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/GVN.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/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51 #include "llvm/Transforms/Utils/Local.h"
52 #include "llvm/Transforms/Utils/SSAUpdater.h"
55 using namespace llvm::gvn;
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 struct llvm::GVN::Expression {
80 SmallVector<uint32_t, 4> varargs;
82 Expression(uint32_t o = ~2U) : opcode(o) {}
84 bool operator==(const Expression &other) const {
85 if (opcode != other.opcode)
87 if (opcode == ~0U || opcode == ~1U)
89 if (type != other.type)
91 if (varargs != other.varargs)
96 friend hash_code hash_value(const Expression &Value) {
98 Value.opcode, Value.type,
99 hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
104 template <> struct DenseMapInfo<GVN::Expression> {
105 static inline GVN::Expression getEmptyKey() { return ~0U; }
107 static inline GVN::Expression getTombstoneKey() { return ~1U; }
109 static unsigned getHashValue(const GVN::Expression &e) {
110 using llvm::hash_value;
111 return static_cast<unsigned>(hash_value(e));
113 static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
117 } // End llvm namespace.
119 /// Represents a particular available value that we know how to materialize.
120 /// Materialization of an AvailableValue never fails. An AvailableValue is
121 /// implicitly associated with a rematerialization point which is the
122 /// location of the instruction from which it was formed.
123 struct llvm::gvn::AvailableValue {
125 SimpleVal, // A simple offsetted value that is accessed.
126 LoadVal, // A value produced by a load.
127 MemIntrin, // A memory intrinsic which is loaded from.
128 UndefVal // A UndefValue representing a value from dead block (which
129 // is not yet physically removed from the CFG).
132 /// V - The value that is live out of the block.
133 PointerIntPair<Value *, 2, ValType> Val;
135 /// Offset - The byte offset in Val that is interesting for the load query.
138 static AvailableValue get(Value *V, unsigned Offset = 0) {
140 Res.Val.setPointer(V);
141 Res.Val.setInt(SimpleVal);
146 static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
148 Res.Val.setPointer(MI);
149 Res.Val.setInt(MemIntrin);
154 static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
156 Res.Val.setPointer(LI);
157 Res.Val.setInt(LoadVal);
162 static AvailableValue getUndef() {
164 Res.Val.setPointer(nullptr);
165 Res.Val.setInt(UndefVal);
170 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
171 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
172 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
173 bool isUndefValue() const { return Val.getInt() == UndefVal; }
175 Value *getSimpleValue() const {
176 assert(isSimpleValue() && "Wrong accessor");
177 return Val.getPointer();
180 LoadInst *getCoercedLoadValue() const {
181 assert(isCoercedLoadValue() && "Wrong accessor");
182 return cast<LoadInst>(Val.getPointer());
185 MemIntrinsic *getMemIntrinValue() const {
186 assert(isMemIntrinValue() && "Wrong accessor");
187 return cast<MemIntrinsic>(Val.getPointer());
190 /// Emit code at the specified insertion point to adjust the value defined
191 /// here to the specified type. This handles various coercion cases.
192 Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
196 /// Represents an AvailableValue which can be rematerialized at the end of
197 /// the associated BasicBlock.
198 struct llvm::gvn::AvailableValueInBlock {
199 /// BB - The basic block in question.
202 /// AV - The actual available value
205 static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV) {
206 AvailableValueInBlock Res;
208 Res.AV = std::move(AV);
212 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
213 unsigned Offset = 0) {
214 return get(BB, AvailableValue::get(V, Offset));
216 static AvailableValueInBlock getUndef(BasicBlock *BB) {
217 return get(BB, AvailableValue::getUndef());
220 /// Emit code at the end of this block to adjust the value defined here to
221 /// the specified type. This handles various coercion cases.
222 Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const {
223 return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
227 //===----------------------------------------------------------------------===//
228 // ValueTable Internal Functions
229 //===----------------------------------------------------------------------===//
231 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
233 e.type = I->getType();
234 e.opcode = I->getOpcode();
235 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
237 e.varargs.push_back(lookupOrAdd(*OI));
238 if (I->isCommutative()) {
239 // Ensure that commutative instructions that only differ by a permutation
240 // of their operands get the same value number by sorting the operand value
241 // numbers. Since all commutative instructions have two operands it is more
242 // efficient to sort by hand rather than using, say, std::sort.
243 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
244 if (e.varargs[0] > e.varargs[1])
245 std::swap(e.varargs[0], e.varargs[1]);
248 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
249 // Sort the operand value numbers so x<y and y>x get the same value number.
250 CmpInst::Predicate Predicate = C->getPredicate();
251 if (e.varargs[0] > e.varargs[1]) {
252 std::swap(e.varargs[0], e.varargs[1]);
253 Predicate = CmpInst::getSwappedPredicate(Predicate);
255 e.opcode = (C->getOpcode() << 8) | Predicate;
256 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
257 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
259 e.varargs.push_back(*II);
265 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
266 CmpInst::Predicate Predicate,
267 Value *LHS, Value *RHS) {
268 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
269 "Not a comparison!");
271 e.type = CmpInst::makeCmpResultType(LHS->getType());
272 e.varargs.push_back(lookupOrAdd(LHS));
273 e.varargs.push_back(lookupOrAdd(RHS));
275 // Sort the operand value numbers so x<y and y>x get the same value number.
276 if (e.varargs[0] > e.varargs[1]) {
277 std::swap(e.varargs[0], e.varargs[1]);
278 Predicate = CmpInst::getSwappedPredicate(Predicate);
280 e.opcode = (Opcode << 8) | Predicate;
284 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
285 assert(EI && "Not an ExtractValueInst?");
287 e.type = EI->getType();
290 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
291 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
292 // EI might be an extract from one of our recognised intrinsics. If it
293 // is we'll synthesize a semantically equivalent expression instead on
294 // an extract value expression.
295 switch (I->getIntrinsicID()) {
296 case Intrinsic::sadd_with_overflow:
297 case Intrinsic::uadd_with_overflow:
298 e.opcode = Instruction::Add;
300 case Intrinsic::ssub_with_overflow:
301 case Intrinsic::usub_with_overflow:
302 e.opcode = Instruction::Sub;
304 case Intrinsic::smul_with_overflow:
305 case Intrinsic::umul_with_overflow:
306 e.opcode = Instruction::Mul;
313 // Intrinsic recognized. Grab its args to finish building the expression.
314 assert(I->getNumArgOperands() == 2 &&
315 "Expect two args for recognised intrinsics.");
316 e.varargs.push_back(lookupOrAdd(I->getArgOperand(0)));
317 e.varargs.push_back(lookupOrAdd(I->getArgOperand(1)));
322 // Not a recognised intrinsic. Fall back to producing an extract value
324 e.opcode = EI->getOpcode();
325 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
327 e.varargs.push_back(lookupOrAdd(*OI));
329 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
331 e.varargs.push_back(*II);
336 //===----------------------------------------------------------------------===//
337 // ValueTable External Functions
338 //===----------------------------------------------------------------------===//
340 GVN::ValueTable::ValueTable() : nextValueNumber(1) {}
341 GVN::ValueTable::ValueTable(const ValueTable &Arg)
342 : valueNumbering(Arg.valueNumbering),
343 expressionNumbering(Arg.expressionNumbering), AA(Arg.AA), MD(Arg.MD),
344 DT(Arg.DT), nextValueNumber(Arg.nextValueNumber) {}
345 GVN::ValueTable::ValueTable(ValueTable &&Arg)
346 : valueNumbering(std::move(Arg.valueNumbering)),
347 expressionNumbering(std::move(Arg.expressionNumbering)),
348 AA(std::move(Arg.AA)), MD(std::move(Arg.MD)), DT(std::move(Arg.DT)),
349 nextValueNumber(std::move(Arg.nextValueNumber)) {}
350 GVN::ValueTable::~ValueTable() {}
352 /// add - Insert a value into the table with a specified value number.
353 void GVN::ValueTable::add(Value *V, uint32_t num) {
354 valueNumbering.insert(std::make_pair(V, num));
357 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
358 if (AA->doesNotAccessMemory(C)) {
359 Expression exp = createExpr(C);
360 uint32_t &e = expressionNumbering[exp];
361 if (!e) e = nextValueNumber++;
362 valueNumbering[C] = e;
364 } else if (AA->onlyReadsMemory(C)) {
365 Expression exp = createExpr(C);
366 uint32_t &e = expressionNumbering[exp];
368 e = nextValueNumber++;
369 valueNumbering[C] = e;
373 e = nextValueNumber++;
374 valueNumbering[C] = e;
378 MemDepResult local_dep = MD->getDependency(C);
380 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
381 valueNumbering[C] = nextValueNumber;
382 return nextValueNumber++;
385 if (local_dep.isDef()) {
386 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
388 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
389 valueNumbering[C] = nextValueNumber;
390 return nextValueNumber++;
393 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
394 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
395 uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
397 valueNumbering[C] = nextValueNumber;
398 return nextValueNumber++;
402 uint32_t v = lookupOrAdd(local_cdep);
403 valueNumbering[C] = v;
408 const MemoryDependenceResults::NonLocalDepInfo &deps =
409 MD->getNonLocalCallDependency(CallSite(C));
410 // FIXME: Move the checking logic to MemDep!
411 CallInst* cdep = nullptr;
413 // Check to see if we have a single dominating call instruction that is
415 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
416 const NonLocalDepEntry *I = &deps[i];
417 if (I->getResult().isNonLocal())
420 // We don't handle non-definitions. If we already have a call, reject
421 // instruction dependencies.
422 if (!I->getResult().isDef() || cdep != nullptr) {
427 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
428 // FIXME: All duplicated with non-local case.
