2297 lines · cpp
1//===- LazyValueInfo.cpp - Value constraint analysis ------------*- C++ -*-===//2//3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.4// See https://llvm.org/LICENSE.txt for license information.5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception6//7//===----------------------------------------------------------------------===//8//9// This file defines the interface for lazy computation of value constraint10// information.11//12//===----------------------------------------------------------------------===//13 14#include "llvm/Analysis/LazyValueInfo.h"15#include "llvm/ADT/DenseSet.h"16#include "llvm/ADT/STLExtras.h"17#include "llvm/Analysis/AssumptionCache.h"18#include "llvm/Analysis/ConstantFolding.h"19#include "llvm/Analysis/InstructionSimplify.h"20#include "llvm/Analysis/Passes.h"21#include "llvm/Analysis/TargetLibraryInfo.h"22#include "llvm/Analysis/ValueLattice.h"23#include "llvm/Analysis/ValueTracking.h"24#include "llvm/IR/AssemblyAnnotationWriter.h"25#include "llvm/IR/CFG.h"26#include "llvm/IR/ConstantRange.h"27#include "llvm/IR/Constants.h"28#include "llvm/IR/DataLayout.h"29#include "llvm/IR/Dominators.h"30#include "llvm/IR/InstrTypes.h"31#include "llvm/IR/Instructions.h"32#include "llvm/IR/IntrinsicInst.h"33#include "llvm/IR/Intrinsics.h"34#include "llvm/IR/LLVMContext.h"35#include "llvm/IR/Module.h"36#include "llvm/IR/PatternMatch.h"37#include "llvm/IR/ValueHandle.h"38#include "llvm/InitializePasses.h"39#include "llvm/Support/Debug.h"40#include "llvm/Support/FormattedStream.h"41#include "llvm/Support/KnownBits.h"42#include "llvm/Support/raw_ostream.h"43#include <optional>44using namespace llvm;45using namespace PatternMatch;46 47#define DEBUG_TYPE "lazy-value-info"48 49// This is the number of worklist items we will process to try to discover an50// answer for a given value.51static const unsigned MaxProcessedPerValue = 500;52 53char LazyValueInfoWrapperPass::ID = 0;54LazyValueInfoWrapperPass::LazyValueInfoWrapperPass() : FunctionPass(ID) {}55INITIALIZE_PASS_BEGIN(LazyValueInfoWrapperPass, "lazy-value-info",56 "Lazy Value Information Analysis", false, true)57INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)58INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)59INITIALIZE_PASS_END(LazyValueInfoWrapperPass, "lazy-value-info",60 "Lazy Value Information Analysis", false, true)61 62static cl::opt<bool> PerPredRanges(63 "lvi-per-pred-ranges", cl::Hidden, cl::init(false),64 cl::desc("Enable tracking of ranges for a value in a block for"65 "each block predecessor (default = false)"));66 67namespace llvm {68FunctionPass *createLazyValueInfoPass() {69 return new LazyValueInfoWrapperPass();70}71} // namespace llvm72 73AnalysisKey LazyValueAnalysis::Key;74 75/// Returns true if this lattice value represents at most one possible value.76/// This is as precise as any lattice value can get while still representing77/// reachable code.78static bool hasSingleValue(const ValueLatticeElement &Val) {79 if (Val.isConstantRange() &&80 Val.getConstantRange().isSingleElement())81 // Integer constants are single element ranges82 return true;83 if (Val.isConstant())84 // Non integer constants85 return true;86 return false;87}88 89//===----------------------------------------------------------------------===//90// LazyValueInfoCache Decl91//===----------------------------------------------------------------------===//92 93namespace {94 /// A callback value handle updates the cache when values are erased.95 class LazyValueInfoCache;96 struct LVIValueHandle final : public CallbackVH {97 LazyValueInfoCache *Parent;98 99 LVIValueHandle(Value *V, LazyValueInfoCache *P = nullptr)100 : CallbackVH(V), Parent(P) { }101 102 void deleted() override;103 void allUsesReplacedWith(Value *V) override {104 deleted();105 }106 };107} // end anonymous namespace108 109namespace {110using NonNullPointerSet = SmallDenseSet<AssertingVH<Value>, 2>;111using BBLatticeElementMap =112 SmallDenseMap<PoisoningVH<BasicBlock>, ValueLatticeElement, 4>;113using PredecessorValueLatticeMap =114 SmallDenseMap<AssertingVH<Value>, BBLatticeElementMap, 2>;115 116/// This is the cache kept by LazyValueInfo which117/// maintains information about queries across the clients' queries.118class LazyValueInfoCache {119 /// This is all of the cached information for one basic block. It contains120 /// the per-value lattice elements, as well as a separate set for121 /// overdefined values to reduce memory usage. Additionally pointers122 /// dereferenced in the block are cached for nullability queries.123 struct BlockCacheEntry {124 SmallDenseMap<AssertingVH<Value>, ValueLatticeElement, 4> LatticeElements;125 SmallDenseSet<AssertingVH<Value>, 4> OverDefined;126 // std::nullopt indicates that the nonnull pointers for this basic block127 // block have not been computed yet.128 std::optional<NonNullPointerSet> NonNullPointers;129 // This is an extension of the above LatticeElements, caching, for each130 // Value, a ValueLatticeElement, for each predecessor of the BB tracked by131 // this entry.132 std::optional<PredecessorValueLatticeMap> PredecessorLatticeElements;133 };134 135 /// Cached information per basic block.136 DenseMap<PoisoningVH<BasicBlock>, std::unique_ptr<BlockCacheEntry>>137 BlockCache;138 /// Set of value handles used to erase values from the cache on deletion.139 DenseSet<LVIValueHandle, DenseMapInfo<Value *>> ValueHandles;140 141 const BlockCacheEntry *getBlockEntry(BasicBlock *BB) const {142 auto It = BlockCache.find_as(BB);143 if (It == BlockCache.end())144 return nullptr;145 return It->second.get();146 }147 148 BlockCacheEntry *getOrCreateBlockEntry(BasicBlock *BB) {149 auto It = BlockCache.find_as(BB);150 if (It == BlockCache.end()) {151 std::unique_ptr<BlockCacheEntry> BCE =152 std::make_unique<BlockCacheEntry>();153 if (PerPredRanges)154 BCE->PredecessorLatticeElements =155 std::make_optional<PredecessorValueLatticeMap>();156 It = BlockCache.insert({BB, std::move(BCE)}).first;157 }158 159 return It->second.get();160 }161 162 void addValueHandle(Value *Val) {163 auto HandleIt = ValueHandles.find_as(Val);164 if (HandleIt == ValueHandles.end())165 ValueHandles.insert({Val, this});166 }167 168public:169 void insertResult(Value *Val, BasicBlock *BB,170 const ValueLatticeElement &Result) {171 BlockCacheEntry *Entry = getOrCreateBlockEntry(BB);172 173 // Insert over-defined values into their own cache to reduce memory174 // overhead.175 if (Result.isOverdefined())176 Entry->OverDefined.insert(Val);177 else178 Entry->LatticeElements.insert({Val, Result});179 180 addValueHandle(Val);181 }182 183 void insertPredecessorResults(Value *Val, BasicBlock *BB,184 BBLatticeElementMap &PredLatticeElements) {185 BlockCacheEntry *Entry = getOrCreateBlockEntry(BB);186 187 Entry->PredecessorLatticeElements->insert({Val, PredLatticeElements});188 189 addValueHandle(Val);190 }191 192 std::optional<BBLatticeElementMap>193 getCachedPredecessorInfo(Value *V, BasicBlock *BB) const {194 const BlockCacheEntry *Entry = getBlockEntry(BB);195 if (!Entry)196 return std::nullopt;197 198 auto LatticeIt = Entry->PredecessorLatticeElements->find_as(V);199 if (LatticeIt == Entry->PredecessorLatticeElements->end())200 return std::nullopt;201 202 return LatticeIt->second;203 }204 205 std::optional<ValueLatticeElement> getCachedValueInfo(Value *V,206 BasicBlock *BB) const {207 const BlockCacheEntry *Entry = getBlockEntry(BB);208 if (!Entry)209 return std::nullopt;210 211 if (Entry->OverDefined.count(V))212 return ValueLatticeElement::getOverdefined();213 214 auto LatticeIt = Entry->LatticeElements.find_as(V);215 if (LatticeIt == Entry->LatticeElements.end())216 return std::nullopt;217 218 return LatticeIt->second;219 }220 221 bool222 isNonNullAtEndOfBlock(Value *V, BasicBlock *BB,223 function_ref<NonNullPointerSet(BasicBlock *)> InitFn) {224 BlockCacheEntry *Entry = getOrCreateBlockEntry(BB);225 if (!Entry->NonNullPointers) {226 Entry->NonNullPointers = InitFn(BB);227 for (Value *V : *Entry->NonNullPointers)228 addValueHandle(V);229 }230 231 return Entry->NonNullPointers->count(V);232 }233 234 /// clear - Empty the cache.235 void clear() {236 BlockCache.clear();237 ValueHandles.clear();238 }239 240 /// Inform the cache that a given value has been deleted.241 void eraseValue(Value *V);242 243 /// This is part of the update interface to inform the cache244 /// that a block has been deleted.245 void eraseBlock(BasicBlock *BB);246 247 /// Updates the cache to remove any influence an overdefined value in248 /// OldSucc might have (unless also overdefined in NewSucc). This just249 /// flushes elements from the cache and does not add any.250 void threadEdgeImpl(BasicBlock *OldSucc, BasicBlock *NewSucc);251};252} // namespace253 254void LazyValueInfoCache::eraseValue(Value *V) {255 for (auto &Pair : BlockCache) {256 Pair.second->LatticeElements.erase(V);257 Pair.second->OverDefined.erase(V);258 if (Pair.second->NonNullPointers)259 Pair.second->NonNullPointers->erase(V);260 if (PerPredRanges)261 Pair.second->PredecessorLatticeElements->erase(V);262 }263 264 auto HandleIt = ValueHandles.find_as(V);265 if (HandleIt != ValueHandles.end())266 ValueHandles.erase(HandleIt);267}268 269void LVIValueHandle::deleted() {270 // This erasure deallocates *this, so it MUST happen after we're done271 // using any and all members of *this.272 Parent->eraseValue(*this);273}274 275void LazyValueInfoCache::eraseBlock(BasicBlock *BB) {276 // Clear all when a BB is removed.277 if (PerPredRanges)278 for (auto &Pair : BlockCache)279 Pair.second->PredecessorLatticeElements->clear();280 BlockCache.erase(BB);281}282 283void LazyValueInfoCache::threadEdgeImpl(BasicBlock *OldSucc,284 BasicBlock *NewSucc) {285 // When an edge in the graph has been threaded, values that we could not286 // determine a value for before (i.e. were marked overdefined) may be287 // possible to solve now. We do NOT try to proactively update these values.288 // Instead, we clear their entries from the cache, and allow lazy updating to289 // recompute them when needed.290 291 // The updating process is fairly simple: we need to drop cached info292 // for all values that were marked overdefined in OldSucc, and for those same293 // values in any successor of OldSucc (except NewSucc) in which they were294 // also marked overdefined.295 std::vector<BasicBlock*> worklist;296 worklist.push_back(OldSucc);297 298 const BlockCacheEntry *Entry = getBlockEntry(OldSucc);299 if (!Entry || Entry->OverDefined.empty())300 return; // Nothing to process here.301 SmallVector<Value *, 4> ValsToClear(Entry->OverDefined.begin(),302 Entry->OverDefined.end());303 304 // Use a worklist to perform a depth-first search of OldSucc's successors.305 // NOTE: We do not need a visited list since any blocks we have already306 // visited will have had their overdefined markers cleared already, and we307 // thus won't loop to their successors.308 while (!worklist.empty()) {309 BasicBlock *ToUpdate = worklist.back();310 worklist.pop_back();311 312 // Skip blocks only accessible through NewSucc.313 if (ToUpdate == NewSucc) continue;314 315 // If a value was marked overdefined in OldSucc, and is here too...316 auto OI = BlockCache.find_as(ToUpdate);317 if (OI == BlockCache.end() || OI->second->OverDefined.empty())318 continue;319 auto &ValueSet = OI->second->OverDefined;320 321 bool changed = false;322 for (Value *V : ValsToClear) {323 if (!ValueSet.erase(V))324 continue;325 326 // If we removed anything, then we potentially need to update327 // blocks successors too.328 changed = true;329 }330 331 if (!changed) continue;332 333 llvm::append_range(worklist, successors(ToUpdate));334 }335}336 337namespace llvm {338namespace {339/// An assembly annotator class to print LazyValueCache information in340/// comments.341class LazyValueInfoAnnotatedWriter : public AssemblyAnnotationWriter {342 LazyValueInfoImpl *LVIImpl;343 // While analyzing which blocks we can solve values for, we need the dominator344 // information.345 DominatorTree &DT;346 347public:348 LazyValueInfoAnnotatedWriter(LazyValueInfoImpl *L, DominatorTree &DTree)349 : LVIImpl(L), DT(DTree) {}350 351 void emitBasicBlockStartAnnot(const BasicBlock *BB,352 formatted_raw_ostream &OS) override;353 354 void emitInstructionAnnot(const Instruction *I,355 formatted_raw_ostream &OS) override;356};357} // namespace358// The actual implementation