815 char &BBVal = FullyAvailableBlocks[BB];
817 // If we didn't speculate on this, just return with it set to false.
823 // If we did speculate on this value, we could have blocks set to 1 that are
824 // incorrect. Walk the (transitive) successors of this block and mark them as
826 SmallVector<BasicBlock*, 32> BBWorklist;
827 BBWorklist.push_back(BB);
830 BasicBlock *Entry = BBWorklist.pop_back_val();
831 // Note that this sets blocks to 0 (unavailable) if they happen to not
832 // already be in FullyAvailableBlocks. This is safe.
833 char &EntryVal = FullyAvailableBlocks[Entry];
834 if (EntryVal == 0) continue; // Already unavailable.
836 // Mark as unavailable.
839 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
840 } while (!BBWorklist.empty());
846 /// Return true if CoerceAvailableValueToLoadType will succeed.
847 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
849 const DataLayout &DL) {
850 // If the loaded or stored value is an first class array or struct, don't try
851 // to transform them. We need to be able to bitcast to integer.
852 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
853 StoredVal->getType()->isStructTy() ||
854 StoredVal->getType()->isArrayTy())
857 // The store has to be at least as big as the load.
858 if (DL.getTypeSizeInBits(StoredVal->getType()) <
859 DL.getTypeSizeInBits(LoadTy))
865 /// If we saw a store of a value to memory, and
866 /// then a load from a must-aliased pointer of a different type, try to coerce
867 /// the stored value. LoadedTy is the type of the load we want to replace.
868 /// IRB is IRBuilder used to insert new instructions.
870 /// If we can't do it, return null.
871 static Value *CoerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy,
873 const DataLayout &DL) {
874 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
877 // If this is already the right type, just return it.
878 Type *StoredValTy = StoredVal->getType();
880 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
881 uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
883 // If the store and reload are the same size, we can always reuse it.
884 if (StoreSize == LoadSize) {
885 // Pointer to Pointer -> use bitcast.
886 if (StoredValTy->getScalarType()->isPointerTy() &&
887 LoadedTy->getScalarType()->isPointerTy())
888 return IRB.CreateBitCast(StoredVal, LoadedTy);
890 // Convert source pointers to integers, which can be bitcast.
891 if (StoredValTy->getScalarType()->isPointerTy()) {
892 StoredValTy = DL.getIntPtrType(StoredValTy);
893 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
896 Type *TypeToCastTo = LoadedTy;
897 if (TypeToCastTo->getScalarType()->isPointerTy())
898 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
900 if (StoredValTy != TypeToCastTo)
901 StoredVal = IRB.CreateBitCast(StoredVal, TypeToCastTo);
903 // Cast to pointer if the load needs a pointer type.
904 if (LoadedTy->getScalarType()->isPointerTy())
905 StoredVal = IRB.CreateIntToPtr(StoredVal, LoadedTy);
910 // If the loaded value is smaller than the available value, then we can
911 // extract out a piece from it. If the available value is too small, then we
912 // can't do anything.
913 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
915 // Convert source pointers to integers, which can be manipulated.
916 if (StoredValTy->getScalarType()->isPointerTy()) {
917 StoredValTy = DL.getIntPtrType(StoredValTy);
918 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
921 // Convert vectors and fp to integer, which can be manipulated.
922 if (!StoredValTy->isIntegerTy()) {
923 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
924 StoredVal = IRB.CreateBitCast(StoredVal, StoredValTy);
927 // If this is a big-endian system, we need to shift the value down to the low
928 // bits so that a truncate will work.
929 if (DL.isBigEndian()) {
930 StoredVal = IRB.CreateLShr(StoredVal, StoreSize - LoadSize, "tmp");
933 // Truncate the integer to the right size now.
934 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
935 StoredVal = IRB.CreateTrunc(StoredVal, NewIntTy, "trunc");
937 if (LoadedTy == NewIntTy)
940 // If the result is a pointer, inttoptr.
941 if (LoadedTy->getScalarType()->isPointerTy())
942 return IRB.CreateIntToPtr(StoredVal, LoadedTy, "inttoptr");
944 // Otherwise, bitcast.
945 return IRB.CreateBitCast(StoredVal, LoadedTy, "bitcast");
948 /// This function is called when we have a
949 /// memdep query of a load that ends up being a clobbering memory write (store,
950 /// memset, memcpy, memmove). This means that the write *may* provide bits used
951 /// by the load but we can't be sure because the pointers don't mustalias.
953 /// Check this case to see if there is anything more we can do before we give
954 /// up. This returns -1 if we have to give up, or a byte number in the stored
955 /// value of the piece that feeds the load.
956 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
958 uint64_t WriteSizeInBits,
959 const DataLayout &DL) {
960 // If the loaded or stored value is a first class array or struct, don't try
961 // to transform them. We need to be able to bitcast to integer.
962 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
965 int64_t StoreOffset = 0, LoadOffset = 0;
967 GetPointerBaseWithConstantOffset(WritePtr, StoreOffset, DL);
968 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, DL);
969 if (StoreBase != LoadBase)
972 // If the load and store are to the exact same address, they should have been
973 // a must alias. AA must have gotten confused.
974 // FIXME: Study to see if/when this happens. One case is forwarding a memset
975 // to a load from the base of the memset.
977 if (LoadOffset == StoreOffset) {
978 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
979 << "Base = " << *StoreBase << "\n"
980 << "Store Ptr = " << *WritePtr << "\n"
981 << "Store Offs = " << StoreOffset << "\n"
982 << "Load Ptr = " << *LoadPtr << "\n";
987 // If the load and store don't overlap at all, the store doesn't provide
988 // anything to the load. In this case, they really don't alias at all, AA
989 // must have gotten confused.
990 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
992 if ((WriteSizeInBits & 7) | (LoadSize & 7))
994 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
998 bool isAAFailure = false;
999 if (StoreOffset < LoadOffset)
1000 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
1002 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1006 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1007 << "Base = " << *StoreBase << "\n"
1008 << "Store Ptr = " << *WritePtr << "\n"
1009 << "Store Offs = " << StoreOffset << "\n"
1010 << "Load Ptr = " << *LoadPtr << "\n";
1016 // If the Load isn't completely contained within the stored bits, we don't
1017 // have all the bits to feed it. We could do something crazy in the future
1018 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1020 if (StoreOffset > LoadOffset ||
1021 StoreOffset+StoreSize < LoadOffset+LoadSize)
1024 // Okay, we can do this transformation. Return the number of bytes into the
1025 // store that the load is.
1026 return LoadOffset-StoreOffset;
1029 /// This function is called when we have a
1030 /// memdep query of a load that ends up being a clobbering store.
1031 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1033 // Cannot handle reading from store of first-class aggregate yet.