429 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
430 cdep = NonLocalDepCall;
439 valueNumbering[C] = nextValueNumber;
440 return nextValueNumber++;
443 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
444 valueNumbering[C] = nextValueNumber;
445 return nextValueNumber++;
447 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
448 uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
449 uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
451 valueNumbering[C] = nextValueNumber;
452 return nextValueNumber++;
456 uint32_t v = lookupOrAdd(cdep);
457 valueNumbering[C] = v;
461 valueNumbering[C] = nextValueNumber;
462 return nextValueNumber++;
466 /// Returns true if a value number exists for the specified value.
467 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
469 /// lookup_or_add - Returns the value number for the specified value, assigning
470 /// it a new number if it did not have one before.
471 uint32_t GVN::ValueTable::lookupOrAdd(Value *V) {
472 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
473 if (VI != valueNumbering.end())
476 if (!isa<Instruction>(V)) {
477 valueNumbering[V] = nextValueNumber;
478 return nextValueNumber++;
481 Instruction* I = cast<Instruction>(V);
483 switch (I->getOpcode()) {
484 case Instruction::Call:
485 return lookupOrAddCall(cast<CallInst>(I));
486 case Instruction::Add:
487 case Instruction::FAdd:
488 case Instruction::Sub:
489 case Instruction::FSub:
490 case Instruction::Mul:
491 case Instruction::FMul:
492 case Instruction::UDiv:
493 case Instruction::SDiv:
494 case Instruction::FDiv:
495 case Instruction::URem:
496 case Instruction::SRem:
497 case Instruction::FRem:
498 case Instruction::Shl:
499 case Instruction::LShr:
500 case Instruction::AShr:
501 case Instruction::And:
502 case Instruction::Or:
503 case Instruction::Xor:
504 case Instruction::ICmp:
505 case Instruction::FCmp:
506 case Instruction::Trunc:
507 case Instruction::ZExt:
508 case Instruction::SExt:
509 case Instruction::FPToUI:
510 case Instruction::FPToSI:
511 case Instruction::UIToFP:
512 case Instruction::SIToFP:
513 case Instruction::FPTrunc:
514 case Instruction::FPExt:
515 case Instruction::PtrToInt:
516 case Instruction::IntToPtr:
517 case Instruction::BitCast:
518 case Instruction::Select:
519 case Instruction::ExtractElement:
520 case Instruction::InsertElement:
521 case Instruction::ShuffleVector:
522 case Instruction::InsertValue:
523 case Instruction::GetElementPtr:
526 case Instruction::ExtractValue:
527 exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
530 valueNumbering[V] = nextValueNumber;
531 return nextValueNumber++;
534 uint32_t& e = expressionNumbering[exp];
535 if (!e) e = nextValueNumber++;
536 valueNumbering[V] = e;
540 /// Returns the value number of the specified value. Fails if
541 /// the value has not yet been numbered.
542 uint32_t GVN::ValueTable::lookup(Value *V) const {
543 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
544 assert(VI != valueNumbering.end() && "Value not numbered?");
548 /// Returns the value number of the given comparison,
549 /// assigning it a new number if it did not have one before. Useful when
550 /// we deduced the result of a comparison, but don't immediately have an
551 /// instruction realizing that comparison to hand.
552 uint32_t GVN::ValueTable::lookupOrAddCmp(unsigned Opcode,
553 CmpInst::Predicate Predicate,
554 Value *LHS, Value *RHS) {
555 Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
556 uint32_t& e = expressionNumbering[exp];
557 if (!e) e = nextValueNumber++;
561 /// Remove all entries from the ValueTable.
562 void GVN::ValueTable::clear() {
563 valueNumbering.clear();
564 expressionNumbering.clear();
568 /// Remove a value from the value numbering.
569 void GVN::ValueTable::erase(Value *V) {
570 valueNumbering.erase(V);
573 /// verifyRemoved - Verify that the value is removed from all internal data
575 void GVN::ValueTable::verifyRemoved(const Value *V) const {
576 for (DenseMap<Value*, uint32_t>::const_iterator
577 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
578 assert(I->first != V && "Inst still occurs in value numbering map!");
582 //===----------------------------------------------------------------------===//
584 //===----------------------------------------------------------------------===//
586 PreservedAnalyses GVN::run(Function &F, AnalysisManager<Function> &AM) {
587 // FIXME: The order of evaluation of these 'getResult' calls is very
588 // significant! Re-ordering these variables will cause GVN when run alone to
589 // be less effective! We should fix memdep and basic-aa to not exhibit this
590 // behavior, but until then don't change the order here.
591 auto &AC = AM.getResult<AssumptionAnalysis>(F);
592 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
593 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
594 auto &AA = AM.getResult<AAManager>(F);
595 auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
596 bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep);
598 return PreservedAnalyses::all();
599 PreservedAnalyses PA;
600 PA.preserve<DominatorTreeAnalysis>();
601 PA.preserve<GlobalsAA>();
606 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
608 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
609 E = d.end(); I != E; ++I) {
610 errs() << I->first << "\n";
616 /// Return true if we can prove that the value
617 /// we're analyzing is fully available in the specified block. As we go, keep
618 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
619 /// map is actually a tri-state map with the following values:
620 /// 0) we know the block *is not* fully available.
621 /// 1) we know the block *is* fully available.
622 /// 2) we do not know whether the block is fully available or not, but we are
623 /// currently speculating that it will be.
624 /// 3) we are speculating for this block and have used that to speculate for
626 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
627 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
628 uint32_t RecurseDepth) {
629 if (RecurseDepth > MaxRecurseDepth)
632 // Optimistically assume that the block is fully available and check to see
633 // if we already know about this block in one lookup.
634 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
635 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
637 // If the entry already existed for this block, return the precomputed value.
639 // If this is a speculative "available" value, mark it as being used for
640 // speculation of other blocks.
641 if (IV.first->second == 2)
642 IV.first->second = 3;
643 return IV.first->second != 0;
646 // Otherwise, see if it is fully available in all predecessors.
647 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
649 // If this block has no predecessors, it isn't live-in here.
651 goto SpeculationFailure;
653 for (; PI != PE; ++PI)
654 // If the value isn't fully available in one of our predecessors, then it
655 // isn't fully available in this block either. Undo our previous
656 // optimistic assumption and bail out.
657 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
658 goto SpeculationFailure;
662 // If we get here, we found out that this is not, after
663 // all, a fully-available block. We have a problem if we speculated on this and
664 // used the speculation to mark other blocks as available.
666 char &BBVal = FullyAvailableBlocks[BB];
668 // If we didn't speculate on this, just return with it set to false.
674 // If we did speculate on this value, we could have blocks set to 1 that are
675 // incorrect. Walk the (transitive) successors of this block and mark them as
677 SmallVector<BasicBlock*, 32> BBWorklist;
678 BBWorklist.push_back(BB);
681 BasicBlock *Entry = BBWorklist.pop_back_val();
682 // Note that this sets blocks to 0 (unavailable) if they happen to not
683 // already be in FullyAvailableBlocks. This is safe.
684 char &EntryVal = FullyAvailableBlocks[Entry];
685 if (EntryVal == 0) continue; // Already unavailable.
687 // Mark as unavailable.
690 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
691 } while (!BBWorklist.empty());
697 /// Return true if CoerceAvailableValueToLoadType will succeed.
698 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
700 const DataLayout &DL) {
701 // If the loaded or stored value is an first class array or struct, don't try
702 // to transform them. We need to be able to bitcast to integer.
703 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
704 StoredVal->getType()->isStructTy() ||
705 StoredVal->getType()->isArrayTy())
708 // The store has to be at least as big as the load.
709 if (DL.getTypeSizeInBits(StoredVal->getType()) <
710 DL.getTypeSizeInBits(LoadTy))
716 /// If we saw a store of a value to memory, and
717 /// then a load from a must-aliased pointer of a different type, try to coerce
718 /// the stored value. LoadedTy is the type of the load we want to replace.
719 /// IRB is IRBuilder used to insert new instructions.
721 /// If we can't do it, return null.
722 static Value *CoerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy,
724 const DataLayout &DL) {
725 assert(CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL) &&
726 "precondition violation - materialization can't fail");
728 if (auto *CExpr = dyn_cast<ConstantExpr>(StoredVal))
729 StoredVal = ConstantFoldConstantExpression(CExpr, DL);
731 // If this is already the right type, just return it.
732 Type *StoredValTy = StoredVal->getType();
734 uint64_t StoredValSize = DL.getTypeSizeInBits(StoredValTy);
735 uint64_t LoadedValSize = DL.getTypeSizeInBits(LoadedTy);
737 // If the store and reload are the same size, we can always reuse it.
738 if (StoredValSize == LoadedValSize) {
739 // Pointer to Pointer -> use bitcast.
740 if (StoredValTy->getScalarType()->isPointerTy() &&
741 LoadedTy->getScalarType()->isPointerTy()) {
742 StoredVal = IRB.CreateBitCast(StoredVal, LoadedTy);
744 // Convert source pointers to integers, which can be bitcast.
745 if (StoredValTy->getScalarType()->isPointerTy()) {
746 StoredValTy = DL.getIntPtrType(StoredValTy);
747 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
750 Type *TypeToCastTo = LoadedTy;
751 if (TypeToCastTo->getScalarType()->isPointerTy())
752 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
754 if (StoredValTy != TypeToCastTo)
755 StoredVal = IRB.CreateBitCast(StoredVal, TypeToCastTo);
757 // Cast to pointer if the load needs a pointer type.