of the lazy analysis and update.359class LazyValueInfoImpl {360 361 /// Cached results from previous queries362 LazyValueInfoCache TheCache;363 364 /// This stack holds the state of the value solver during a query.365 /// It basically emulates the callstack of the naive366 /// recursive value lookup process.367 SmallVector<std::pair<BasicBlock*, Value*>, 8> BlockValueStack;368 369 /// Keeps track of which block-value pairs are in BlockValueStack.370 DenseSet<std::pair<BasicBlock*, Value*> > BlockValueSet;371 372 /// Push BV onto BlockValueStack unless it's already in there.373 /// Returns true on success.374 bool pushBlockValue(const std::pair<BasicBlock *, Value *> &BV) {375 if (!BlockValueSet.insert(BV).second)376 return false; // It's already in the stack.377 378 LLVM_DEBUG(dbgs() << "PUSH: " << *BV.second << " in "379 << BV.first->getName() << "\n");380 BlockValueStack.push_back(BV);381 return true;382 }383 384 AssumptionCache *AC; ///< A pointer to the cache of @llvm.assume calls.385 const DataLayout &DL; ///< A mandatory DataLayout386 387 /// Declaration of the llvm.experimental.guard() intrinsic,388 /// if it exists in the module.389 Function *GuardDecl;390 391 std::optional<ValueLatticeElement> getBlockValue(Value *Val, BasicBlock *BB,392 Instruction *CxtI);393 std::optional<ValueLatticeElement> getEdgeValue(Value *V, BasicBlock *F,394 BasicBlock *T,395 Instruction *CxtI = nullptr);396 397 // These methods process one work item and may add more. A false value398 // returned means that the work item was not completely processed and must399 // be revisited after going through the new items.400 bool solveBlockValue(Value *Val, BasicBlock *BB);401 std::optional<ValueLatticeElement> solveBlockValueImpl(Value *Val,402 BasicBlock *BB);403 std::optional<ValueLatticeElement> solveBlockValueNonLocal(Value *Val,404 BasicBlock *BB);405 std::optional<ValueLatticeElement> solveBlockValuePHINode(PHINode *PN,406 BasicBlock *BB);407 std::optional<ValueLatticeElement> solveBlockValueSelect(SelectInst *S,408 BasicBlock *BB);409 std::optional<ConstantRange> getRangeFor(Value *V, Instruction *CxtI,410 BasicBlock *BB);411 std::optional<ValueLatticeElement> solveBlockValueBinaryOpImpl(412 Instruction *I, BasicBlock *BB,413 std::function<ConstantRange(const ConstantRange &, const ConstantRange &)>414 OpFn);415 std::optional<ValueLatticeElement>416 solveBlockValueBinaryOp(BinaryOperator *BBI, BasicBlock *BB);417 std::optional<ValueLatticeElement> solveBlockValueCast(CastInst *CI,418 BasicBlock *BB);419 std::optional<ValueLatticeElement>420 solveBlockValueOverflowIntrinsic(WithOverflowInst *WO, BasicBlock *BB);421 std::optional<ValueLatticeElement> solveBlockValueIntrinsic(IntrinsicInst *II,422 BasicBlock *BB);423 std::optional<ValueLatticeElement>424 solveBlockValueInsertElement(InsertElementInst *IEI, BasicBlock *BB);425 std::optional<ValueLatticeElement>426 solveBlockValueExtractValue(ExtractValueInst *EVI, BasicBlock *BB);427 bool isNonNullAtEndOfBlock(Value *Val, BasicBlock *BB);428 void intersectAssumeOrGuardBlockValueConstantRange(Value *Val,429 ValueLatticeElement &BBLV,430 Instruction *BBI);431 432 void solve();433 434 // For the following methods, if UseBlockValue is true, the function may435 // push additional values to the worklist and return nullopt. If436 // UseBlockValue is false, it will never return nullopt.437 438 std::optional<ValueLatticeElement>439 getValueFromSimpleICmpCondition(CmpInst::Predicate Pred, Value *RHS,440 const APInt &Offset, Instruction *CxtI,441 bool UseBlockValue);442 443 std::optional<ValueLatticeElement>444 getValueFromICmpCondition(Value *Val, ICmpInst *ICI, bool isTrueDest,445 bool UseBlockValue);446 ValueLatticeElement getValueFromTrunc(Value *Val, TruncInst *Trunc,447 bool IsTrueDest);448 449 std::optional<ValueLatticeElement>450 getValueFromCondition(Value *Val, Value *Cond, bool IsTrueDest,451 bool UseBlockValue, unsigned Depth = 0);452 453 std::optional<ValueLatticeElement> getEdgeValueLocal(Value *Val,454 BasicBlock *BBFrom,455 BasicBlock *BBTo,456 bool UseBlockValue);457 458public:459 /// This is the query interface to determine the lattice value for the460 /// specified Value* at the context instruction (if specified) or at the461 /// start of the block.462 ValueLatticeElement getValueInBlock(Value *V, BasicBlock *BB,463 Instruction *CxtI = nullptr);464 465 /// This is the query interface to determine the lattice value for the466 /// specified Value* at the specified instruction using only information467 /// from assumes/guards and range metadata. Unlike getValueInBlock(), no468 /// recursive query is performed.469 ValueLatticeElement getValueAt(Value *V, Instruction *CxtI);470 471 /// This is the query interface to determine the lattice472 /// value for the specified Value* that is true on the specified edge.473 ValueLatticeElement getValueOnEdge(Value *V, BasicBlock *FromBB,474 BasicBlock *ToBB,475 Instruction *CxtI = nullptr);476 477 ValueLatticeElement getValueAtUse(const Use &U);478 479 /// Complete flush all previously computed values480 void clear() {481 TheCache.clear();482 }483 484 /// Printing the LazyValueInfo Analysis.485 void printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) {486 LazyValueInfoAnnotatedWriter Writer(this, DTree);487 F.print(OS, &Writer);488 }489 490 /// This is part of the update interface to remove information related to this491 /// value from the cache.492 void forgetValue(Value *V) { TheCache.eraseValue(V); }493 494 /// This is part of the update interface to inform the cache495 /// that a block has been deleted.496 void eraseBlock(BasicBlock *BB) {497 TheCache.eraseBlock(BB);498 }499 500 /// This is the update interface to inform the cache that an edge from501 /// PredBB to OldSucc has been threaded to be from PredBB to NewSucc.502 void threadEdge(BasicBlock *PredBB,BasicBlock *OldSucc,BasicBlock *NewSucc);503 504 LazyValueInfoImpl(AssumptionCache *AC, const DataLayout &DL,505 Function *GuardDecl)506 : AC(AC), DL(DL), GuardDecl(GuardDecl) {}507};508} // namespace llvm509 510void LazyValueInfoImpl::solve() {511 SmallVector<std::pair<BasicBlock *, Value *>, 8> StartingStack =512 BlockValueStack;513 514 