1034 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1035 DepSI->getValueOperand()->getType()->isArrayTy())
1038 const DataLayout &DL = DepSI->getModule()->getDataLayout();
1039 Value *StorePtr = DepSI->getPointerOperand();
1040 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1041 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1042 StorePtr, StoreSize, DL);
1045 /// This function is called when we have a
1046 /// memdep query of a load that ends up being clobbered by another load. See if
1047 /// the other load can feed into the second load.
1048 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1049 LoadInst *DepLI, const DataLayout &DL){
1050 // Cannot handle reading from store of first-class aggregate yet.
1051 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1054 Value *DepPtr = DepLI->getPointerOperand();
1055 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
1056 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
1057 if (R != -1) return R;
1059 // If we have a load/load clobber an DepLI can be widened to cover this load,
1060 // then we should widen it!
1061 int64_t LoadOffs = 0;
1062 const Value *LoadBase =
1063 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, DL);
1064 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1066 unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
1067 LoadBase, LoadOffs, LoadSize, DepLI);
1068 if (Size == 0) return -1;
1070 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
1075 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1077 const DataLayout &DL) {
1078 // If the mem operation is a non-constant size, we can't handle it.
1079 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1080 if (!SizeCst) return -1;
1081 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1083 // If this is memset, we just need to see if the offset is valid in the size
1085 if (MI->getIntrinsicID() == Intrinsic::memset)
1086 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1089 // If we have a memcpy/memmove, the only case we can handle is if this is a
1090 // copy from constant memory. In that case, we can read directly from the
1092 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1094 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1095 if (!Src) return -1;
1097 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, DL));
1098 if (!GV || !GV->isConstant()) return -1;
1100 // See if the access is within the bounds of the transfer.
1101 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1102 MI->getDest(), MemSizeInBits, DL);
1106 unsigned AS = Src->getType()->getPointerAddressSpace();
1107 // Otherwise, see if we can constant fold a load from the constant with the
1108 // offset applied as appropriate.
1109 Src = ConstantExpr::getBitCast(Src,
1110 Type::getInt8PtrTy(Src->getContext(), AS));
1111 Constant *OffsetCst =
1112 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1113 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1115 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1116 if (ConstantFoldLoadFromConstPtr(Src, DL))
1122 /// This function is called when we have a
1123 /// memdep query of a load that ends up being a clobbering store. This means
1124 /// that the store provides bits used by the load but we the pointers don't
1125 /// mustalias. Check this case to see if there is anything more we can do
1126 /// before we give up.
1127 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1129 Instruction *InsertPt, const DataLayout &DL){
1130 LLVMContext &Ctx = SrcVal->getType()->getContext();
1132 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1133 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
1135 IRBuilder<> Builder(InsertPt);
1137 // Compute which bits of the stored value are being used by the load. Convert
1138 // to an integer type to start with.
1139 if (SrcVal->getType()->getScalarType()->isPointerTy())
1140 SrcVal = Builder.CreatePtrToInt(SrcVal,
1141 DL.getIntPtrType(SrcVal->getType()));
1142 if (!SrcVal->getType()->isIntegerTy())
1143 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1145 // Shift the bits to the least significant depending on endianness.
1147 if (DL.isLittleEndian())
1148 ShiftAmt = Offset*8;
1150 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1153 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1155 if (LoadSize != StoreSize)
1156 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1158 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, Builder, DL);
1161 /// This function is called when we have a
1162 /// memdep query of a load that ends up being a clobbering load. This means
1163 /// that the load *may* provide bits used by the load but we can't be sure
1164 /// because the pointers don't mustalias. Check this case to see if there is
1165 /// anything more we can do before we give up.
1166 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1167 Type *LoadTy, Instruction *InsertPt,
1169 const DataLayout &DL = SrcVal->getModule()->getDataLayout();
1170 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1171 // widen SrcVal out to a larger load.
1172 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
1173 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1174 if (Offset+LoadSize > SrcValSize) {
1175 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1176 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1177 // If we have a load/load clobber an DepLI can be widened to cover this
1178 // load, then we should widen it to the next power of 2 size big enough!
1179 unsigned NewLoadSize = Offset+LoadSize;
1180 if (!isPowerOf2_32(NewLoadSize))
1181 NewLoadSize = NextPowerOf2(NewLoadSize);
1183 Value *PtrVal = SrcVal->getPointerOperand();
1185 // Insert the new load after the old load. This ensures that subsequent
1186 // memdep queries will find the new load. We can't easily remove the old
1187 // load completely because it is already in the value numbering table.
1188 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1190 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1191 DestPTy = PointerType::get(DestPTy,
1192 PtrVal->getType()->getPointerAddressSpace());
1193 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1194 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1195 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1196 NewLoad->takeName(SrcVal);
1197 NewLoad->setAlignment(SrcVal->getAlignment());
1199 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1200 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1202 // Replace uses of the original load with the wider load. On a big endian
1203 // system, we need to shift down to get the relevant bits.
1204 Value *RV = NewLoad;
1205 if (DL.isBigEndian())
1206 RV = Builder.CreateLShr(RV,
1207 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1208 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1209 SrcVal->replaceAllUsesWith(RV);
1211 // We would like to use gvn.markInstructionForDeletion here, but we can't
1212 // because the load is already memoized into the leader map table that GVN
1213 // tracks. It is potentially possible to remove the load from the table,
1214 // but then there all of the operations based on it would need to be
1215 // rehashed. Just leave the dead load around.
1216 gvn.getMemDep().removeInstruction(SrcVal);
1220 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1224 /// This function is called when we have a
1225 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1226 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1227 Type *LoadTy, Instruction *InsertPt,
1228 const DataLayout &DL){
1229 LLVMContext &Ctx = LoadTy->getContext();
1230 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1232 IRBuilder<> Builder(InsertPt);
1234 // We know that this method is only called when the mem transfer fully
1235 // provides the bits for the load.
1236 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1237 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1238 // independently of what the offset is.
1239 Value *Val = MSI->getValue();
1241 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1243 Value *OneElt = Val;
1245 // Splat the value out to the right number of bits.
1246 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1247 // If we can double the number of bytes set, do it.
1248 if (NumBytesSet*2 <= LoadSize) {
1249 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1250 Val = Builder.CreateOr(Val, ShVal);
1255 // Otherwise insert one byte at a time.
1256 Value *ShVal = Builder.CreateShl(Val, 1*8);
1257 Val = Builder.CreateOr(OneElt, ShVal);
1261 return CoerceAvailableValueToLoadType(Val, LoadTy, Builder, DL);
1264 // Otherwise, this is a memcpy/memmove from a constant global.
1265 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1266 Constant *Src = cast<Constant>(MTI->getSource());
1267 unsigned AS = Src->getType()->getPointerAddressSpace();
1269 // Otherwise, see if we can constant fold a load from the constant with the
1270 // offset applied as appropriate.
1271 Src = ConstantExpr::getBitCast(Src,
1272 Type::getInt8PtrTy(Src->getContext(), AS));
1273 Constant *OffsetCst =
1274 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1275 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1277 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1278 return ConstantFoldLoadFromConstPtr(Src, DL);
1282 /// Given a set of loads specified by ValuesPerBlock,
1283 /// construct SSA form, allowing us to eliminate LI. This returns the value
1284 /// that should be used at LI's definition site.