758 if (LoadedTy->getScalarType()->isPointerTy())
759 StoredVal = IRB.CreateIntToPtr(StoredVal, LoadedTy);
762 if (auto *CExpr = dyn_cast<ConstantExpr>(StoredVal))
763 StoredVal = ConstantFoldConstantExpression(CExpr, DL);
768 // If the loaded value is smaller than the available value, then we can
769 // extract out a piece from it. If the available value is too small, then we
770 // can't do anything.
771 assert(StoredValSize >= LoadedValSize &&
772 "CanCoerceMustAliasedValueToLoad fail");
774 // Convert source pointers to integers, which can be manipulated.
775 if (StoredValTy->getScalarType()->isPointerTy()) {
776 StoredValTy = DL.getIntPtrType(StoredValTy);
777 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
780 // Convert vectors and fp to integer, which can be manipulated.
781 if (!StoredValTy->isIntegerTy()) {
782 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoredValSize);
783 StoredVal = IRB.CreateBitCast(StoredVal, StoredValTy);
786 // If this is a big-endian system, we need to shift the value down to the low
787 // bits so that a truncate will work.
788 if (DL.isBigEndian()) {
789 uint64_t ShiftAmt = DL.getTypeStoreSizeInBits(StoredValTy) -
790 DL.getTypeStoreSizeInBits(LoadedTy);
791 StoredVal = IRB.CreateLShr(StoredVal, ShiftAmt, "tmp");
794 // Truncate the integer to the right size now.
795 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadedValSize);
796 StoredVal = IRB.CreateTrunc(StoredVal, NewIntTy, "trunc");
798 if (LoadedTy != NewIntTy) {
799 // If the result is a pointer, inttoptr.
800 if (LoadedTy->getScalarType()->isPointerTy())
801 StoredVal = IRB.CreateIntToPtr(StoredVal, LoadedTy, "inttoptr");
803 // Otherwise, bitcast.
804 StoredVal = IRB.CreateBitCast(StoredVal, LoadedTy, "bitcast");
807 if (auto *CExpr = dyn_cast<ConstantExpr>(StoredVal))
808 StoredVal = ConstantFoldConstantExpression(CExpr, DL);
813 /// This function is called when we have a
814 /// memdep query of a load that ends up being a clobbering memory write (store,
815 /// memset, memcpy, memmove). This means that the write *may* provide bits used
816 /// by the load but we can't be sure because the pointers don't mustalias.
818 /// Check this case to see if there is anything more we can do before we give
819 /// up. This returns -1 if we have to give up, or a byte number in the stored
820 /// value of the piece that feeds the load.
821 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
823 uint64_t WriteSizeInBits,
824 const DataLayout &DL) {
825 // If the loaded or stored value is a first class array or struct, don't try
826 // to transform them. We need to be able to bitcast to integer.
827 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
830 int64_t StoreOffset = 0, LoadOffset = 0;
832 GetPointerBaseWithConstantOffset(WritePtr, StoreOffset, DL);
833 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, DL);
834 if (StoreBase != LoadBase)
837 // If the load and store are to the exact same address, they should have been
838 // a must alias. AA must have gotten confused.
839 // FIXME: Study to see if/when this happens. One case is forwarding a memset
840 // to a load from the base of the memset.
842 if (LoadOffset == StoreOffset) {
843 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
844 << "Base = " << *StoreBase << "\n"
845 << "Store Ptr = " << *WritePtr << "\n"
846 << "Store Offs = " << StoreOffset << "\n"
847 << "Load Ptr = " << *LoadPtr << "\n";
852 // If the load and store don't overlap at all, the store doesn't provide
853 // anything to the load. In this case, they really don't alias at all, AA
854 // must have gotten confused.
855 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
857 if ((WriteSizeInBits & 7) | (LoadSize & 7))
859 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
863 bool isAAFailure = false;
864 if (StoreOffset < LoadOffset)
865 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
867 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
871 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
872 << "Base = " << *StoreBase << "\n"
873 << "Store Ptr = " << *WritePtr << "\n"
874 << "Store Offs = " << StoreOffset << "\n"
875 << "Load Ptr = " << *LoadPtr << "\n";
881 // If the Load isn't completely contained within the stored bits, we don't
882 // have all the bits to feed it. We could do something crazy in the future
883 // (issue a smaller load then merge the bits in) but this seems unlikely to be
885 if (StoreOffset > LoadOffset ||
886 StoreOffset+StoreSize < LoadOffset+LoadSize)
889 // Okay, we can do this transformation. Return the number of bytes into the
890 // store that the load is.
891 return LoadOffset-StoreOffset;
894 /// This function is called when we have a
895 /// memdep query of a load that ends up being a clobbering store.
896 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
898 // Cannot handle reading from store of first-class aggregate yet.
899 if (DepSI->getValueOperand()->getType()->isStructTy() ||
900 DepSI->getValueOperand()->getType()->isArrayTy())
903 const DataLayout &DL = DepSI->getModule()->getDataLayout();
904 Value *StorePtr = DepSI->getPointerOperand();
905 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
906 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
907 StorePtr, StoreSize, DL);
910 /// This function is called when we have a
911 /// memdep query of a load that ends up being clobbered by another load. See if
912 /// the other load can feed into the second load.
913 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
914 LoadInst *DepLI, const DataLayout &DL){
915 // Cannot handle reading from store of first-class aggregate yet.
916 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
919 Value *DepPtr = DepLI->getPointerOperand();
920 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
921 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
922 if (R != -1) return R;
924 // If we have a load/load clobber an DepLI can be widened to cover this load,
925 // then we should widen it!
926 int64_t LoadOffs = 0;
927 const Value *LoadBase =
928 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, DL);
929 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
931 unsigned Size = MemoryDependenceResults::getLoadLoadClobberFullWidthSize(
932 LoadBase, LoadOffs, LoadSize, DepLI);
933 if (Size == 0) return -1;
935 // Check non-obvious conditions enforced by MDA which we rely on for being
936 // able to materialize this potentially available value
937 assert(DepLI->isSimple() && "Cannot widen volatile/atomic load!");
938 assert(DepLI->getType()->isIntegerTy() && "Can't widen non-integer load");
940 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
945 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
947 const DataLayout &DL) {
948 // If the mem operation is a non-constant size, we can't handle it.
949 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
950 if (!SizeCst) return -1;
951 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
953 // If this is memset, we just need to see if the offset is valid in the size
955 if (MI->getIntrinsicID() == Intrinsic::memset)
956 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
959 // If we have a memcpy/memmove, the only case we can handle is if this is a
960 // copy from constant memory. In that case, we can read directly from the
962 MemTransferInst *MTI = cast<MemTransferInst>(MI);
964 Constant *Src = dyn_cast<Constant>(MTI->getSource());
967 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, DL));
968 if (!GV || !GV->isConstant()) return -1;
970 // See if the access is within the bounds of the transfer.
971 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
972 MI->getDest(), MemSizeInBits, DL);
976 unsigned AS = Src->getType()->getPointerAddressSpace();
977 // Otherwise, see if we can constant fold a load from the constant with the
978 // offset applied as appropriate.
979 Src = ConstantExpr::getBitCast(Src,
980 Type::getInt8PtrTy(Src->getContext(), AS));
981 Constant *OffsetCst =
982 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
983 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
985 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
986 if (ConstantFoldLoadFromConstPtr(Src, LoadTy, DL))
992 /// This function is called when we have a
993 /// memdep query of a load that ends up being a clobbering store. This means
994 /// that the store provides bits used by the load but we the pointers don't
995 /// mustalias. Check this case to see if there is anything more we can do
996 /// before we give up.
997 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
999 Instruction *InsertPt, const DataLayout &DL){
1000 LLVMContext &Ctx = SrcVal->getType()->getContext();
1002 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1003 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
1005 IRBuilder<> Builder(InsertPt);
1007 // Compute which bits of the stored value are being used by the load. Convert
1008 // to an integer type to start with.
1009 if (SrcVal->getType()->getScalarType()->isPointerTy())
1010 SrcVal = Builder.CreatePtrToInt(SrcVal,
1011 DL.getIntPtrType(SrcVal->getType()));
1012 if (!SrcVal->getType()->isIntegerTy())
1013 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1015 // Shift the bits to the least significant depending on endianness.
1017 if (DL.isLittleEndian())
1018 ShiftAmt = Offset*8;
1020 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1023 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1025 if (LoadSize != StoreSize)
1026 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1028 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, Builder, DL);
1031 /// This function is called when we have a
1032 /// memdep query of a load that ends up being a clobbering load. This means
1033 /// that the load *may* provide bits used by the load but we can't be sure
1034 /// because the pointers don't mustalias. Check this case to see if there is
1035 /// anything more we can do before we give up.
1036 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1037 Type *LoadTy, Instruction *InsertPt,
1039 const DataLayout &DL = SrcVal->getModule()->getDataLayout();
1040 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1041 // widen SrcVal out to a larger load.
1042 unsigned SrcValStoreSize = DL.getTypeStoreSize(SrcVal->getType());
1043 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1044 if (Offset+LoadSize > SrcValStoreSize) {
1045 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1046 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1047 // If we have a load/load clobber an DepLI can be widened to cover this
1048 // load, then we should widen it to the next power of 2 size big enough!
1049 unsigned NewLoadSize = Offset+LoadSize;
1050 if (!isPowerOf2_32(NewLoadSize))
1051 NewLoadSize = NextPowerOf2(NewLoadSize);
1053 Value *PtrVal = SrcVal->getPointerOperand();
1055 // Insert the new load after the old load. This ensures that subsequent
1056 // memdep queries will find the new load. We can't easily remove the old
1057 // load completely because it is already in the value numbering table.