unsigned processedCount = 0;515 while (!BlockValueStack.empty()) {516 processedCount++;517 // Abort if we have to process too many values to get a result for this one.518 // Because of the design of the overdefined cache currently being per-block519 // to avoid naming-related issues (IE it wants to try to give different520 // results for the same name in different blocks), overdefined results don't521 // get cached globally, which in turn means we will often try to rediscover522 // the same overdefined result again and again. Once something like523 // PredicateInfo is used in LVI or CVP, we should be able to make the524 // overdefined cache global, and remove this throttle.525 if (processedCount > MaxProcessedPerValue) {526 LLVM_DEBUG(527 dbgs() << "Giving up on stack because we are getting too deep\n");528 // Fill in the original values529 while (!StartingStack.empty()) {530 std::pair<BasicBlock *, Value *> &e = StartingStack.back();531 TheCache.insertResult(e.second, e.first,532 ValueLatticeElement::getOverdefined());533 StartingStack.pop_back();534 }535 BlockValueSet.clear();536 BlockValueStack.clear();537 return;538 }539 std::pair<BasicBlock *, Value *> e = BlockValueStack.back();540 assert(BlockValueSet.count(e) && "Stack value should be in BlockValueSet!");541 unsigned StackSize = BlockValueStack.size();542 (void) StackSize;543 544 if (solveBlockValue(e.second, e.first)) {545 // The work item was completely processed.546 assert(BlockValueStack.size() == StackSize &&547 BlockValueStack.back() == e && "Nothing should have been pushed!");548#ifndef NDEBUG549 std::optional<ValueLatticeElement> BBLV =550 TheCache.getCachedValueInfo(e.second, e.first);551 assert(BBLV && "Result should be in cache!");552 LLVM_DEBUG(553 dbgs() << "POP " << *e.second << " in " << e.first->getName() << " = "554 << *BBLV << "\n");555#endif556 557 BlockValueStack.pop_back();558 BlockValueSet.erase(e);559 } else {560 // More work needs to be done before revisiting.561 assert(BlockValueStack.size() == StackSize + 1 &&562 "Exactly one element should have been pushed!");563 }564 }565}566 567std::optional<ValueLatticeElement>568LazyValueInfoImpl::getBlockValue(Value *Val, BasicBlock *BB,569 Instruction *CxtI) {570 // If already a constant, there is nothing to compute.571 if (Constant *VC = dyn_cast<Constant>(Val))572 return ValueLatticeElement::get(VC);573 574 if (std::optional<ValueLatticeElement> OptLatticeVal =575 TheCache.getCachedValueInfo(Val, BB)) {576 intersectAssumeOrGuardBlockValueConstantRange(Val, *OptLatticeVal, CxtI);577 return OptLatticeVal;578 }579 580 // We have hit a cycle, assume overdefined.581 if (!pushBlockValue({ BB, Val }))582 return ValueLatticeElement::getOverdefined();583 584 // Yet to be resolved.585 return std::nullopt;586}587 588static ValueLatticeElement getFromRangeMetadata(Instruction *BBI) {589 switch (BBI->getOpcode()) {590 default:591 break;592 case Instruction::Call:593 case Instruction::Invoke:594 if (std::optional<ConstantRange> Range = cast<CallBase>(BBI)->getRange())595 return ValueLatticeElement::getRange(*Range);596 [[fallthrough]];597 case Instruction::Load:598 if (MDNode *Ranges = BBI->getMetadata(LLVMContext::MD_range))599 if (isa<IntegerType>(BBI->getType())) {600 return ValueLatticeElement::getRange(601 getConstantRangeFromMetadata(*Ranges));602 }603 break;604 };605 // Nothing known - will be intersected with other facts606 return ValueLatticeElement::getOverdefined();607}608 609bool LazyValueInfoImpl::solveBlockValue(Value *Val, BasicBlock *BB) {610 assert(!isa<Constant>(Val) && "Value should not be constant");611 assert(!TheCache.getCachedValueInfo(Val, BB) &&612 "Value should not be in cache");613 614 // Hold off inserting this value into the Cache in case we have to return615 // false and come back later.616 std::optional<ValueLatticeElement> Res = solveBlockValueImpl(Val, BB);617 if (!Res)618 // Work pushed, will revisit619 return false;620 621 TheCache.insertResult(Val, BB, *Res);622 return true;623}624 625std::optional<ValueLatticeElement>626LazyValueInfoImpl::solveBlockValueImpl(Value *Val, BasicBlock *BB) {627 Instruction *BBI = dyn_cast<Instruction>(Val);628 if (!BBI || BBI->getParent() != BB)629 return solveBlockValueNonLocal(Val, BB);630 631 if (PHINode *PN = dyn_cast<PHINode>(BBI))632 return solveBlockValuePHINode(PN, BB);633 634 if (auto *SI = dyn_cast<SelectInst>(BBI))635 return solveBlockValueSelect(SI, BB);636 637 // If this value is a nonnull pointer, record it's range and bailout. Note638 // that for all other pointer typed values, we terminate the search at the639 // definition. We could easily extend this to look through geps, bitcasts,640 // and the like to prove non-nullness, but it's not clear that's worth it641 // compile time wise. The context-insensitive value walk done inside642 // isKnownNonZero gets most of the profitable cases at much less expense.643 // This does mean that we have a sensitivity to where the defining644 // instruction is placed, even if it could legally be hoisted much higher.645 // That is unfortunate.646 PointerType *PT = dyn_cast<PointerType>(BBI->getType());647 if (PT && isKnownNonZero(BBI, DL))648 return ValueLatticeElement::getNot(ConstantPointerNull::get(PT));649 650 if (BBI->getType()->isIntOrIntVectorTy()) {651 if (auto *CI = dyn_cast<CastInst>(BBI))652 return solveBlockValueCast(CI, BB);653 654 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(BBI))655 return solveBlockValueBinaryOp(BO, BB);656 657 if (auto *IEI = dyn_cast<InsertElementInst>(BBI))658 return solveBlockValueInsertElement(IEI, BB);659 660 if (auto *EVI = dyn_cast<ExtractValueInst>(BBI))661 return solveBlockValueExtractValue(EVI, BB);662 663 if (auto *II = dyn_cast<IntrinsicInst>(BBI))664 return solveBlockValueIntrinsic(II, BB);665 }666 667 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()668 << "' - unknown inst def found.