1285 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1286 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1288 // Check for the fully redundant, dominating load case. In this case, we can
1289 // just use the dominating value directly.
1290 if (ValuesPerBlock.size() == 1 &&
1291 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1293 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1294 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
1297 // Otherwise, we have to construct SSA form.
1298 SmallVector<PHINode*, 8> NewPHIs;
1299 SSAUpdater SSAUpdate(&NewPHIs);
1300 SSAUpdate.Initialize(LI->getType(), LI->getName());
1302 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1303 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1304 BasicBlock *BB = AV.BB;
1306 if (SSAUpdate.HasValueForBlock(BB))
1309 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
1312 // Perform PHI construction.
1313 return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1316 Value *AvailableValueInBlock::MaterializeAdjustedValue(LoadInst *LI,
1319 Type *LoadTy = LI->getType();
1320 const DataLayout &DL = LI->getModule()->getDataLayout();
1321 if (isSimpleValue()) {
1322 Res = getSimpleValue();
1323 if (Res->getType() != LoadTy) {
1324 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), DL);
1326 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1327 << *getSimpleValue() << '\n'
1328 << *Res << '\n' << "\n\n\n");
1330 } else if (isCoercedLoadValue()) {
1331 LoadInst *Load = getCoercedLoadValue();
1332 if (Load->getType() == LoadTy && Offset == 0) {
1335 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1338 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1339 << *getCoercedLoadValue() << '\n'
1340 << *Res << '\n' << "\n\n\n");
1342 } else if (isMemIntrinValue()) {
1343 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
1344 BB->getTerminator(), DL);
1345 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1346 << " " << *getMemIntrinValue() << '\n'
1347 << *Res << '\n' << "\n\n\n");
1349 assert(isUndefValue() && "Should be UndefVal");
1350 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1351 return UndefValue::get(LoadTy);
1356 static bool isLifetimeStart(const Instruction *Inst) {
1357 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1358 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1362 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1363 AvailValInBlkVect &ValuesPerBlock,
1364 UnavailBlkVect &UnavailableBlocks) {
1366 // Filter out useless results (non-locals, etc). Keep track of the blocks
1367 // where we have a value available in repl, also keep track of whether we see
1368 // dependencies that produce an unknown value for the load (such as a call
1369 // that could potentially clobber the load).
1370 unsigned NumDeps = Deps.size();
1371 const DataLayout &DL = LI->getModule()->getDataLayout();
1372 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1373 BasicBlock *DepBB = Deps[i].getBB();
1374 MemDepResult DepInfo = Deps[i].getResult();
1376 if (DeadBlocks.count(DepBB)) {
1377 // Dead dependent mem-op disguise as a load evaluating the same value
1378 // as the load in question.
1379 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1383 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1384 UnavailableBlocks.push_back(DepBB);
1388 if (DepInfo.isClobber()) {
1389 // The address being loaded in this non-local block may not be the same as
1390 // the pointer operand of the load if PHI translation occurs. Make sure
1391 // to consider the right address.
1392 Value *Address = Deps[i].getAddress();
1394 // If the dependence is to a store that writes to a superset of the bits
1395 // read by the load, we can extract the bits we need for the load from the
1397 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1400 AnalyzeLoadFromClobberingStore(LI->getType(), Address, DepSI);
1402 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1403 DepSI->getValueOperand(),
1410 // Check to see if we have something like this:
1413 // if we have this, replace the later with an extraction from the former.
1414 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1415 // If this is a clobber and L is the first instruction in its block, then
1416 // we have the first instruction in the entry block.
1417 if (DepLI != LI && Address) {
1419 AnalyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
1422 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1429 // If the clobbering value is a memset/memcpy/memmove, see if we can
1430 // forward a value on from it.
1431 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1433 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1436 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1443 UnavailableBlocks.push_back(DepBB);
1447 // DepInfo.isDef() here
1449 Instruction *DepInst = DepInfo.getInst();
1451 // Loading the allocation -> undef.
1452 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1453 // Loading immediately after lifetime begin -> undef.
1454 isLifetimeStart(DepInst)) {
1455 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1456 UndefValue::get(LI->getType())));
1460 // Loading from calloc (which zero initializes memory) -> zero
1461 if (isCallocLikeFn(DepInst, TLI)) {
1462 ValuesPerBlock.push_back(AvailableValueInBlock::get(
1463 DepBB, Constant::getNullValue(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 (!CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1474 LI->getType(), DL)) {
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 (!CanCoerceMustAliasedValueToLoad(LD, LI->getType(), DL)) {
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 MapVector<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)) {
1558 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1559 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1560 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1561 << Pred->getName() << "': " << *LI << '\n');
1565 if (LoadBB->isEHPad()) {
1567 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1568 << Pred->getName() << "': " << *LI << '\n');
1572 CriticalEdgePred.push_back(Pred);
1574 // Only add the predecessors that will not be split for now.
1575 PredLoads[Pred] = nullptr;
1579 // Decide whether PRE is profitable for this load.
1580 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1581 assert(NumUnavailablePreds != 0 &&
1582 "Fully available value should already be eliminated!");
1584 // If this load is unavailable in multiple predecessors, reject it.
1585 // FIXME: If we could restructure the CFG, we could make a common pred with
1586 // all the preds that don't have an available LI and insert a new load into
1588 if (NumUnavailablePreds != 1)
1591 // Split critical edges, and update the unavailable predecessors accordingly.
1592 for (BasicBlock *OrigPred : CriticalEdgePred) {
1593 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1594 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1595 PredLoads[NewPred] = nullptr;
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 const DataLayout &DL = LI->getModule()->getDataLayout();
1603 SmallVector<Instruction*, 8> NewInsts;
1604 for (auto &PredLoad : PredLoads) {
1605 BasicBlock *UnavailablePred = PredLoad.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(), DL, AC);
1614 Value *LoadPtr = nullptr;
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 PredLoad.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 critical 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 (const auto &PredLoad : PredLoads) {
1659 BasicBlock *UnavailablePred = PredLoad.first;
1660 Value *LoadPtr = PredLoad.second;
1662 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1664 UnavailablePred->getTerminator());
1666 // Transfer the old load's AA tags to the new load.
1668 LI->getAAMetadata(Tags);
1670 NewLoad->setAAMetadata(Tags);
1672 // Transfer DebugLoc.
1673 NewLoad->setDebugLoc(LI->getDebugLoc());
1675 // Add the newly created load.
1676 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1678 MD->invalidateCachedPointerInfo(LoadPtr);
1679 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1682 // Perform PHI construction.