1058 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1060 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1061 DestPTy = PointerType::get(DestPTy,
1062 PtrVal->getType()->getPointerAddressSpace());
1063 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1064 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1065 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1066 NewLoad->takeName(SrcVal);
1067 NewLoad->setAlignment(SrcVal->getAlignment());
1069 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1070 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1072 // Replace uses of the original load with the wider load. On a big endian
1073 // system, we need to shift down to get the relevant bits.
1074 Value *RV = NewLoad;
1075 if (DL.isBigEndian())
1076 RV = Builder.CreateLShr(RV, (NewLoadSize - SrcValStoreSize) * 8);
1077 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1078 SrcVal->replaceAllUsesWith(RV);
1080 // We would like to use gvn.markInstructionForDeletion here, but we can't
1081 // because the load is already memoized into the leader map table that GVN
1082 // tracks. It is potentially possible to remove the load from the table,
1083 // but then there all of the operations based on it would need to be
1084 // rehashed. Just leave the dead load around.
1085 gvn.getMemDep().removeInstruction(SrcVal);
1089 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1093 /// This function is called when we have a
1094 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1095 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1096 Type *LoadTy, Instruction *InsertPt,
1097 const DataLayout &DL){
1098 LLVMContext &Ctx = LoadTy->getContext();
1099 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1101 IRBuilder<> Builder(InsertPt);
1103 // We know that this method is only called when the mem transfer fully
1104 // provides the bits for the load.
1105 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1106 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1107 // independently of what the offset is.
1108 Value *Val = MSI->getValue();
1110 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1112 Value *OneElt = Val;
1114 // Splat the value out to the right number of bits.
1115 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1116 // If we can double the number of bytes set, do it.
1117 if (NumBytesSet*2 <= LoadSize) {
1118 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1119 Val = Builder.CreateOr(Val, ShVal);
1124 // Otherwise insert one byte at a time.
1125 Value *ShVal = Builder.CreateShl(Val, 1*8);
1126 Val = Builder.CreateOr(OneElt, ShVal);
1130 return CoerceAvailableValueToLoadType(Val, LoadTy, Builder, DL);
1133 // Otherwise, this is a memcpy/memmove from a constant global.
1134 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1135 Constant *Src = cast<Constant>(MTI->getSource());
1136 unsigned AS = Src->getType()->getPointerAddressSpace();
1138 // Otherwise, see if we can constant fold a load from the constant with the
1139 // offset applied as appropriate.
1140 Src = ConstantExpr::getBitCast(Src,
1141 Type::getInt8PtrTy(Src->getContext(), AS));
1142 Constant *OffsetCst =
1143 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1144 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1146 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1147 return ConstantFoldLoadFromConstPtr(Src, LoadTy, DL);
1151 /// Given a set of loads specified by ValuesPerBlock,
1152 /// construct SSA form, allowing us to eliminate LI. This returns the value
1153 /// that should be used at LI's definition site.
1154 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1155 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1157 // Check for the fully redundant, dominating load case. In this case, we can
1158 // just use the dominating value directly.
1159 if (ValuesPerBlock.size() == 1 &&
1160 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1162 assert(!ValuesPerBlock[0].AV.isUndefValue() &&
1163 "Dead BB dominate this block");
1164 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
1167 // Otherwise, we have to construct SSA form.
1168 SmallVector<PHINode*, 8> NewPHIs;
1169 SSAUpdater SSAUpdate(&NewPHIs);
1170 SSAUpdate.Initialize(LI->getType(), LI->getName());
1172 for (const AvailableValueInBlock &AV : ValuesPerBlock) {
1173 BasicBlock *BB = AV.BB;
1175 if (SSAUpdate.HasValueForBlock(BB))
1178 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
1181 // Perform PHI construction.
1182 return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1185 Value *AvailableValue::MaterializeAdjustedValue(LoadInst *LI,
1186 Instruction *InsertPt,
1189 Type *LoadTy = LI->getType();
1190 const DataLayout &DL = LI->getModule()->getDataLayout();
1191 if (isSimpleValue()) {
1192 Res = getSimpleValue();
1193 if (Res->getType() != LoadTy) {
1194 Res = GetStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
1196 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1197 << *getSimpleValue() << '\n'
1198 << *Res << '\n' << "\n\n\n");
1200 } else if (isCoercedLoadValue()) {
1201 LoadInst *Load = getCoercedLoadValue();
1202 if (Load->getType() == LoadTy && Offset == 0) {
1205 Res = GetLoadValueForLoad(Load, Offset, LoadTy, InsertPt, gvn);
1207 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1208 << *getCoercedLoadValue() << '\n'
1209 << *Res << '\n' << "\n\n\n");
1211 } else if (isMemIntrinValue()) {
1212 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
1214 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1215 << " " << *getMemIntrinValue() << '\n'
1216 << *Res << '\n' << "\n\n\n");
1218 assert(isUndefValue() && "Should be UndefVal");
1219 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1220 return UndefValue::get(LoadTy);
1222 assert(Res && "failed to materialize?");
1226 static bool isLifetimeStart(const Instruction *Inst) {
1227 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1228 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1232 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
1233 Value *Address, AvailableValue &Res) {
1235 assert((DepInfo.isDef() || DepInfo.isClobber()) &&
1236 "expected a local dependence");
1237 assert(LI->isUnordered() && "rules below are incorrect for ordered access");
1239 const DataLayout &DL = LI->getModule()->getDataLayout();
1241 if (DepInfo.isClobber()) {
1242 // If the dependence is to a store that writes to a superset of the bits
1243 // read by the load, we can extract the bits we need for the load from the
1245 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1246 // Can't forward from non-atomic to atomic without violating memory model.
1247 if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
1249 AnalyzeLoadFromClobberingStore(LI->getType(), Address, DepSI);
1251 Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
1257 // Check to see if we have something like this:
1260 // if we have this, replace the later with an extraction from the former.
1261 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1262 // If this is a clobber and L is the first instruction in its block, then
1263 // we have the first instruction in the entry block.
1264 // Can't forward from non-atomic to atomic without violating memory model.
1265 if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
1267 AnalyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
1270 Res = AvailableValue::getLoad(DepLI, Offset);
1276 // If the clobbering value is a memset/memcpy/memmove, see if we can
1277 // forward a value on from it.
1278 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1279 if (Address && !LI->isAtomic()) {
1280 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1283 Res = AvailableValue::getMI(DepMI, Offset);
1288 // Nothing known about this clobber, have to be conservative
1290 // fast print dep, using operator<< on instruction is too slow.
1291 dbgs() << "GVN: load ";
1292 LI->printAsOperand(dbgs());
1293 Instruction *I = DepInfo.getInst();
1294 dbgs() << " is clobbered by " << *I << '\n';
1298 assert(DepInfo.isDef() && "follows from above");
1300 Instruction *DepInst = DepInfo.getInst();
1302 // Loading the allocation -> undef.
1303 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1304 // Loading immediately after lifetime begin -> undef.
1305 isLifetimeStart(DepInst)) {
1306 Res = AvailableValue::get(UndefValue::get(LI->getType()));
1310 // Loading from calloc (which zero initializes memory) -> zero
1311 if (isCallocLikeFn(DepInst, TLI)) {
1312 Res = AvailableValue::get(Constant::getNullValue(LI->getType()));
1316 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1317 // Reject loads and stores that are to the same address but are of
1318 // different types if we have to. If the stored value is larger or equal to
1319 // the loaded value, we can reuse it.
1320 if (S->getValueOperand()->getType() != LI->getType() &&
1321 !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1325 // Can't forward from non-atomic to atomic without violating memory model.
1326 if (S->isAtomic() < LI->isAtomic())
1329 Res = AvailableValue::get(S->getValueOperand());
1333 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1334 // If the types mismatch and we can't handle it, reject reuse of the load.
1335 // If the stored value is larger or equal to the loaded value, we can reuse
1337 if (LD->getType() != LI->getType() &&
1338 !CanCoerceMustAliasedValueToLoad(LD, LI->getType(), DL))
1341 // Can't forward from non-atomic to atomic without violating memory model.
1342 if (LD->isAtomic() < LI->isAtomic())
1345 Res = AvailableValue::getLoad(LD);
1349 // Unknown def - must be conservative
1351 // fast print dep, using operator<< on instruction is too slow.
1352 dbgs() << "GVN: load ";
1353 LI->printAsOperand(dbgs());
1354 dbgs() << " has unknown def " << *DepInst << '\n';
1359 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1360 AvailValInBlkVect &ValuesPerBlock,
1361 UnavailBlkVect &UnavailableBlocks) {
1363 // Filter out useless results (non-locals, etc). Keep track of the blocks
1364 // where we have a value available in repl, also keep track of whether we see
1365 // dependencies that produce an unknown value for the load (such as a call
1366 // that could potentially clobber the load).
1367 unsigned NumDeps = Deps.size();
1368 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1369 BasicBlock *DepBB = Deps[i].getBB();
1370 MemDepResult DepInfo = Deps[i].getResult();
1372 if (DeadBlocks.count(DepBB)) {
1373 // Dead dependent mem-op disguise as a load evaluating the same value
1374 // as the load in question.
1375 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1379 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1380 UnavailableBlocks.push_back(DepBB);
1384 // The address being loaded in this non-local block may not be the same as
1385 // the pointer operand of the load if PHI translation occurs. Make sure
1386 // to consider the right address.
1387 Value *Address = Deps[i].getAddress();
1390 if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1391 // subtlety: because we know this was a non-local dependency, we know
1392 // it's safe to materialize anywhere between the instruction within
1393 // DepInfo and the end of it's block.