\n");669 return getFromRangeMetadata(BBI);670}671 672static void AddNonNullPointer(Value *Ptr, NonNullPointerSet &PtrSet,673 bool IsDereferenced = true) {674 // TODO: Use NullPointerIsDefined instead.675 if (Ptr->getType()->getPointerAddressSpace() == 0)676 PtrSet.insert(IsDereferenced ? getUnderlyingObject(Ptr)677 : Ptr->stripInBoundsOffsets());678}679 680static void AddNonNullPointersByInstruction(681 Instruction *I, NonNullPointerSet &PtrSet) {682 if (LoadInst *L = dyn_cast<LoadInst>(I)) {683 AddNonNullPointer(L->getPointerOperand(), PtrSet);684 } else if (StoreInst *S = dyn_cast<StoreInst>(I)) {685 AddNonNullPointer(S->getPointerOperand(), PtrSet);686 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) {687 if (MI->isVolatile()) return;688 689 // FIXME: check whether it has a valuerange that excludes zero?690 ConstantInt *Len = dyn_cast<ConstantInt>(MI->getLength());691 if (!Len || Len->isZero()) return;692 693 AddNonNullPointer(MI->getRawDest(), PtrSet);694 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))695 AddNonNullPointer(MTI->getRawSource(), PtrSet);696 } else if (auto *CB = dyn_cast<CallBase>(I)) {697 for (auto &U : CB->args()) {698 if (U->getType()->isPointerTy() &&699 CB->paramHasNonNullAttr(CB->getArgOperandNo(&U),700 /*AllowUndefOrPoison=*/false))701 AddNonNullPointer(U.get(), PtrSet, /*IsDereferenced=*/false);702 }703 }704}705 706bool LazyValueInfoImpl::isNonNullAtEndOfBlock(Value *Val, BasicBlock *BB) {707 if (NullPointerIsDefined(BB->getParent(),708 Val->getType()->getPointerAddressSpace()))709 return false;710 711 Val = Val->stripInBoundsOffsets();712 return TheCache.isNonNullAtEndOfBlock(Val, BB, [](BasicBlock *BB) {713 NonNullPointerSet NonNullPointers;714 for (Instruction &I : *BB)715 AddNonNullPointersByInstruction(&I, NonNullPointers);716 return NonNullPointers;717 });718}719 720std::optional<ValueLatticeElement>721LazyValueInfoImpl::solveBlockValueNonLocal(Value *Val, BasicBlock *BB) {722 ValueLatticeElement Result; // Start Undefined.723 724 // If this is the entry block, we must be asking about an argument.725 if (BB->isEntryBlock()) {726 assert(isa<Argument>(Val) && "Unknown live-in to the entry block");727 if (std::optional<ConstantRange> Range = cast<Argument>(Val)->getRange())728 return ValueLatticeElement::getRange(*Range);729 return ValueLatticeElement::getOverdefined();730 }731 732 // Loop over all of our predecessors, merging what we know from them into733 // result. If we encounter an unexplored predecessor, we eagerly explore it734 // in a depth first manner. In practice, this has the effect of discovering735 // paths we can't analyze eagerly without spending compile times analyzing736 // other paths. This heuristic benefits from the fact that predecessors are737 // frequently arranged such that dominating ones come first and we quickly738 // find a path to function entry. TODO: We should consider explicitly739 // canonicalizing to make this true rather than relying on this happy740 // accident.741 std::optional<BBLatticeElementMap> PredLatticeElements;742 if (PerPredRanges)743 PredLatticeElements = std::make_optional<BBLatticeElementMap>();744 for (BasicBlock *Pred : predecessors(BB)) {745 // Skip self loops.746 if (Pred == BB)747 continue;748 std::optional<ValueLatticeElement> EdgeResult = getEdgeValue(Val, Pred, BB);749 if (!EdgeResult)750 // Explore that input, then return here751 return std::nullopt;752 753 Result.mergeIn(*EdgeResult);754 755 // If we hit overdefined, exit early. The BlockVals entry is already set756 // to overdefined.757 if (Result.isOverdefined()) {758 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()759 << "' - overdefined because of pred '"760 << Pred->getName() << "' (non local).\n");761 return Result;762 }763 if (PerPredRanges)764 PredLatticeElements->insert({Pred, *EdgeResult});765 }766 767 if (PerPredRanges)768 TheCache.insertPredecessorResults(Val, BB, *PredLatticeElements);769 770 // Return the merged value, which is more precise than 'overdefined'.771 assert(!Result.isOverdefined());772 return Result;773}774 775std::optional<ValueLatticeElement>776LazyValueInfoImpl::solveBlockValuePHINode(PHINode *PN, BasicBlock *BB) {777 ValueLatticeElement Result; // Start Undefined.778 779 // Loop over all of our predecessors, merging what we know from them into780 // result. See the comment about the chosen traversal order in781 // solveBlockValueNonLocal; the same reasoning applies here.782 std::optional<BBLatticeElementMap> PredLatticeElements;783 if (PerPredRanges)784 PredLatticeElements = std::make_optional<BBLatticeElementMap>();785 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {786 BasicBlock *PhiBB = PN->getIncomingBlock(i);787 Value *PhiVal = PN->getIncomingValue(i);788 // Note that we can provide PN as the context value to getEdgeValue, even789 // though the results will be cached, because PN is the value being used as790 // the cache key in the caller.791 std::optional<ValueLatticeElement> EdgeResult =792 getEdgeValue(PhiVal, PhiBB, BB, PN);793 if (!EdgeResult)794 // Explore that input, then return here795 return std::nullopt;796 797 Result.mergeIn(*EdgeResult);798 799 // If we hit overdefined, exit early. The BlockVals entry is already set800 // to overdefined.801 if (Result.isOverdefined()) {802 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()803 << "' - overdefined because of pred (local).\n");804 805 return Result;806 }807 808 if (PerPredRanges)809 PredLatticeElements->insert({PhiBB, *EdgeResult});810 }811 812 if (PerPredRanges)813 TheCache.insertPredecessorResults(PN, BB, *PredLatticeElements);814 815 // Return the merged value, which is more precise than 'overdefined'.816 assert(!Result.isOverdefined() && "Possible PHI in entry block?");817 return Result;818}819 820// If we can determine a constraint on the value given conditions assumed by821// the program, intersect those constraints with BBLV822void LazyValueInfoImpl::intersectAssumeOrGuardBlockValueConstantRange(823 Value *Val, ValueLatticeElement &BBLV, Instruction *BBI) {824 BBI = BBI ? BBI : dyn_cast<Instruction>(Val);825 if (!BBI)826 return;827 828 BasicBlock *BB = BBI->getParent();829 for (auto &AssumeVH : AC->assumptionsFor(Val)) {830 if (!AssumeVH)831 continue;832 833 // Only check assumes in the block of the context instruction. Other834 // assumes will have already been taken into account when the value was835 // propagated from predecessor blocks.836 auto *I = cast<CallInst>(AssumeVH);837 if (I->getParent() != BB || !isValidAssumeForContext(I, BBI))838 continue;839 840 BBLV = BBLV.intersect(*getValueFromCondition(Val, I->getArgOperand(0),841 /*IsTrueDest*/ true,842 /*UseBlockValue*/ false));843 }844 845 // If guards are not used in the module, don't spend time looking for them846 if (GuardDecl && !GuardDecl->use_empty() &&847 BBI->getIterator() != BB->begin()) {848 for (Instruction &I :849 make_range(std::next(BBI->getIterator().getReverse()), BB->rend())) {850 Value *Cond = nullptr;851 if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>(m_Value(Cond))))852 BBLV = BBLV.intersect(*getValueFromCondition(Val, Cond,853 /*IsTrueDest*/ true,854 /*UseBlockValue*/ false));855 }856 }857 858 if (BBLV.isOverdefined()) {859 // Check whether we're checking at the terminator, and the pointer has860 // been dereferenced in this block.861 PointerType *PTy = dyn_cast<PointerType>(Val->getType());862 if (PTy && BB->getTerminator() == BBI &&863 isNonNullAtEndOfBlock(Val, BB))864 BBLV = ValueLatticeElement::getNot(ConstantPointerNull::get(PTy));865 }866}867 868std::optional<ValueLatticeElement>869LazyValueInfoImpl::solveBlockValueSelect(SelectInst *SI, BasicBlock *BB) {870 // Recurse on our inputs if needed871 std::optional<ValueLatticeElement> OptTrueVal =872 getBlockValue(SI->getTrueValue(), BB, SI);873 if (!OptTrueVal)874 return std::nullopt;875 ValueLatticeElement &TrueVal = *OptTrueVal;876 877 std::optional<ValueLatticeElement> OptFalseVal =878 getBlockValue(SI->getFalseValue(), BB, SI);879 if (!OptFalseVal)880 return std::nullopt;881 ValueLatticeElement &FalseVal = *OptFalseVal;882 883 if (TrueVal.isConstantRange() || FalseVal.isConstantRange()) {884 const ConstantRange &TrueCR = TrueVal.asConstantRange(SI->getType());885 const ConstantRange &FalseCR = FalseVal.asConstantRange(SI->getType());886 Value *LHS = nullptr;887 Value *RHS = nullptr;888 SelectPatternResult SPR = matchSelectPattern(SI, LHS, RHS);889 // Is this a min specifically of our two inputs? (Avoid the risk of890 // ValueTracking getting smarter looking back past our immediate inputs.)891 if (SelectPatternResult::isMinOrMax(SPR.Flavor) &&892 ((LHS == SI->getTrueValue() && RHS == SI->getFalseValue()) ||893 (RHS == SI->getTrueValue() && LHS == SI->getFalseValue()))) {894 ConstantRange ResultCR = [&]() {895 switch (SPR.Flavor) {896 default:897 llvm_unreachable("unexpected minmax type!");898 case SPF_SMIN: /// Signed minimum899 return TrueCR.smin(FalseCR);900 case SPF_UMIN: /// Unsigned minimum901 return TrueCR.umin(FalseCR);902 case SPF_SMAX: /// Signed maximum903 return TrueCR.smax(FalseCR);904 case SPF_UMAX: /// Unsigned maximum905 return TrueCR.umax(FalseCR);906 };907 }();908 return ValueLatticeElement::getRange(909 ResultCR, TrueVal.isConstantRangeIncludingUndef() ||910 FalseVal.isConstantRangeIncludingUndef());911 }912 913 if (SPR.Flavor == SPF_ABS) {914 if (LHS == SI->getTrueValue())915 return ValueLatticeElement::getRange(916 TrueCR.abs(), TrueVal.isConstantRangeIncludingUndef());917 if (LHS == SI->getFalseValue())918 return ValueLatticeElement::getRange(919 FalseCR.abs(), FalseVal.isConstantRangeIncludingUndef());920 }921 922 if (SPR.Flavor == SPF_NABS) {923 ConstantRange Zero(APInt::getZero(TrueCR.getBitWidth()));924 if (LHS == SI->getTrueValue())925 return ValueLatticeElement::getRange(926 Zero.sub(TrueCR.abs()), FalseVal.isConstantRangeIncludingUndef());927 if (LHS == SI->getFalseValue())928 return ValueLatticeElement::getRange(929 Zero.sub(FalseCR.abs()), FalseVal.isConstantRangeIncludingUndef());930 }931 }932 933 // Can we constrain the facts about the true and false values by using the934 // condition itself? This shows up with idioms like e.g. select(a > 5, a, 5).935 // TODO: We could potentially refine an overdefined true value above.936 Value *Cond = SI->getCondition();937 // If the value is undef, a different value may be chosen in938 // the select condition.939 if (isGuaranteedNotToBeUndef(Cond, AC)) {940 TrueVal =941 TrueVal.intersect(*getValueFromCondition(SI->getTrueValue(), Cond,942 /*IsTrueDest*/ true,943 /*UseBlockValue*/ false));944 FalseVal =945 FalseVal.intersect(*getValueFromCondition(SI->getFalseValue(), Cond,946 /*IsTrueDest*/ false,947 /*UseBlockValue*/ false));948 }949 950 TrueVal.mergeIn(FalseVal);951 return TrueVal;952}953 954std::optional<ConstantRange>955LazyValueInfoImpl::getRangeFor(Value *V, Instruction *CxtI, BasicBlock *BB) {956 std::optional<ValueLatticeElement> OptVal = getBlockValue(V, BB, CxtI);957 if (!OptVal)958 return std::nullopt;959 return OptVal->asConstantRange(V->getType());960}961 962std::optional<ValueLatticeElement>963LazyValueInfoImpl::solveBlockValueCast(CastInst *CI, BasicBlock *BB) {964 // Filter out casts we don't know how to reason about before attempting to965 // recurse on our operand. This can cut a long search short if we know we're966 // not going to be able to get any useful information anways.967 switch (CI->getOpcode()) {968 case Instruction::Trunc:969 case Instruction::SExt:970 case Instruction::ZExt:971 break;972 default:973 // Unhandled instructions are overdefined.974 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()975 << "' - overdefined (unknown cast).\n");976 return ValueLatticeElement::getOverdefined();977 }978 979 // Figure out the range of the LHS. If that fails, we still apply the980 // transfer rule on the full set since we may be able to locally infer981 // interesting facts.982 std::optional<ConstantRange> LHSRes = getRangeFor(CI->getOperand(0), CI, BB);983 if (!LHSRes)984 // More work to do before applying this transfer rule.985 return std::nullopt;986 const ConstantRange &LHSRange = *LHSRes;987 988 const unsigned ResultBitWidth = CI->getType()->getScalarSizeInBits();989 990 // NOTE: We're currently limited by the set of operations that ConstantRange991 // can evaluate symbolically. Enhancing that set will allows us to analyze992 // more definitions.993 ConstantRange Res = ConstantRange::getEmpty(ResultBitWidth);994 if (auto *Trunc = dyn_cast<TruncInst>(CI))995 Res = LHSRange.truncate(ResultBitWidth, Trunc->getNoWrapKind());996 else997 Res = LHSRange.castOp(CI->getOpcode(), ResultBitWidth);998 999 return ValueLatticeElement::getRange(Res);1000}1001 