1683 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1684 LI->replaceAllUsesWith(V);
1685 if (isa<PHINode>(V))
1687 if (Instruction *I = dyn_cast<Instruction>(V))
1688 I->setDebugLoc(LI->getDebugLoc());
1689 if (V->getType()->getScalarType()->isPointerTy())
1690 MD->invalidateCachedPointerInfo(V);
1691 markInstructionForDeletion(LI);
1696 /// Attempt to eliminate a load whose dependencies are
1697 /// non-local by performing PHI construction.
1698 bool GVN::processNonLocalLoad(LoadInst *LI) {
1699 // Step 1: Find the non-local dependencies of the load.
1701 MD->getNonLocalPointerDependency(LI, Deps);
1703 // If we had to process more than one hundred blocks to find the
1704 // dependencies, this load isn't worth worrying about. Optimizing
1705 // it will be too expensive.
1706 unsigned NumDeps = Deps.size();
1710 // If we had a phi translation failure, we'll have a single entry which is a
1711 // clobber in the current block. Reject this early.
1713 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1715 dbgs() << "GVN: non-local load ";
1716 LI->printAsOperand(dbgs());
1717 dbgs() << " has unknown dependencies\n";
1722 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1723 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1724 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1725 OE = GEP->idx_end();
1727 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1728 performScalarPRE(I);
1731 // Step 2: Analyze the availability of the load
1732 AvailValInBlkVect ValuesPerBlock;
1733 UnavailBlkVect UnavailableBlocks;
1734 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1736 // If we have no predecessors that produce a known value for this load, exit
1738 if (ValuesPerBlock.empty())
1741 // Step 3: Eliminate fully redundancy.
1743 // If all of the instructions we depend on produce a known value for this
1744 // load, then it is fully redundant and we can use PHI insertion to compute
1745 // its value. Insert PHIs and remove the fully redundant value now.
1746 if (UnavailableBlocks.empty()) {
1747 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1749 // Perform PHI construction.
1750 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1751 LI->replaceAllUsesWith(V);
1753 if (isa<PHINode>(V))
1755 if (Instruction *I = dyn_cast<Instruction>(V))
1756 if (LI->getDebugLoc())
1757 I->setDebugLoc(LI->getDebugLoc());
1758 if (V->getType()->getScalarType()->isPointerTy())
1759 MD->invalidateCachedPointerInfo(V);
1760 markInstructionForDeletion(LI);
1765 // Step 4: Eliminate partial redundancy.
1766 if (!EnablePRE || !EnableLoadPRE)
1769 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1772 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1773 assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1774 "This function can only be called with llvm.assume intrinsic");
1775 Value *V = IntrinsicI->getArgOperand(0);
1777 if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1778 if (Cond->isZero()) {
1779 Type *Int8Ty = Type::getInt8Ty(V->getContext());
1780 // Insert a new store to null instruction before the load to indicate that
1781 // this code is not reachable. FIXME: We could insert unreachable
1782 // instruction directly because we can modify the CFG.
1783 new StoreInst(UndefValue::get(Int8Ty),
1784 Constant::getNullValue(Int8Ty->getPointerTo()),
1787 markInstructionForDeletion(IntrinsicI);
1791 Constant *True = ConstantInt::getTrue(V->getContext());
1792 bool Changed = false;
1794 for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1795 BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1797 // This property is only true in dominated successors, propagateEquality
1798 // will check dominance for us.
1799 Changed |= propagateEquality(V, True, Edge, false);
1802 // We can replace assume value with true, which covers cases like this:
1803 // call void @llvm.assume(i1 %cmp)
1804 // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1805 ReplaceWithConstMap[V] = True;
1807 // If one of *cmp *eq operand is const, adding it to map will cover this:
1808 // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1809 // call void @llvm.assume(i1 %cmp)
1810 // ret float %0 ; will change it to ret float 3.000000e+00
1811 if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1812 if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1813 CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1814 (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1815 CmpI->getFastMathFlags().noNaNs())) {
1816 Value *CmpLHS = CmpI->getOperand(0);
1817 Value *CmpRHS = CmpI->getOperand(1);
1818 if (isa<Constant>(CmpLHS))
1819 std::swap(CmpLHS, CmpRHS);
1820 auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1822 // If only one operand is constant.
1823 if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1824 ReplaceWithConstMap[CmpLHS] = RHSConst;
1830 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1831 // Patch the replacement so that it is not more restrictive than the value
1833 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1834 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1836 ReplOp->andIRFlags(Op);
1838 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1839 // FIXME: If both the original and replacement value are part of the
1840 // same control-flow region (meaning that the execution of one
1841 // guarantees the execution of the other), then we can combine the
1842 // noalias scopes here and do better than the general conservative
1843 // answer used in combineMetadata().
1845 // In general, GVN unifies expressions over different control-flow
1846 // regions, and so we need a conservative combination of the noalias
1848 static const unsigned KnownIDs[] = {
1849 LLVMContext::MD_tbaa,
1850 LLVMContext::MD_alias_scope,
1851 LLVMContext::MD_noalias,
1852 LLVMContext::MD_range,
1853 LLVMContext::MD_fpmath,
1854 LLVMContext::MD_invariant_load,
1856 combineMetadata(ReplInst, I, KnownIDs);
1860 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1861 patchReplacementInstruction(I, Repl);
1862 I->replaceAllUsesWith(Repl);
1865 /// Attempt to eliminate a load, first by eliminating it
1866 /// locally, and then attempting non-local elimination if that fails.
1867 bool GVN::processLoad(LoadInst *L) {
1874 if (L->use_empty()) {
1875 markInstructionForDeletion(L);
1879 // ... to a pointer that has been loaded from before...
1880 MemDepResult Dep = MD->getDependency(L);
1881 const DataLayout &DL = L->getModule()->getDataLayout();
1883 // If we have a clobber and target data is around, see if this is a clobber
1884 // that we can fix up through code synthesis.
1885 if (Dep.isClobber()) {
1886 // Check to see if we have something like this:
1887 // store i32 123, i32* %P
1888 // %A = bitcast i32* %P to i8*
1889 // %B = gep i8* %A, i32 1
1892 // We could do that by recognizing if the clobber instructions are obviously
1893 // a common base + constant offset, and if the previous store (or memset)
1894 // completely covers this load. This sort of thing can happen in bitfield
1896 Value *AvailVal = nullptr;
1897 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1898 int Offset = AnalyzeLoadFromClobberingStore(
1899 L->getType(), L->getPointerOperand(), DepSI);
1901 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1902 L->getType(), L, DL);
1905 // Check to see if we have something like this:
1908 // if we have this, replace the later with an extraction from the former.
1909 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1910 // If this is a clobber and L is the first instruction in its block, then
1911 // we have the first instruction in the entry block.
1915 int Offset = AnalyzeLoadFromClobberingLoad(
1916 L->getType(), L->getPointerOperand(), DepLI, DL);
1918 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1921 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1922 // a value on from it.