1394 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1397 UnavailableBlocks.push_back(DepBB);
1401 assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1402 "post condition violation");
1405 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1406 UnavailBlkVect &UnavailableBlocks) {
1407 // Okay, we have *some* definitions of the value. This means that the value
1408 // is available in some of our (transitive) predecessors. Lets think about
1409 // doing PRE of this load. This will involve inserting a new load into the
1410 // predecessor when it's not available. We could do this in general, but
1411 // prefer to not increase code size. As such, we only do this when we know
1412 // that we only have to insert *one* load (which means we're basically moving
1413 // the load, not inserting a new one).
1415 SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1416 UnavailableBlocks.end());
1418 // Let's find the first basic block with more than one predecessor. Walk
1419 // backwards through predecessors if needed.
1420 BasicBlock *LoadBB = LI->getParent();
1421 BasicBlock *TmpBB = LoadBB;
1423 while (TmpBB->getSinglePredecessor()) {
1424 TmpBB = TmpBB->getSinglePredecessor();
1425 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1427 if (Blockers.count(TmpBB))
1430 // If any of these blocks has more than one successor (i.e. if the edge we
1431 // just traversed was critical), then there are other paths through this
1432 // block along which the load may not be anticipated. Hoisting the load
1433 // above this block would be adding the load to execution paths along
1434 // which it was not previously executed.
1435 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1442 // Check to see how many predecessors have the loaded value fully
1444 MapVector<BasicBlock *, Value *> PredLoads;
1445 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1446 for (const AvailableValueInBlock &AV : ValuesPerBlock)
1447 FullyAvailableBlocks[AV.BB] = true;
1448 for (BasicBlock *UnavailableBB : UnavailableBlocks)
1449 FullyAvailableBlocks[UnavailableBB] = false;
1451 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1452 for (BasicBlock *Pred : predecessors(LoadBB)) {
1453 // If any predecessor block is an EH pad that does not allow non-PHI
1454 // instructions before the terminator, we can't PRE the load.
1455 if (Pred->getTerminator()->isEHPad()) {
1457 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1458 << Pred->getName() << "': " << *LI << '\n');
1462 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1466 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1467 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1468 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1469 << Pred->getName() << "': " << *LI << '\n');
1473 if (LoadBB->isEHPad()) {
1475 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1476 << Pred->getName() << "': " << *LI << '\n');
1480 CriticalEdgePred.push_back(Pred);
1482 // Only add the predecessors that will not be split for now.
1483 PredLoads[Pred] = nullptr;
1487 // Decide whether PRE is profitable for this load.
1488 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1489 assert(NumUnavailablePreds != 0 &&
1490 "Fully available value should already be eliminated!");
1492 // If this load is unavailable in multiple predecessors, reject it.
1493 // FIXME: If we could restructure the CFG, we could make a common pred with
1494 // all the preds that don't have an available LI and insert a new load into
1496 if (NumUnavailablePreds != 1)
1499 // Split critical edges, and update the unavailable predecessors accordingly.
1500 for (BasicBlock *OrigPred : CriticalEdgePred) {
1501 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1502 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1503 PredLoads[NewPred] = nullptr;
1504 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1505 << LoadBB->getName() << '\n');
1508 // Check if the load can safely be moved to all the unavailable predecessors.
1509 bool CanDoPRE = true;
1510 const DataLayout &DL = LI->getModule()->getDataLayout();
1511 SmallVector<Instruction*, 8> NewInsts;
1512 for (auto &PredLoad : PredLoads) {
1513 BasicBlock *UnavailablePred = PredLoad.first;
1515 // Do PHI translation to get its value in the predecessor if necessary. The
1516 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1518 // If all preds have a single successor, then we know it is safe to insert
1519 // the load on the pred (?!?), so we can insert code to materialize the
1520 // pointer if it is not available.
1521 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1522 Value *LoadPtr = nullptr;
1523 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1526 // If we couldn't find or insert a computation of this phi translated value,
1529 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1530 << *LI->getPointerOperand() << "\n");
1535 PredLoad.second = LoadPtr;
1539 while (!NewInsts.empty()) {
1540 Instruction *I = NewInsts.pop_back_val();
1541 if (MD) MD->removeInstruction(I);
1542 I->eraseFromParent();
1544 // HINT: Don't revert the edge-splitting as following transformation may
1545 // also need to split these critical edges.
1546 return !CriticalEdgePred.empty();
1549 // Okay, we can eliminate this load by inserting a reload in the predecessor
1550 // and using PHI construction to get the value in the other predecessors, do
1552 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1553 DEBUG(if (!NewInsts.empty())
1554 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1555 << *NewInsts.back() << '\n');
1557 // Assign value numbers to the new instructions.
1558 for (Instruction *I : NewInsts) {
1559 // FIXME: We really _ought_ to insert these value numbers into their
1560 // parent's availability map. However, in doing so, we risk getting into
1561 // ordering issues. If a block hasn't been processed yet, we would be
1562 // marking a value as AVAIL-IN, which isn't what we intend.
1566 for (const auto &PredLoad : PredLoads) {
1567 BasicBlock *UnavailablePred = PredLoad.first;
1568 Value *LoadPtr = PredLoad.second;
1570 auto *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre",
1571 LI->isVolatile(), LI->getAlignment(),
1572 LI->getOrdering(), LI->getSynchScope(),
1573 UnavailablePred->getTerminator());
1575 // Transfer the old load's AA tags to the new load.
1577 LI->getAAMetadata(Tags);
1579 NewLoad->setAAMetadata(Tags);
1581 if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1582 NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1583 if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1584 NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1585 if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1586 NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1588 // Transfer DebugLoc.
1589 NewLoad->setDebugLoc(LI->getDebugLoc());
1591 // Add the newly created load.
1592 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1594 MD->invalidateCachedPointerInfo(LoadPtr);
1595 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1598 // Perform PHI construction.
1599 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1600 LI->replaceAllUsesWith(V);
1601 if (isa<PHINode>(V))
1603 if (Instruction *I = dyn_cast<Instruction>(V))
1604 I->setDebugLoc(LI->getDebugLoc());
1605 if (V->getType()->getScalarType()->isPointerTy())
1606 MD->invalidateCachedPointerInfo(V);
1607 markInstructionForDeletion(LI);
1612 /// Attempt to eliminate a load whose dependencies are
1613 /// non-local by performing PHI construction.
1614 bool GVN::processNonLocalLoad(LoadInst *LI) {
1615 // non-local speculations are not allowed under asan.
1616 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeAddress))
1619 // Step 1: Find the non-local dependencies of the load.
1621 MD->getNonLocalPointerDependency(LI, Deps);
1623 // If we had to process more than one hundred blocks to find the
1624 // dependencies, this load isn't worth worrying about. Optimizing
1625 // it will be too expensive.
1626 unsigned NumDeps = Deps.size();
1630 // If we had a phi translation failure, we'll have a single entry which is a
1631 // clobber in the current block. Reject this early.
1633 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1635 dbgs() << "GVN: non-local load ";
1636 LI->printAsOperand(dbgs());
1637 dbgs() << " has unknown dependencies\n";
1642 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1643 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1644 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1645 OE = GEP->idx_end();
1647 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1648 performScalarPRE(I);
1651 // Step 2: Analyze the availability of the load
1652 AvailValInBlkVect ValuesPerBlock;
1653 UnavailBlkVect UnavailableBlocks;
1654 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1656 // If we have no predecessors that produce a known value for this load, exit
1658 if (ValuesPerBlock.empty())
1661 // Step 3: Eliminate fully redundancy.
1663 // If all of the instructions we depend on produce a known value for this
1664 // load, then it is fully redundant and we can use PHI insertion to compute
1665 // its value. Insert PHIs and remove the fully redundant value now.
1666 if (UnavailableBlocks.empty()) {
1667 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1669 // Perform PHI construction.
1670 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1671 LI->replaceAllUsesWith(V);
1673 if (isa<PHINode>(V))
1675 if (Instruction *I = dyn_cast<Instruction>(V))
1676 if (LI->getDebugLoc())
1677 I->setDebugLoc(LI->getDebugLoc());
1678 if (V->getType()->getScalarType()->isPointerTy())
1679 MD->invalidateCachedPointerInfo(V);
1680 markInstructionForDeletion(LI);
1685 // Step 4: Eliminate partial redundancy.
1686 if (!EnablePRE || !EnableLoadPRE)
1689 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1692 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1693 assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1694 "This function can only be called with llvm.assume intrinsic");
1695 Value *V = IntrinsicI->getArgOperand(0);
1697 if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1698 if (Cond->isZero()) {
1699 Type *Int8Ty = Type::getInt8Ty(V->getContext());
1700 // Insert a new store to null instruction before the load to indicate that
1701 // this code is not reachable. FIXME: We could insert unreachable
1702 // instruction directly because we can modify the CFG.
1703 new StoreInst(UndefValue::get(Int8Ty),
1704 Constant::getNullValue(Int8Ty->getPointerTo()),
1707 markInstructionForDeletion(IntrinsicI);
1711 Constant *True = ConstantInt::getTrue(V->getContext());
1712 bool Changed = false;
1714 for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1715 BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1717 // This property is only true in dominated successors, propagateEquality
1718 // will check dominance for us.
1719 Changed |= propagateEquality(V, True, Edge, false);
1722 // We can replace assume value with true, which covers cases like this:
1723 // call void @llvm.assume(i1 %cmp)
1724 // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1725 ReplaceWithConstMap[V] = True;
1727 // If one of *cmp *eq operand is const, adding it to map will cover this:
1728 // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1729 // call void @llvm.assume(i1 %cmp)
1730 // ret float %0 ; will change it to ret float 3.000000e+00
1731 if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1732 if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1733 CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1734 (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1735 CmpI->getFastMathFlags().noNaNs())) {
1736 Value *CmpLHS = CmpI->getOperand(0);
1737 Value *CmpRHS = CmpI->getOperand(1);
1738 if (isa<Constant>(CmpLHS))
1739 std::swap(CmpLHS, CmpRHS);
1740 auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1742 // If only one operand is constant.