1002std::optional<ValueLatticeElement>1003LazyValueInfoImpl::solveBlockValueBinaryOpImpl(1004 Instruction *I, BasicBlock *BB,1005 std::function<ConstantRange(const ConstantRange &, const ConstantRange &)>1006 OpFn) {1007 Value *LHS = I->getOperand(0);1008 Value *RHS = I->getOperand(1);1009 1010 auto ThreadBinOpOverSelect =1011 [&](Value *X, const ConstantRange &CRX, SelectInst *Y,1012 bool XIsLHS) -> std::optional<ValueLatticeElement> {1013 Value *Cond = Y->getCondition();1014 // Only handle selects with constant values.1015 Constant *TrueC = dyn_cast<Constant>(Y->getTrueValue());1016 if (!TrueC)1017 return std::nullopt;1018 Constant *FalseC = dyn_cast<Constant>(Y->getFalseValue());1019 if (!FalseC)1020 return std::nullopt;1021 if (!isGuaranteedNotToBeUndef(Cond, AC))1022 return std::nullopt;1023 1024 ConstantRange TrueX =1025 CRX.intersectWith(getValueFromCondition(X, Cond, /*CondIsTrue=*/true,1026 /*UseBlockValue=*/false)1027 ->asConstantRange(X->getType()));1028 ConstantRange FalseX =1029 CRX.intersectWith(getValueFromCondition(X, Cond, /*CondIsTrue=*/false,1030 /*UseBlockValue=*/false)1031 ->asConstantRange(X->getType()));1032 ConstantRange TrueY = TrueC->toConstantRange();1033 ConstantRange FalseY = FalseC->toConstantRange();1034 1035 if (XIsLHS)1036 return ValueLatticeElement::getRange(1037 OpFn(TrueX, TrueY).unionWith(OpFn(FalseX, FalseY)));1038 return ValueLatticeElement::getRange(1039 OpFn(TrueY, TrueX).unionWith(OpFn(FalseY, FalseX)));1040 };1041 1042 // Figure out the ranges of the operands. If that fails, use a1043 // conservative range, but apply the transfer rule anyways. This1044 // lets us pick up facts from expressions like "and i32 (call i321045 // @foo()), 32"1046 std::optional<ConstantRange> LHSRes = getRangeFor(LHS, I, BB);1047 if (!LHSRes)1048 return std::nullopt;1049 1050 // Try to thread binop over rhs select1051 if (auto *SI = dyn_cast<SelectInst>(RHS)) {1052 if (auto Res = ThreadBinOpOverSelect(LHS, *LHSRes, SI, /*XIsLHS=*/true))1053 return *Res;1054 }1055 1056 std::optional<ConstantRange> RHSRes = getRangeFor(RHS, I, BB);1057 if (!RHSRes)1058 return std::nullopt;1059 1060 // Try to thread binop over lhs select1061 if (auto *SI = dyn_cast<SelectInst>(LHS)) {1062 if (auto Res = ThreadBinOpOverSelect(RHS, *RHSRes, SI, /*XIsLHS=*/false))1063 return *Res;1064 }1065 1066 const ConstantRange &LHSRange = *LHSRes;1067 const ConstantRange &RHSRange = *RHSRes;1068 1069 std::optional<ValueLatticeElement> MergedResult =1070 ValueLatticeElement::getRange(OpFn(LHSRange, RHSRange));1071 1072 if (!PerPredRanges)1073 return MergedResult;1074 1075 std::optional<BBLatticeElementMap> PredLHS =1076 TheCache.getCachedPredecessorInfo(LHS, BB);1077 if (!PredLHS)1078 return MergedResult;1079 std::optional<BBLatticeElementMap> PredRHS =1080 TheCache.getCachedPredecessorInfo(RHS, BB);1081 if (!PredRHS)1082 return MergedResult;1083 1084 const BBLatticeElementMap &LHSPredMap = *PredLHS;1085 const BBLatticeElementMap &RHSPredMap = *PredRHS;1086 1087 BBLatticeElementMap PredLatticeElements;1088 ValueLatticeElement OverallPredResult;1089 for (auto *Pred : predecessors(BB)) {1090 auto LHSIt = LHSPredMap.find_as(Pred);1091 if (LHSIt == LHSPredMap.end())1092 return MergedResult;1093 const ValueLatticeElement &LHSFromPred = LHSIt->second;1094 std::optional<ConstantRange> LHSFromPredRes =1095 LHSFromPred.asConstantRange(LHS->getType());1096 if (!LHSFromPredRes)1097 return MergedResult;1098 1099 auto RHSIt = RHSPredMap.find_as(Pred);1100 if (RHSIt == RHSPredMap.end())1101 return MergedResult;1102 const ValueLatticeElement &RHSFromPred = RHSIt->second;1103 std::optional<ConstantRange> RHSFromPredRes =1104 RHSFromPred.asConstantRange(RHS->getType());1105 if (!RHSFromPredRes)1106 return MergedResult;1107 1108 const ConstantRange &LHSFromPredRange = *LHSFromPredRes;1109 const ConstantRange &RHSFromPredRange = *RHSFromPredRes;1110 std::optional<ValueLatticeElement> PredResult =1111 ValueLatticeElement::getRange(OpFn(LHSFromPredRange, RHSFromPredRange));1112 if (!PredResult)1113 return MergedResult;1114 if (PredResult->isOverdefined()) {1115 LLVM_DEBUG(1116 dbgs() << " pred BB '" << Pred->getName() << "' for BB '"1117 << BB->getName()1118 << "' overdefined. Discarding all predecessor intervals.\n");1119 return MergedResult;1120 }1121 PredLatticeElements.insert({Pred, *PredResult});1122 OverallPredResult.mergeIn(*PredResult);1123 }1124 1125 // If this point is reached, all predecessors for both LHS and RHS have1126 // constant ranges previously computed. Can cache result and use the1127 // OverallPredResult;1128 TheCache.insertPredecessorResults(I, BB, PredLatticeElements);1129 1130 LLVM_DEBUG(dbgs() << " Using predecessor intervals, evaluated " << *I1131 << " to: " << OverallPredResult << ".\n");1132 1133 if (!MergedResult)1134 return OverallPredResult;1135 1136 LLVM_DEBUG(dbgs() << " Intersecting intervals for " << *I << ": "1137 << OverallPredResult << " and " << MergedResult << ".\n");1138 return MergedResult->intersect(OverallPredResult);1139}1140 1141std::optional<ValueLatticeElement>1142LazyValueInfoImpl::solveBlockValueBinaryOp(BinaryOperator *BO, BasicBlock *BB) {1143 assert(BO->getOperand(0)->getType()->isSized() &&1144 "all operands to binary operators are sized");1145 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(BO)) {1146 unsigned NoWrapKind = OBO->getNoWrapKind();1147 return solveBlockValueBinaryOpImpl(1148 BO, BB,1149 [BO, NoWrapKind](const ConstantRange &CR1, const ConstantRange &CR2) {1150 return CR1.overflowingBinaryOp(BO->getOpcode(), CR2, NoWrapKind);1151 });1152 }1153 1154 return solveBlockValueBinaryOpImpl(1155 BO, BB, [BO](const ConstantRange &CR1, const ConstantRange &CR2) {1156 return CR1.binaryOp(BO->getOpcode(), CR2);1157 });1158}1159 1160std::optional<ValueLatticeElement>1161LazyValueInfoImpl::solveBlockValueOverflowIntrinsic(WithOverflowInst *WO,1162 BasicBlock *BB) {1163 return solveBlockValueBinaryOpImpl(1164 WO, BB, [WO](const ConstantRange &CR1, const ConstantRange &CR2) {1165 return CR1.binaryOp(WO->getBinaryOp(), CR2);1166 });1167}1168 1169std::optional<ValueLatticeElement>1170LazyValueInfoImpl::solveBlockValueIntrinsic(IntrinsicInst *II, BasicBlock *BB) {1171 ValueLatticeElement MetadataVal = getFromRangeMetadata(II);1172 if (!ConstantRange::isIntrinsicSupported(II->getIntrinsicID())) {1173 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()1174 << "' - unknown intrinsic.