1923 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1924 int Offset = AnalyzeLoadFromClobberingMemInst(
1925 L->getType(), L->getPointerOperand(), DepMI, DL);
1927 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, DL);
1931 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1932 << *AvailVal << '\n' << *L << "\n\n\n");
1934 // Replace the load!
1935 L->replaceAllUsesWith(AvailVal);
1936 if (AvailVal->getType()->getScalarType()->isPointerTy())
1937 MD->invalidateCachedPointerInfo(AvailVal);
1938 markInstructionForDeletion(L);
1944 // If the value isn't available, don't do anything!
1945 if (Dep.isClobber()) {
1947 // fast print dep, using operator<< on instruction is too slow.
1948 dbgs() << "GVN: load ";
1949 L->printAsOperand(dbgs());
1950 Instruction *I = Dep.getInst();
1951 dbgs() << " is clobbered by " << *I << '\n';
1956 // If it is defined in another block, try harder.
1957 if (Dep.isNonLocal())
1958 return processNonLocalLoad(L);
1962 // fast print dep, using operator<< on instruction is too slow.
1963 dbgs() << "GVN: load ";
1964 L->printAsOperand(dbgs());
1965 dbgs() << " has unknown dependence\n";
1970 Instruction *DepInst = Dep.getInst();
1971 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1972 Value *StoredVal = DepSI->getValueOperand();
1974 // The store and load are to a must-aliased pointer, but they may not
1975 // actually have the same type. See if we know how to reuse the stored
1976 // value (depending on its type).
1977 if (StoredVal->getType() != L->getType()) {
1978 IRBuilder<> Builder(L);
1980 CoerceAvailableValueToLoadType(StoredVal, L->getType(), Builder, DL);
1984 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1985 << '\n' << *L << "\n\n\n");
1989 L->replaceAllUsesWith(StoredVal);
1990 if (StoredVal->getType()->getScalarType()->isPointerTy())
1991 MD->invalidateCachedPointerInfo(StoredVal);
1992 markInstructionForDeletion(L);
1997 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1998 Value *AvailableVal = DepLI;
2000 // The loads are of a must-aliased pointer, but they may not actually have
2001 // the same type. See if we know how to reuse the previously loaded value
2002 // (depending on its type).
2003 if (DepLI->getType() != L->getType()) {
2004 IRBuilder<> Builder(L);
2006 CoerceAvailableValueToLoadType(DepLI, L->getType(), Builder, DL);
2010 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
2011 << "\n" << *L << "\n\n\n");
2015 patchAndReplaceAllUsesWith(L, AvailableVal);
2016 if (DepLI->getType()->getScalarType()->isPointerTy())
2017 MD->invalidateCachedPointerInfo(DepLI);
2018 markInstructionForDeletion(L);
2023 // If this load really doesn't depend on anything, then we must be loading an
2024 // undef value. This can happen when loading for a fresh allocation with no
2025 // intervening stores, for example.
2026 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
2027 L->replaceAllUsesWith(UndefValue::get(L->getType()));
2028 markInstructionForDeletion(L);
2033 // If this load occurs either right after a lifetime begin,
2034 // then the loaded value is undefined.
2035 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
2036 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
2037 L->replaceAllUsesWith(UndefValue::get(L->getType()));
2038 markInstructionForDeletion(L);
2044 // If this load follows a calloc (which zero initializes memory),
2045 // then the loaded value is zero
2046 if (isCallocLikeFn(DepInst, TLI)) {
2047 L->replaceAllUsesWith(Constant::getNullValue(L->getType()));
2048 markInstructionForDeletion(L);
2056 // In order to find a leader for a given value number at a
2057 // specific basic block, we first obtain the list of all Values for that number,
2058 // and then scan the list to find one whose block dominates the block in
2059 // question. This is fast because dominator tree queries consist of only
2060 // a few comparisons of DFS numbers.
2061 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
2062 LeaderTableEntry Vals = LeaderTable[num];
2063 if (!Vals.Val) return nullptr;
2065 Value *Val = nullptr;
2066 if (DT->dominates(Vals.BB, BB)) {
2068 if (isa<Constant>(Val)) return Val;
2071 LeaderTableEntry* Next = Vals.Next;
2073 if (DT->dominates(Next->BB, BB)) {
2074 if (isa<Constant>(Next->Val)) return Next->Val;
2075 if (!Val) Val = Next->Val;
2084 /// There is an edge from 'Src' to 'Dst'. Return
2085 /// true if every path from the entry block to 'Dst' passes via this edge. In
2086 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2087 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2088 DominatorTree *DT) {
2089 // While in theory it is interesting to consider the case in which Dst has
2090 // more than one predecessor, because Dst might be part of a loop which is
2091 // only reachable from Src, in practice it is pointless since at the time
2092 // GVN runs all such loops have preheaders, which means that Dst will have
2093 // been changed to have only one predecessor, namely Src.
2094 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2095 const BasicBlock *Src = E.getStart();
2096 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2098 return Pred != nullptr;
2101 // Tries to replace instruction with const, using information from
2102 // ReplaceWithConstMap.
2103 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
2104 bool Changed = false;
2105 for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
2106 Value *Operand = Instr->getOperand(OpNum);
2107 auto it = ReplaceWithConstMap.find(Operand);
2108 if (it != ReplaceWithConstMap.end()) {
2109 assert(!isa<Constant>(Operand) &&
2110 "Replacing constants with constants is invalid");
2111 Instr->setOperand(OpNum, it->second);
2118 /// The given values are known to be equal in every block
2119 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2120 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2121 /// If DominatesByEdge is false, then it means that it is dominated by Root.End.
2122 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
2123 bool DominatesByEdge) {
2124 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2125 Worklist.push_back(std::make_pair(LHS, RHS));
2126 bool Changed = false;
2127 // For speed, compute a conservative fast approximation to
2128 // DT->dominates(Root, Root.getEnd());
2129 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2131 while (!Worklist.empty()) {
2132 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2133 LHS = Item.first; RHS = Item.second;
2137 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2139 // Don't try to propagate equalities between constants.
2140 if (isa<Constant>(LHS) && isa<Constant>(RHS))
2143 // Prefer a constant on the right-hand side, or an Argument if no constants.
2144 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2145 std::swap(LHS, RHS);
2146 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2148 // If there is no obvious reason to prefer the left-hand side over the
2149 // right-hand side, ensure the longest lived term is on the right-hand side,
2150 // so the shortest lived term will be replaced by the longest lived.
2151 // This tends to expose more simplifications.
2152 uint32_t LVN = VN.lookup_or_add(LHS);
2153 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2154 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2155 // Move the 'oldest' value to the right-hand side, using the value number
2156 // as a proxy for age.
2157 uint32_t RVN = VN.lookup_or_add(RHS);
2159 std::swap(LHS, RHS);
2164 // If value numbering later sees that an instruction in the scope is equal
2165 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2166 // the invariant that instructions only occur in the leader table for their
2167 // own value number (this is used by removeFromLeaderTable), do not do this
2168 // if RHS is an instruction (if an instruction in the scope is morphed into
2169 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2170 // using the leader table is about compiling faster, not optimizing better).