1743 if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1744 ReplaceWithConstMap[CmpLHS] = RHSConst;
1750 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1751 auto *ReplInst = dyn_cast<Instruction>(Repl);
1755 // Patch the replacement so that it is not more restrictive than the value
1757 ReplInst->andIRFlags(I);
1759 // FIXME: If both the original and replacement value are part of the
1760 // same control-flow region (meaning that the execution of one
1761 // guarantees the execution of the other), then we can combine the
1762 // noalias scopes here and do better than the general conservative
1763 // answer used in combineMetadata().
1765 // In general, GVN unifies expressions over different control-flow
1766 // regions, and so we need a conservative combination of the noalias
1768 static const unsigned KnownIDs[] = {
1769 LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
1770 LLVMContext::MD_noalias, LLVMContext::MD_range,
1771 LLVMContext::MD_fpmath, LLVMContext::MD_invariant_load,
1772 LLVMContext::MD_invariant_group};
1773 combineMetadata(ReplInst, I, KnownIDs);
1776 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1777 patchReplacementInstruction(I, Repl);
1778 I->replaceAllUsesWith(Repl);
1781 /// Attempt to eliminate a load, first by eliminating it
1782 /// locally, and then attempting non-local elimination if that fails.
1783 bool GVN::processLoad(LoadInst *L) {
1787 // This code hasn't been audited for ordered or volatile memory access
1788 if (!L->isUnordered())
1791 if (L->use_empty()) {
1792 markInstructionForDeletion(L);
1796 // ... to a pointer that has been loaded from before...
1797 MemDepResult Dep = MD->getDependency(L);
1799 // If it is defined in another block, try harder.
1800 if (Dep.isNonLocal())
1801 return processNonLocalLoad(L);
1803 // Only handle the local case below
1804 if (!Dep.isDef() && !Dep.isClobber()) {
1805 // This might be a NonFuncLocal or an Unknown
1807 // fast print dep, using operator<< on instruction is too slow.
1808 dbgs() << "GVN: load ";
1809 L->printAsOperand(dbgs());
1810 dbgs() << " has unknown dependence\n";
1816 if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1817 Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1819 // Replace the load!
1820 patchAndReplaceAllUsesWith(L, AvailableValue);
1821 markInstructionForDeletion(L);
1823 // Tell MDA to rexamine the reused pointer since we might have more
1824 // information after forwarding it.
1825 if (MD && AvailableValue->getType()->getScalarType()->isPointerTy())
1826 MD->invalidateCachedPointerInfo(AvailableValue);
1833 // In order to find a leader for a given value number at a
1834 // specific basic block, we first obtain the list of all Values for that number,
1835 // and then scan the list to find one whose block dominates the block in
1836 // question. This is fast because dominator tree queries consist of only
1837 // a few comparisons of DFS numbers.
1838 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1839 LeaderTableEntry Vals = LeaderTable[num];
1840 if (!Vals.Val) return nullptr;
1842 Value *Val = nullptr;
1843 if (DT->dominates(Vals.BB, BB)) {
1845 if (isa<Constant>(Val)) return Val;
1848 LeaderTableEntry* Next = Vals.Next;
1850 if (DT->dominates(Next->BB, BB)) {
1851 if (isa<Constant>(Next->Val)) return Next->Val;
1852 if (!Val) Val = Next->Val;
1861 /// There is an edge from 'Src' to 'Dst'. Return
1862 /// true if every path from the entry block to 'Dst' passes via this edge. In
1863 /// particular 'Dst' must not be reachable via another edge from 'Src'.
1864 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
1865 DominatorTree *DT) {
1866 // While in theory it is interesting to consider the case in which Dst has
1867 // more than one predecessor, because Dst might be part of a loop which is
1868 // only reachable from Src, in practice it is pointless since at the time
1869 // GVN runs all such loops have preheaders, which means that Dst will have
1870 // been changed to have only one predecessor, namely Src.
1871 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1872 assert((!Pred || Pred == E.getStart()) &&
1873 "No edge between these basic blocks!");
1874 return Pred != nullptr;
1877 // Tries to replace instruction with const, using information from
1878 // ReplaceWithConstMap.
1879 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1880 bool Changed = false;
1881 for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1882 Value *Operand = Instr->getOperand(OpNum);
1883 auto it = ReplaceWithConstMap.find(Operand);
1884 if (it != ReplaceWithConstMap.end()) {
1885 assert(!isa<Constant>(Operand) &&
1886 "Replacing constants with constants is invalid");
1887 DEBUG(dbgs() << "GVN replacing: " << *Operand << " with " << *it->second
1888 << " in instruction " << *Instr << '\n');
1889 Instr->setOperand(OpNum, it->second);
1896 /// The given values are known to be equal in every block
1897 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
1898 /// 'RHS' everywhere in the scope. Returns whether a change was made.
1899 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1900 /// value starting from the end of Root.Start.
1901 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1902 bool DominatesByEdge) {
1903 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
1904 Worklist.push_back(std::make_pair(LHS, RHS));
1905 bool Changed = false;
1906 // For speed, compute a conservative fast approximation to
1907 // DT->dominates(Root, Root.getEnd());
1908 const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1910 while (!Worklist.empty()) {
1911 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1912 LHS = Item.first; RHS = Item.second;
1916 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1918 // Don't try to propagate equalities between constants.
1919 if (isa<Constant>(LHS) && isa<Constant>(RHS))
1922 // Prefer a constant on the right-hand side, or an Argument if no constants.
1923 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1924 std::swap(LHS, RHS);
1925 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1927 // If there is no obvious reason to prefer the left-hand side over the
1928 // right-hand side, ensure the longest lived term is on the right-hand side,
1929 // so the shortest lived term will be replaced by the longest lived.
1930 // This tends to expose more simplifications.
1931 uint32_t LVN = VN.lookupOrAdd(LHS);
1932 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1933 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1934 // Move the 'oldest' value to the right-hand side, using the value number
1935 // as a proxy for age.
1936 uint32_t RVN = VN.lookupOrAdd(RHS);
1938 std::swap(LHS, RHS);
1943 // If value numbering later sees that an instruction in the scope is equal
1944 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
1945 // the invariant that instructions only occur in the leader table for their
1946 // own value number (this is used by removeFromLeaderTable), do not do this
1947 // if RHS is an instruction (if an instruction in the scope is morphed into
1948 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1949 // using the leader table is about compiling faster, not optimizing better).
1950 // The leader table only tracks basic blocks, not edges. Only add to if we
1951 // have the simple case where the edge dominates the end.
1952 if (RootDominatesEnd && !isa<Instruction>(RHS))
1953 addToLeaderTable(LVN, RHS, Root.getEnd());
1955 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
1956 // LHS always has at least one use that is not dominated by Root, this will
1957 // never do anything if LHS has only one use.
1958 if (!LHS->hasOneUse()) {
1959 unsigned NumReplacements =
1961 ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1962 : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1964 Changed |= NumReplacements > 0;
1965 NumGVNEqProp += NumReplacements;
1968 // Now try to deduce additional equalities from this one. For example, if
1969 // the known equality was "(A != B)" == "false" then it follows that A and B
1970 // are equal in the scope. Only boolean equalities with an explicit true or
1971 // false RHS are currently supported.
1972 if (!RHS->getType()->isIntegerTy(1))
1973 // Not a boolean equality - bail out.
1975 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1977 // RHS neither 'true' nor 'false' - bail out.
1979 // Whether RHS equals 'true'. Otherwise it equals 'false'.
1980 bool isKnownTrue = CI->isAllOnesValue();
1981 bool isKnownFalse = !isKnownTrue;
1983 // If "A && B" is known true then both A and B are known true. If "A || B"
1984 // is known false then both A and B are known false.
1986 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1987 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1988 Worklist.push_back(std::make_pair(A, RHS));
1989 Worklist.push_back(std::make_pair(B, RHS));
1993 // If we are propagating an equality like "(A == B)" == "true" then also
1994 // propagate the equality A == B. When propagating a comparison such as
1995 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1996 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1997 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1999 // If "A == B" is known true, or "A != B" is known false, then replace
2000 // A with B everywhere in the scope.
2001 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2002 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2003 Worklist.push_back(std::make_pair(Op0, Op1));
2005 // Handle the floating point versions of equality comparisons too.
2006 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
2007 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
2009 // Floating point -0.0 and 0.0 compare equal, so we can only
2010 // propagate values if we know that we have a constant and that
2011 // its value is non-zero.
2013 // FIXME: We should do this optimization if 'no signed zeros' is
2014 // applicable via an instruction-level fast-math-flag or some other
2015 // indicator that relaxed FP semantics are being used.
2017 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
2018 Worklist.push_back(std::make_pair(Op0, Op1));
2021 // If "A >= B" is known true, replace "A < B" with false everywhere.
2022 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2023 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2024 // Since we don't have the instruction "A < B" immediately to hand, work
2025 // out the value number that it would have and use that to find an
2026 // appropriate instruction (if any).
2027 uint32_t NextNum = VN.getNextUnusedValueNumber();
2028 uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2029 // If the number we were assigned was brand new then there is no point in
2030 // looking for an instruction realizing it: there cannot be one!
2031 if (Num < NextNum) {
2032 Value *NotCmp = findLeader(Root.getEnd(), Num);
2033 if (NotCmp && isa<Instruction>(NotCmp)) {
2034 unsigned NumReplacements =
2036 ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
2037 : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
2039 Changed |= NumReplacements > 0;
2040 NumGVNEqProp += NumReplacements;
2043 // Ensure that any instruction in scope that gets the "A < B" value number
2044 // is replaced with false.
2045 // The leader table only tracks basic blocks, not edges. Only add to if we
2046 // have the simple case where the edge dominates the end.
2047 if (RootDominatesEnd)
2048 addToLeaderTable(Num, NotVal, Root.getEnd());
2057 /// When calculating availability, handle an instruction
2058 /// by inserting it into the appropriate sets
2059 bool GVN::processInstruction(Instruction *I) {
2060 // Ignore dbg info intrinsics.
2061 if (isa<DbgInfoIntrinsic>(I))
2064 // If the instruction can be easily simplified then do so now in preference
2065 // to value numbering it. Value numbering often exposes redundancies, for
2066 // example if it determines that %y is equal to %x then the instruction
2067 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2068 const DataLayout &DL = I->getModule()->getDataLayout();
2069 if (Value *V = SimplifyInstruction(I, DL, TLI, DT, AC)) {
2070 bool Changed = false;
2071 if (!I->use_empty()) {
2072 I->replaceAllUsesWith(V);
2075 if (isInstructionTriviallyDead(I, TLI)) {
2076 markInstructionForDeletion(I);
2080 if (MD && V->getType()->getScalarType()->isPointerTy())
2081 MD->invalidateCachedPointerInfo(V);
2087 if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
2088 if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
2089 return processAssumeIntrinsic(IntrinsicI);
2091 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2092 if (processLoad(LI))
2095 unsigned Num = VN.lookupOrAdd(LI);
2096 addToLeaderTable(Num, LI, LI->getParent());
2100 // For conditional branches, we can perform simple conditional propagation on
2101 // the condition value itself.
2102 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2103 if (!BI->isConditional())
2106 if (isa<Constant>(BI->getCondition()))
2107 return processFoldableCondBr(BI);
2109 Value *BranchCond = BI->getCondition();
2110 BasicBlock *TrueSucc = BI->getSuccessor(0);
2111 BasicBlock *FalseSucc = BI->getSuccessor(1);
2112 // Avoid multiple edges early.
2113 if (TrueSucc == FalseSucc)
2116 BasicBlock *Parent = BI->getParent();
2117 bool Changed = false;
2119 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2120 BasicBlockEdge TrueE(Parent, TrueSucc);
2121 Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
2123 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2124 BasicBlockEdge FalseE(Parent, FalseSucc);
2125 Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
2130 // For switches, propagate the case values into the case destinations.
2131 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2132 Value *SwitchCond = SI->getCondition();
2133 BasicBlock *Parent = SI->getParent();
2134 bool Changed = false;
2136 // Remember how many outgoing edges there are to every successor.
2137 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2138 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2139 ++SwitchEdges[SI->getSuccessor(i)];
2141 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2143 BasicBlock *Dst = i.getCaseSuccessor();
2144 // If there is only a single edge, propagate the case value into it.
2145 if (SwitchEdges.lookup(Dst) == 1) {
2146 BasicBlockEdge E(Parent, Dst);
2147 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E, true);
2153 // Instructions with void type don't return a value, so there's
2154 // no point in trying to find redundancies in them.
2155 if (I->getType()->isVoidTy())
2158 uint32_t NextNum = VN.getNextUnusedValueNumber();
2159 unsigned Num = VN.lookupOrAdd(I);
2161 // Allocations are always uniquely numbered, so we can save time and memory
2162 // by fast failing them.
2163 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2164 addToLeaderTable(Num, I, I->getParent());
2168 // If the number we were assigned was a brand new VN, then we don't
2169 // need to do a lookup to see if the number already exists
2170 // somewhere in the domtree: it can't!
2171 if (Num >= NextNum) {
2172 addToLeaderTable(Num, I, I->getParent());
2176 // Perform fast-path value-number based elimination of values inherited from
2178 Value *Repl = findLeader(I->getParent(), Num);
2180 // Failure, just remember this instance for future use.
2181 addToLeaderTable(Num, I, I->getParent());
2183 } else if (Repl == I) {
2184 // If I was the result of a shortcut PRE, it might already be in the table
2185 // and the best replacement for itself. Nothing to do.
2190 patchAndReplaceAllUsesWith(I, Repl);
2191 if (MD && Repl->getType()->getScalarType()->isPointerTy())
2192 MD->invalidateCachedPointerInfo(Repl);
2193 markInstructionForDeletion(I);
2197 /// runOnFunction - This is the main transformation entry point for a function.
2198 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
2199 const TargetLibraryInfo &RunTLI, AAResults &RunAA,
2200 MemoryDependenceResults *RunMD) {
2205 VN.setAliasAnalysis(&RunAA);
2209 bool Changed = false;
2210 bool ShouldContinue = true;
2212 // Merge unconditional branches, allowing PRE to catch more
2213 // optimization opportunities.
2214 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2215 BasicBlock *BB = &*FI++;
2218 MergeBlockIntoPredecessor(BB, DT, /* LoopInfo */ nullptr, MD);
2219 if (removedBlock) ++NumGVNBlocks;
2221 Changed |= removedBlock;
2224 unsigned Iteration = 0;
2225 while (ShouldContinue) {
2226 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2227 ShouldContinue = iterateOnFunction(F);
2228 Changed |= ShouldContinue;
2233 // Fabricate val-num for dead-code in order to suppress assertion in
2235 assignValNumForDeadCode();
2236 bool PREChanged = true;
2237 while (PREChanged) {
2238 PREChanged = performPRE(F);
2239 Changed |= PREChanged;
2243 // FIXME: Should perform GVN again after PRE does something. PRE can move
2244 // computations into blocks where they become fully redundant. Note that
2245 // we can't do this until PRE's critical edge splitting updates memdep.
2246 // Actually, when this happens, we should just fully integrate PRE into GVN.
2248 cleanupGlobalSets();
2249 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2256 bool GVN::processBlock(BasicBlock *BB) {
2257 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2258 // (and incrementing BI before processing an instruction).
2259 assert(InstrsToErase.empty() &&
2260 "We expect InstrsToErase to be empty across iterations");
2261 if (DeadBlocks.count(BB))
2264 // Clearing map before every BB because it can be used only for single BB.
2265 ReplaceWithConstMap.clear();
2266 bool ChangedFunction = false;
2268 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2270 if (!ReplaceWithConstMap.empty())
2271 ChangedFunction |= replaceOperandsWithConsts(&*BI);
2272 ChangedFunction |= processInstruction(&*BI);
2274 if (InstrsToErase.empty()) {
2279 // If we need some instructions deleted, do it now.
2280 NumGVNInstr += InstrsToErase.size();
2282 // Avoid iterator invalidation.
2283 bool AtStart = BI == BB->begin();
2287 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2288 E = InstrsToErase.end(); I != E; ++I) {
2289 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2290 if (MD) MD->removeInstruction(*I);
2291 DEBUG(verifyRemoved(*I));
2292 (*I)->eraseFromParent();
2294 InstrsToErase.clear();
2302 return ChangedFunction;
2305 // Instantiate an expression in a predecessor that lacked it.
2306 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2307 unsigned int ValNo) {
2308 // Because we are going top-down through the block, all value numbers
2309 // will be available in the predecessor by the time we need them. Any
2310 // that weren't originally present will have been instantiated earlier
2312 bool success = true;
2313 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2314 Value *Op = Instr->getOperand(i);
2315 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2317 // This could be a newly inserted instruction, in which case, we won't
2318 // find a value number, and should give up before we hurt ourselves.
2319 // FIXME: Rewrite the infrastructure to let it easier to value number
2320 // and process newly inserted instructions.
2321 if (!VN.exists(Op)) {
2325 if (Value *V = findLeader(Pred, VN.lookup(Op))) {
2326 Instr->setOperand(i, V);
2333 // Fail out if we encounter an operand that is not available in
2334 // the PRE predecessor. This is typically because of loads which
2335 // are not value numbered precisely.
2339 Instr->insertBefore(Pred->getTerminator());
2340 Instr->setName(Instr->getName() + ".pre");
2341 Instr->setDebugLoc(Instr->getDebugLoc());
2342 VN.add(Instr, ValNo);
2344 // Update the availability map to include the new instruction.
2345 addToLeaderTable(ValNo, Instr, Pred);
2349 bool GVN::performScalarPRE(Instruction *CurInst) {
2350 if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2351 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2352 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2353 isa<DbgInfoIntrinsic>(CurInst))
2356 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2357 // sinking the compare again, and it would force the code generator to
2358 // move the i1 from processor flags or predicate registers into a general
2359 // purpose register.
2360 if (isa<CmpInst>(CurInst))
2363 // We don't currently value number ANY inline asm calls.
2364 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2365 if (CallI->isInlineAsm())
2368 uint32_t ValNo = VN.lookup(CurInst);
2370 // Look for the predecessors for PRE opportunities. We're
2371 // only trying to solve the basic diamond case, where
2372 // a value is computed in the successor and one predecessor,
2373 // but not the other. We also explicitly disallow cases
2374 // where the successor is its own predecessor, because they're
2375 // more complicated to get right.
2376 unsigned NumWith = 0;
2377 unsigned NumWithout = 0;
2378 BasicBlock *PREPred = nullptr;
2379 BasicBlock *CurrentBlock = CurInst->getParent();
2381 SmallVector<std::pair<Value *, BasicBlock *>, 8> predMap;
2382 for (BasicBlock *P : predecessors(CurrentBlock)) {
2383 // We're not interested in PRE where the block is its
2384 // own predecessor, or in blocks with predecessors
2385 // that are not reachable.