2171 // The leader table only tracks basic blocks, not edges. Only add to if we
2172 // have the simple case where the edge dominates the end.
2173 if (RootDominatesEnd && !isa<Instruction>(RHS))
2174 addToLeaderTable(LVN, RHS, Root.getEnd());
2176 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2177 // LHS always has at least one use that is not dominated by Root, this will
2178 // never do anything if LHS has only one use.
2179 if (!LHS->hasOneUse()) {
2180 unsigned NumReplacements =
2182 ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
2183 : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getEnd());
2185 Changed |= NumReplacements > 0;
2186 NumGVNEqProp += NumReplacements;
2189 // Now try to deduce additional equalities from this one. For example, if
2190 // the known equality was "(A != B)" == "false" then it follows that A and B
2191 // are equal in the scope. Only boolean equalities with an explicit true or
2192 // false RHS are currently supported.
2193 if (!RHS->getType()->isIntegerTy(1))
2194 // Not a boolean equality - bail out.
2196 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2198 // RHS neither 'true' nor 'false' - bail out.
2200 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2201 bool isKnownTrue = CI->isAllOnesValue();
2202 bool isKnownFalse = !isKnownTrue;
2204 // If "A && B" is known true then both A and B are known true. If "A || B"
2205 // is known false then both A and B are known false.
2207 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2208 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2209 Worklist.push_back(std::make_pair(A, RHS));
2210 Worklist.push_back(std::make_pair(B, RHS));
2214 // If we are propagating an equality like "(A == B)" == "true" then also
2215 // propagate the equality A == B. When propagating a comparison such as
2216 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2217 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
2218 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2220 // If "A == B" is known true, or "A != B" is known false, then replace
2221 // A with B everywhere in the scope.
2222 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2223 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2224 Worklist.push_back(std::make_pair(Op0, Op1));
2226 // Handle the floating point versions of equality comparisons too.
2227 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
2228 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
2230 // Floating point -0.0 and 0.0 compare equal, so we can only
2231 // propagate values if we know that we have a constant and that
2232 // its value is non-zero.
2234 // FIXME: We should do this optimization if 'no signed zeros' is
2235 // applicable via an instruction-level fast-math-flag or some other
2236 // indicator that relaxed FP semantics are being used.
2238 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
2239 Worklist.push_back(std::make_pair(Op0, Op1));
2242 // If "A >= B" is known true, replace "A < B" with false everywhere.
2243 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2244 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2245 // Since we don't have the instruction "A < B" immediately to hand, work
2246 // out the value number that it would have and use that to find an
2247 // appropriate instruction (if any).
2248 uint32_t NextNum = VN.getNextUnusedValueNumber();
2249 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2250 // If the number we were assigned was brand new then there is no point in
2251 // looking for an instruction realizing it: there cannot be one!
2252 if (Num < NextNum) {
2253 Value *NotCmp = findLeader(Root.getEnd(), Num);
2254 if (NotCmp && isa<Instruction>(NotCmp)) {
2255 unsigned NumReplacements =
2257 ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
2258 : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
2260 Changed |= NumReplacements > 0;
2261 NumGVNEqProp += NumReplacements;
2264 // Ensure that any instruction in scope that gets the "A < B" value number
2265 // is replaced with false.
2266 // The leader table only tracks basic blocks, not edges. Only add to if we
2267 // have the simple case where the edge dominates the end.
2268 if (RootDominatesEnd)
2269 addToLeaderTable(Num, NotVal, Root.getEnd());
2278 /// When calculating availability, handle an instruction
2279 /// by inserting it into the appropriate sets
2280 bool GVN::processInstruction(Instruction *I) {
2281 // Ignore dbg info intrinsics.
2282 if (isa<DbgInfoIntrinsic>(I))
2285 // If the instruction can be easily simplified then do so now in preference
2286 // to value numbering it. Value numbering often exposes redundancies, for
2287 // example if it determines that %y is equal to %x then the instruction
2288 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2289 const DataLayout &DL = I->getModule()->getDataLayout();
2290 if (Value *V = SimplifyInstruction(I, DL, TLI, DT, AC)) {
2291 I->replaceAllUsesWith(V);
2292 if (MD && V->getType()->getScalarType()->isPointerTy())
2293 MD->invalidateCachedPointerInfo(V);
2294 markInstructionForDeletion(I);
2299 if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
2300 if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
2301 return processAssumeIntrinsic(IntrinsicI);
2303 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2304 if (processLoad(LI))
2307 unsigned Num = VN.lookup_or_add(LI);
2308 addToLeaderTable(Num, LI, LI->getParent());
2312 // For conditional branches, we can perform simple conditional propagation on
2313 // the condition value itself.
2314 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2315 if (!BI->isConditional())
2318 if (isa<Constant>(BI->getCondition()))
2319 return processFoldableCondBr(BI);
2321 Value *BranchCond = BI->getCondition();
2322 BasicBlock *TrueSucc = BI->getSuccessor(0);
2323 BasicBlock *FalseSucc = BI->getSuccessor(1);
2324 // Avoid multiple edges early.
2325 if (TrueSucc == FalseSucc)
2328 BasicBlock *Parent = BI->getParent();
2329 bool Changed = false;
2331 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2332 BasicBlockEdge TrueE(Parent, TrueSucc);
2333 Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
2335 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2336 BasicBlockEdge FalseE(Parent, FalseSucc);
2337 Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
2342 // For switches, propagate the case values into the case destinations.
2343 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2344 Value *SwitchCond = SI->getCondition();
2345 BasicBlock *Parent = SI->getParent();
2346 bool Changed = false;
2348 // Remember how many outgoing edges there are to every successor.
2349 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2350 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2351 ++SwitchEdges[SI->getSuccessor(i)];
2353 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2355 BasicBlock *Dst = i.getCaseSuccessor();
2356 // If there is only a single edge, propagate the case value into it.
2357 if (SwitchEdges.lookup(Dst) == 1) {
2358 BasicBlockEdge E(Parent, Dst);
2359 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E, true);
2365 // Instructions with void type don't return a value, so there's
2366 // no point in trying to find redundancies in them.
2367 if (I->getType()->isVoidTy())
2370 uint32_t NextNum = VN.getNextUnusedValueNumber();
2371 unsigned Num = VN.lookup_or_add(I);
2373 // Allocations are always uniquely numbered, so we can save time and memory
2374 // by fast failing them.
2375 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2376 addToLeaderTable(Num, I, I->getParent());
2380 // If the number we were assigned was a brand new VN, then we don't
2381 // need to do a lookup to see if the number already exists
2382 // somewhere in the domtree: it can't!
2383 if (Num >= NextNum) {
2384 addToLeaderTable(Num, I, I->getParent());
2388 // Perform fast-path value-number based elimination of values inherited from
2390 Value *repl = findLeader(I->getParent(), Num);
2392 // Failure, just remember this instance for future use.