2386 if (P == CurrentBlock) {
2389 } else if (!DT->isReachableFromEntry(P)) {
2394 Value *predV = findLeader(P, ValNo);
2396 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2399 } else if (predV == CurInst) {
2400 /* CurInst dominates this predecessor. */
2404 predMap.push_back(std::make_pair(predV, P));
2409 // Don't do PRE when it might increase code size, i.e. when
2410 // we would need to insert instructions in more than one pred.
2411 if (NumWithout > 1 || NumWith == 0)
2414 // We may have a case where all predecessors have the instruction,
2415 // and we just need to insert a phi node. Otherwise, perform
2417 Instruction *PREInstr = nullptr;
2419 if (NumWithout != 0) {
2420 // Don't do PRE across indirect branch.
2421 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2424 // We can't do PRE safely on a critical edge, so instead we schedule
2425 // the edge to be split and perform the PRE the next time we iterate
2427 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2428 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2429 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2432 // We need to insert somewhere, so let's give it a shot
2433 PREInstr = CurInst->clone();
2434 if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
2435 // If we failed insertion, make sure we remove the instruction.
2436 DEBUG(verifyRemoved(PREInstr));
2442 // Either we should have filled in the PRE instruction, or we should
2443 // not have needed insertions.
2444 assert (PREInstr != nullptr || NumWithout == 0);
2448 // Create a PHI to make the value available in this block.
2450 PHINode::Create(CurInst->getType(), predMap.size(),
2451 CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2452 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2453 if (Value *V = predMap[i].first)
2454 Phi->addIncoming(V, predMap[i].second);
2456 Phi->addIncoming(PREInstr, PREPred);
2460 addToLeaderTable(ValNo, Phi, CurrentBlock);
2461 Phi->setDebugLoc(CurInst->getDebugLoc());
2462 CurInst->replaceAllUsesWith(Phi);
2463 if (MD && Phi->getType()->getScalarType()->isPointerTy())
2464 MD->invalidateCachedPointerInfo(Phi);
2466 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2468 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2470 MD->removeInstruction(CurInst);
2471 DEBUG(verifyRemoved(CurInst));
2472 CurInst->eraseFromParent();
2478 /// Perform a purely local form of PRE that looks for diamond
2479 /// control flow patterns and attempts to perform simple PRE at the join point.
2480 bool GVN::performPRE(Function &F) {
2481 bool Changed = false;
2482 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2483 // Nothing to PRE in the entry block.
2484 if (CurrentBlock == &F.getEntryBlock())
2487 // Don't perform PRE on an EH pad.
2488 if (CurrentBlock->isEHPad())
2491 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2492 BE = CurrentBlock->end();
2494 Instruction *CurInst = &*BI++;
2495 Changed |= performScalarPRE(CurInst);
2499 if (splitCriticalEdges())
2505 /// Split the critical edge connecting the given two blocks, and return
2506 /// the block inserted to the critical edge.
2507 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2509 SplitCriticalEdge(Pred, Succ, CriticalEdgeSplittingOptions(DT));
2511 MD->invalidateCachedPredecessors();
2515 /// Split critical edges found during the previous
2516 /// iteration that may enable further optimization.
2517 bool GVN::splitCriticalEdges() {
2518 if (toSplit.empty())
2521 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2522 SplitCriticalEdge(Edge.first, Edge.second,
2523 CriticalEdgeSplittingOptions(DT));
2524 } while (!toSplit.empty());
2525 if (MD) MD->invalidateCachedPredecessors();
2529 /// Executes one iteration of GVN
2530 bool GVN::iterateOnFunction(Function &F) {
2531 cleanupGlobalSets();
2533 // Top-down walk of the dominator tree
2534 bool Changed = false;
2535 // Save the blocks this function have before transformation begins. GVN may
2536 // split critical edge, and hence may invalidate the RPO/DT iterator.
2538 std::vector<BasicBlock *> BBVect;
2539 BBVect.reserve(256);
2540 // Needed for value numbering with phi construction to work.
2541 ReversePostOrderTraversal<Function *> RPOT(&F);
2542 for (ReversePostOrderTraversal<Function *>::rpo_iterator RI = RPOT.begin(),
2545 BBVect.push_back(*RI);
2547 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2549 Changed |= processBlock(*I);
2554 void GVN::cleanupGlobalSets() {
2556 LeaderTable.clear();
2557 TableAllocator.Reset();
2560 /// Verify that the specified instruction does not occur in our
2561 /// internal data structures.
2562 void GVN::verifyRemoved(const Instruction *Inst) const {
2563 VN.verifyRemoved(Inst);
2565 // Walk through the value number scope to make sure the instruction isn't
2566 // ferreted away in it.
2567 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2568 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2569 const LeaderTableEntry *Node = &I->second;
2570 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2572 while (Node->Next) {
2574 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2579 /// BB is declared dead, which implied other blocks become dead as well. This
2580 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2581 /// live successors, update their phi nodes by replacing the operands
2582 /// corresponding to dead blocks with UndefVal.
2583 void GVN::addDeadBlock(BasicBlock *BB) {
2584 SmallVector<BasicBlock *, 4> NewDead;
2585 SmallSetVector<BasicBlock *, 4> DF;
2587 NewDead.push_back(BB);
2588 while (!NewDead.empty()) {
2589 BasicBlock *D = NewDead.pop_back_val();
2590 if (DeadBlocks.count(D))
2593 // All blocks dominated by D are dead.
2594 SmallVector<BasicBlock *, 8> Dom;
2595 DT->getDescendants(D, Dom);
2596 DeadBlocks.insert(Dom.begin(), Dom.end());
2598 // Figure out the dominance-frontier(D).
2599 for (BasicBlock *B : Dom) {
2600 for (BasicBlock *S : successors(B)) {
2601 if (DeadBlocks.count(S))
2604 bool AllPredDead = true;
2605 for (BasicBlock *P : predecessors(S))
2606 if (!DeadBlocks.count(P)) {
2607 AllPredDead = false;
2612 // S could be proved dead later on. That is why we don't update phi
2613 // operands at this moment.
2616 // While S is not dominated by D, it is dead by now. This could take
2617 // place if S already have a dead predecessor before D is declared
2619 NewDead.push_back(S);
2625 // For the dead blocks' live successors, update their phi nodes by replacing
2626 // the operands corresponding to dead blocks with UndefVal.
2627 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2630 if (DeadBlocks.count(B))
2633 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2634 for (BasicBlock *P : Preds) {
2635 if (!DeadBlocks.count(P))
2638 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2639 if (BasicBlock *S = splitCriticalEdges(P, B))
2640 DeadBlocks.insert(P = S);
2643 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2644 PHINode &Phi = cast<PHINode>(*II);
2645 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2646 UndefValue::get(Phi.getType()));
2652 // If the given branch is recognized as a foldable branch (i.e. conditional
2653 // branch with constant condition), it will perform following analyses and
2655 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2656 // R be the target of the dead out-coming edge.
2657 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2658 // edge. The result of this step will be {X| X is dominated by R}
2659 // 2) Identify those blocks which haves at least one dead predecessor. The
2660 // result of this step will be dominance-frontier(R).
2661 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2662 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2664 // Return true iff *NEW* dead code are found.
2665 bool GVN::processFoldableCondBr(BranchInst *BI) {
2666 if (!BI || BI->isUnconditional())
2669 // If a branch has two identical successors, we cannot declare either dead.
2670 if (BI->getSuccessor(0) == BI->getSuccessor(1))
2673 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2677 BasicBlock *DeadRoot =
2678 Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2679 if (DeadBlocks.count(DeadRoot))
2682 if (!DeadRoot->getSinglePredecessor())
2683 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2685 addDeadBlock(DeadRoot);
2689 // performPRE() will trigger assert if it comes across an instruction without
2690 // associated val-num. As it normally has far more live instructions than dead
2691 // instructions, it makes more sense just to "fabricate" a val-number for the
2692 // dead code than checking if instruction involved is dead or not.
2693 void GVN::assignValNumForDeadCode() {
2694 for (BasicBlock *BB : DeadBlocks) {
2695 for (Instruction &Inst : *BB) {
2696 unsigned ValNum = VN.lookupOrAdd(&Inst);
2697 addToLeaderTable(ValNum, &Inst, BB);
2702 class llvm::gvn::GVNLegacyPass : public FunctionPass {
2704 static char ID; // Pass identification, replacement for typeid
2705 explicit GVNLegacyPass(bool NoLoads = false)
2706 : FunctionPass(ID), NoLoads(NoLoads) {
2707 initializeGVNLegacyPassPass(*PassRegistry::getPassRegistry());
2710 bool runOnFunction(Function &F) override {
2711 if (skipFunction(F))
2714 return Impl.runImpl(
2715 F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2716 getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2717 getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2718 getAnalysis<AAResultsWrapperPass>().getAAResults(),
2720 : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep());
2723 void getAnalysisUsage(AnalysisUsage &AU) const override {
2724 AU.addRequired<AssumptionCacheTracker>();
2725 AU.addRequired<DominatorTreeWrapperPass>();
2726 AU.addRequired<TargetLibraryInfoWrapperPass>();
2728 AU.addRequired<MemoryDependenceWrapperPass>();
2729 AU.addRequired<AAResultsWrapperPass>();
2731 AU.addPreserved<DominatorTreeWrapperPass>();
2732 AU.addPreserved<GlobalsAAWrapperPass>();
2740 char GVNLegacyPass::ID = 0;
2742 // The public interface to this file...
2743 FunctionPass *llvm::createGVNPass(bool NoLoads) {
2744 return new GVNLegacyPass(NoLoads);
2747 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2748 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
2749 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
2750 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2751 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
2752 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2753 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
2754 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)