2393 addToLeaderTable(Num, I, I->getParent());
2398 patchAndReplaceAllUsesWith(I, repl);
2399 if (MD && repl->getType()->getScalarType()->isPointerTy())
2400 MD->invalidateCachedPointerInfo(repl);
2401 markInstructionForDeletion(I);
2405 /// runOnFunction - This is the main transformation entry point for a function.
2406 bool GVN::runOnFunction(Function& F) {
2407 if (skipOptnoneFunction(F))
2411 MD = &getAnalysis<MemoryDependenceAnalysis>();
2412 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2413 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2414 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
2415 VN.setAliasAnalysis(&getAnalysis<AAResultsWrapperPass>().getAAResults());
2419 bool Changed = false;
2420 bool ShouldContinue = true;
2422 // Merge unconditional branches, allowing PRE to catch more
2423 // optimization opportunities.
2424 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2425 BasicBlock *BB = FI++;
2428 MergeBlockIntoPredecessor(BB, DT, /* LoopInfo */ nullptr, MD);
2429 if (removedBlock) ++NumGVNBlocks;
2431 Changed |= removedBlock;
2434 unsigned Iteration = 0;
2435 while (ShouldContinue) {
2436 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2437 ShouldContinue = iterateOnFunction(F);
2438 Changed |= ShouldContinue;
2443 // Fabricate val-num for dead-code in order to suppress assertion in
2445 assignValNumForDeadCode();
2446 bool PREChanged = true;
2447 while (PREChanged) {
2448 PREChanged = performPRE(F);
2449 Changed |= PREChanged;
2453 // FIXME: Should perform GVN again after PRE does something. PRE can move
2454 // computations into blocks where they become fully redundant. Note that
2455 // we can't do this until PRE's critical edge splitting updates memdep.
2456 // Actually, when this happens, we should just fully integrate PRE into GVN.
2458 cleanupGlobalSets();
2459 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2467 bool GVN::processBlock(BasicBlock *BB) {
2468 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2469 // (and incrementing BI before processing an instruction).
2470 assert(InstrsToErase.empty() &&
2471 "We expect InstrsToErase to be empty across iterations");
2472 if (DeadBlocks.count(BB))
2475 // Clearing map before every BB because it can be used only for single BB.
2476 ReplaceWithConstMap.clear();
2477 bool ChangedFunction = false;
2479 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2481 if (!ReplaceWithConstMap.empty())
2482 ChangedFunction |= replaceOperandsWithConsts(BI);
2483 ChangedFunction |= processInstruction(BI);
2485 if (InstrsToErase.empty()) {
2490 // If we need some instructions deleted, do it now.
2491 NumGVNInstr += InstrsToErase.size();
2493 // Avoid iterator invalidation.
2494 bool AtStart = BI == BB->begin();
2498 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2499 E = InstrsToErase.end(); I != E; ++I) {
2500 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2501 if (MD) MD->removeInstruction(*I);
2502 DEBUG(verifyRemoved(*I));
2503 (*I)->eraseFromParent();
2505 InstrsToErase.clear();
2513 return ChangedFunction;
2516 // Instantiate an expression in a predecessor that lacked it.
2517 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2518 unsigned int ValNo) {
2519 // Because we are going top-down through the block, all value numbers
2520 // will be available in the predecessor by the time we need them. Any
2521 // that weren't originally present will have been instantiated earlier
2523 bool success = true;
2524 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2525 Value *Op = Instr->getOperand(i);
2526 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2529 if (Value *V = findLeader(Pred, VN.lookup(Op))) {
2530 Instr->setOperand(i, V);
2537 // Fail out if we encounter an operand that is not available in
2538 // the PRE predecessor. This is typically because of loads which
2539 // are not value numbered precisely.
2543 Instr->insertBefore(Pred->getTerminator());
2544 Instr->setName(Instr->getName() + ".pre");
2545 Instr->setDebugLoc(Instr->getDebugLoc());
2546 VN.add(Instr, ValNo);
2548 // Update the availability map to include the new instruction.
2549 addToLeaderTable(ValNo, Instr, Pred);
2553 bool GVN::performScalarPRE(Instruction *CurInst) {
2554 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2556 if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2557 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2558 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2559 isa<DbgInfoIntrinsic>(CurInst))
2562 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2563 // sinking the compare again, and it would force the code generator to
2564 // move the i1 from processor flags or predicate registers into a general
2565 // purpose register.
2566 if (isa<CmpInst>(CurInst))
2569 // We don't currently value number ANY inline asm calls.
2570 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2571 if (CallI->isInlineAsm())
2574 uint32_t ValNo = VN.lookup(CurInst);
2576 // Look for the predecessors for PRE opportunities. We're
2577 // only trying to solve the basic diamond case, where
2578 // a value is computed in the successor and one predecessor,
2579 // but not the other. We also explicitly disallow cases
2580 // where the successor is its own predecessor, because they're
2581 // more complicated to get right.
2582 unsigned NumWith = 0;
2583 unsigned NumWithout = 0;
2584 BasicBlock *PREPred = nullptr;
2585 BasicBlock *CurrentBlock = CurInst->getParent();
2588 for (pred_iterator PI = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2590 BasicBlock *P = *PI;
2591 // We're not interested in PRE where the block is its
2592 // own predecessor, or in blocks with predecessors
2593 // that are not reachable.
2594 if (P == CurrentBlock) {
2597 } else if (!DT->isReachableFromEntry(P)) {
2602 Value *predV = findLeader(P, ValNo);
2604 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2607 } else if (predV == CurInst) {
2608 /* CurInst dominates this predecessor. */
2612 predMap.push_back(std::make_pair(predV, P));
2617 // Don't do PRE when it might increase code size, i.e. when
2618 // we would need to insert instructions in more than one pred.
2619 if (NumWithout > 1 || NumWith == 0)
2622 // We may have a case where all predecessors have the instruction,
2623 // and we just need to insert a phi node. Otherwise, perform
2625 Instruction *PREInstr = nullptr;
2627 if (NumWithout != 0) {
2628 // Don't do PRE across indirect branch.
2629 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2632 // We can't do PRE safely on a critical edge, so instead we schedule
2633 // the edge to be split and perform the PRE the next time we iterate
2635 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2636 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2637 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2640 // We need to insert somewhere, so let's give it a shot
2641 PREInstr = CurInst->clone();
2642 if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
2643 // If we failed insertion, make sure we remove the instruction.
2644 DEBUG(verifyRemoved(PREInstr));
2650 // Either we should have filled in the PRE instruction, or we should
2651 // not have needed insertions.
2652 assert (PREInstr != nullptr || NumWithout == 0);
2656 // Create a PHI to make the value available in this block.
2658 PHINode::Create(CurInst->getType(), predMap.size(),
2659 CurInst->getName() + ".pre-phi", CurrentBlock->begin());
2660 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2661 if (Value *V = predMap[i].first)
2662 Phi->addIncoming(V, predMap[i].second);
2664 Phi->addIncoming(PREInstr, PREPred);
2668 addToLeaderTable(ValNo, Phi, CurrentBlock);
2669 Phi->setDebugLoc(CurInst->getDebugLoc());
2670 CurInst->replaceAllUsesWith(Phi);
2671 if (MD && Phi->getType()->getScalarType()->isPointerTy())
2672 MD->invalidateCachedPointerInfo(Phi);
2674 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2676 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2678 MD->removeInstruction(CurInst);
2679 DEBUG(verifyRemoved(CurInst));
2680 CurInst->eraseFromParent();
2686 /// Perform a purely local form of PRE that looks for diamond
2687 /// control flow patterns and attempts to perform simple PRE at the join point.
2688 bool GVN::performPRE(Function &F) {
2689 bool Changed = false;
2690 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2691 // Nothing to PRE in the entry block.
2692 if (CurrentBlock == &F.getEntryBlock())
2695 // Don't perform PRE on an EH pad.
2696 if (CurrentBlock->isEHPad())
2699 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2700 BE = CurrentBlock->end();
2702 Instruction *CurInst = BI++;
2703 Changed = performScalarPRE(CurInst);
2707 if (splitCriticalEdges())
2713 /// Split the critical edge connecting the given two blocks, and return
2714 /// the block inserted to the critical edge.
2715 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2717 SplitCriticalEdge(Pred, Succ, CriticalEdgeSplittingOptions(DT));
2719 MD->invalidateCachedPredecessors();
2723 /// Split critical edges found during the previous
2724 /// iteration that may enable further optimization.
2725 bool GVN::splitCriticalEdges() {
2726 if (toSplit.empty())
2729 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2730 SplitCriticalEdge(Edge.first, Edge.second,
2731 CriticalEdgeSplittingOptions(DT));
2732 } while (!toSplit.empty());
2733 if (MD) MD->invalidateCachedPredecessors();
2737 /// Executes one iteration of GVN
2738 bool GVN::iterateOnFunction(Function &F) {
2739 cleanupGlobalSets();
2741 // Top-down walk of the dominator tree
2742 bool Changed = false;
2743 // Save the blocks this function have before transformation begins. GVN may
2744 // split critical edge, and hence may invalidate the RPO/DT iterator.
2746 std::vector<BasicBlock *> BBVect;
2747 BBVect.reserve(256);
2748 // Needed for value numbering with phi construction to work.
2749 ReversePostOrderTraversal<Function *> RPOT(&F);
2750 for (ReversePostOrderTraversal<Function *>::rpo_iterator RI = RPOT.begin(),
2753 BBVect.push_back(*RI);
2755 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2757 Changed |= processBlock(*I);
2762 void GVN::cleanupGlobalSets() {
2764 LeaderTable.clear();
2765 TableAllocator.Reset();
2768 /// Verify that the specified instruction does not occur in our
2769 /// internal data structures.
2770 void GVN::verifyRemoved(const Instruction *Inst) const {
2771 VN.verifyRemoved(Inst);
2773 // Walk through the value number scope to make sure the instruction isn't
2774 // ferreted away in it.
2775 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2776 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2777 const LeaderTableEntry *Node = &I->second;
2778 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2780 while (Node->Next) {
2782 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2787 /// BB is declared dead, which implied other blocks become dead as well. This
2788 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2789 /// live successors, update their phi nodes by replacing the operands
2790 /// corresponding to dead blocks with UndefVal.
2791 void GVN::addDeadBlock(BasicBlock *BB) {
2792 SmallVector<BasicBlock *, 4> NewDead;
2793 SmallSetVector<BasicBlock *, 4> DF;
2795 NewDead.push_back(BB);
2796 while (!NewDead.empty()) {
2797 BasicBlock *D = NewDead.pop_back_val();
2798 if (DeadBlocks.count(D))
2801 // All blocks dominated by D are dead.
2802 SmallVector<BasicBlock *, 8> Dom;
2803 DT->getDescendants(D, Dom);
2804 DeadBlocks.insert(Dom.begin(), Dom.end());
2806 // Figure out the dominance-frontier(D).
2807 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
2808 E = Dom.end(); I != E; I++) {
2810 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
2811 BasicBlock *S = *SI;
2812 if (DeadBlocks.count(S))
2815 bool AllPredDead = true;
2816 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
2817 if (!DeadBlocks.count(*PI)) {
2818 AllPredDead = false;
2823 // S could be proved dead later on. That is why we don't update phi
2824 // operands at this moment.
2827 // While S is not dominated by D, it is dead by now. This could take
2828 // place if S already have a dead predecessor before D is declared
2830 NewDead.push_back(S);
2836 // For the dead blocks' live successors, update their phi nodes by replacing
2837 // the operands corresponding to dead blocks with UndefVal.
2838 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2841 if (DeadBlocks.count(B))
2844 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2845 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
2846 PE = Preds.end(); PI != PE; PI++) {
2847 BasicBlock *P = *PI;
2849 if (!DeadBlocks.count(P))
2852 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2853 if (BasicBlock *S = splitCriticalEdges(P, B))
2854 DeadBlocks.insert(P = S);
2857 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2858 PHINode &Phi = cast<PHINode>(*II);
2859 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2860 UndefValue::get(Phi.getType()));
2866 // If the given branch is recognized as a foldable branch (i.e. conditional
2867 // branch with constant condition), it will perform following analyses and
2869 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2870 // R be the target of the dead out-coming edge.
2871 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2872 // edge. The result of this step will be {X| X is dominated by R}
2873 // 2) Identify those blocks which haves at least one dead predecessor. The
2874 // result of this step will be dominance-frontier(R).
2875 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2876 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2878 // Return true iff *NEW* dead code are found.
2879 bool GVN::processFoldableCondBr(BranchInst *BI) {
2880 if (!BI || BI->isUnconditional())
2883 // If a branch has two identical successors, we cannot declare either dead.
2884 if (BI->getSuccessor(0) == BI->getSuccessor(1))
2887 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2891 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2892 BI->getSuccessor(1) : BI->getSuccessor(0);
2893 if (DeadBlocks.count(DeadRoot))
2896 if (!DeadRoot->getSinglePredecessor())
2897 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2899 addDeadBlock(DeadRoot);
2903 // performPRE() will trigger assert if it comes across an instruction without
2904 // associated val-num. As it normally has far more live instructions than dead
2905 // instructions, it makes more sense just to "fabricate" a val-number for the
2906 // dead code than checking if instruction involved is dead or not.
2907 void GVN::assignValNumForDeadCode() {
2908 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
2909 E = DeadBlocks.end(); I != E; I++) {
2910 BasicBlock *BB = *I;
2911 for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
2913 Instruction *Inst = &*II;
2914 unsigned ValNum = VN.lookup_or_add(Inst);
2915 addToLeaderTable(ValNum, Inst, BB);