\n");1175 return MetadataVal;1176 }1177 1178 SmallVector<ConstantRange, 2> OpRanges;1179 for (Value *Op : II->args()) {1180 std::optional<ConstantRange> Range = getRangeFor(Op, II, BB);1181 if (!Range)1182 return std::nullopt;1183 OpRanges.push_back(*Range);1184 }1185 1186 return ValueLatticeElement::getRange(1187 ConstantRange::intrinsic(II->getIntrinsicID(), OpRanges))1188 .intersect(MetadataVal);1189}1190 1191std::optional<ValueLatticeElement>1192LazyValueInfoImpl::solveBlockValueInsertElement(InsertElementInst *IEI,1193 BasicBlock *BB) {1194 std::optional<ValueLatticeElement> OptEltVal =1195 getBlockValue(IEI->getOperand(1), BB, IEI);1196 if (!OptEltVal)1197 return std::nullopt;1198 ValueLatticeElement &Res = *OptEltVal;1199 1200 std::optional<ValueLatticeElement> OptVecVal =1201 getBlockValue(IEI->getOperand(0), BB, IEI);1202 if (!OptVecVal)1203 return std::nullopt;1204 1205 // Bail out if the inserted element is a constant expression. Unlike other1206 // ValueLattice types, these are not considered an implicit splat when a1207 // vector type is used.1208 // We could call ConstantFoldInsertElementInstruction here to handle these.1209 if (OptEltVal->isConstant())1210 return ValueLatticeElement::getOverdefined();1211 1212 Res.mergeIn(*OptVecVal);1213 return Res;1214}1215 1216std::optional<ValueLatticeElement>1217LazyValueInfoImpl::solveBlockValueExtractValue(ExtractValueInst *EVI,1218 BasicBlock *BB) {1219 if (auto *WO = dyn_cast<WithOverflowInst>(EVI->getAggregateOperand()))1220 if (EVI->getNumIndices() == 1 && *EVI->idx_begin() == 0)1221 return solveBlockValueOverflowIntrinsic(WO, BB);1222 1223 // Handle extractvalue of insertvalue to allow further simplification1224 // based on replaced with.overflow intrinsics.1225 if (Value *V = simplifyExtractValueInst(1226 EVI->getAggregateOperand(), EVI->getIndices(),1227 EVI->getDataLayout()))1228 return getBlockValue(V, BB, EVI);1229 1230 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName()1231 << "' - overdefined (unknown extractvalue).\n");1232 return ValueLatticeElement::getOverdefined();1233}1234 1235static bool matchICmpOperand(APInt &Offset, Value *LHS, Value *Val,1236 ICmpInst::Predicate Pred) {1237 if (LHS == Val)1238 return true;1239 1240 // Handle range checking idiom produced by InstCombine. We will subtract the1241 // offset from the allowed range for RHS in this case.1242 const APInt *C;1243 if (match(LHS, m_AddLike(m_Specific(Val), m_APInt(C)))) {1244 Offset = *C;1245 return true;1246 }1247 1248 // Handle the symmetric case. This appears in saturation patterns like1249 // (x == 16) ? 16 : (x + 1).1250 if (match(Val, m_AddLike(m_Specific(LHS), m_APInt(C)))) {1251 Offset = -*C;1252 return true;1253 }1254 1255 // If (x | y) < C, then (x < C) && (y < C).1256 if (match(LHS, m_c_Or(m_Specific(Val), m_Value())) &&1257 (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE))1258 return true;1259 1260 // If (x & y) > C, then (x > C) && (y > C).1261 if (match(LHS, m_c_And(m_Specific(Val), m_Value())) &&1262 (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE))1263 return true;1264 1265 return false;1266}1267 1268/// Get value range for a "(Val + Offset) Pred RHS" condition.1269std::optional<ValueLatticeElement>1270LazyValueInfoImpl::getValueFromSimpleICmpCondition(CmpInst::Predicate Pred,1271 Value *RHS,1272 const APInt &Offset,1273 Instruction *CxtI,1274 bool UseBlockValue) {1275 ConstantRange RHSRange(RHS->getType()->getScalarSizeInBits(),1276 /*isFullSet=*/true);1277 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {1278 RHSRange = ConstantRange(CI->getValue());1279 } else if (UseBlockValue) {1280 std::optional<ValueLatticeElement> R =1281 getBlockValue(RHS, CxtI->getParent(), CxtI);1282 if (!R)1283 return std::nullopt;1284 RHSRange = R->asConstantRange(RHS->getType());1285 }1286 1287 ConstantRange TrueValues =1288 ConstantRange::makeAllowedICmpRegion(Pred, RHSRange);1289 return ValueLatticeElement::getRange(TrueValues.subtract(Offset));1290}1291 1292static std::optional<ConstantRange>1293getRangeViaSLT(CmpInst::Predicate Pred, APInt RHS,1294 function_ref<std::optional<ConstantRange>(const APInt &)> Fn) {1295 bool Invert = false;1296 if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) {1297 Pred = ICmpInst::getInversePredicate(Pred);1298 Invert = true;1299 }1300 if (Pred == ICmpInst::ICMP_SLE) {1301 Pred = ICmpInst::ICMP_SLT;1302 if (RHS.isMaxSignedValue())1303 return std::nullopt; // Could also return full/empty here, if we wanted.1304 ++RHS;1305 }1306 assert(Pred == ICmpInst::ICMP_SLT && "Must be signed predicate");1307 if (auto CR = Fn(RHS))1308 return Invert ? CR->inverse() : CR;1309 return std::nullopt;1310}1311 1312/// Get value range for a "ctpop(Val) Pred RHS" condition.1313static ValueLatticeElement getValueFromICmpCtpop(ICmpInst::Predicate Pred,1314 Value *RHS) {1315 unsigned BitWidth = RHS->getType()->getScalarSizeInBits();1316 1317 auto *RHSConst = dyn_cast<ConstantInt>(RHS);1318 if (!RHSConst)1319 return ValueLatticeElement::getOverdefined();1320 1321 ConstantRange ResValRange =1322 ConstantRange::makeExactICmpRegion(Pred, RHSConst->getValue());1323 1324 unsigned ResMin = ResValRange.getUnsignedMin().getLimitedValue(BitWidth);1325 unsigned ResMax = ResValRange.getUnsignedMax().getLimitedValue(BitWidth);1326 1327 APInt ValMin = APInt::getLowBitsSet(BitWidth, ResMin);1328 APInt ValMax = APInt::getHighBitsSet(BitWidth, ResMax);1329 return ValueLatticeElement::getRange(1330 ConstantRange::getNonEmpty(std::move(ValMin), ValMax + 1));1331}1332 1333std::optional<ValueLatticeElement> LazyValueInfoImpl::getValueFromICmpCondition(1334 Value *Val, ICmpInst *ICI, bool isTrueDest, bool UseBlockValue) {1335 Value *LHS = ICI->getOperand(0);1336 Value *RHS = ICI->getOperand(1);1337 1338 // Get the predicate that must hold along the considered edge.1339 CmpInst::Predicate EdgePred =1340 isTrueDest ? ICI->getPredicate() : ICI->getInversePredicate();1341 1342 if (isa<Constant>(RHS)) {1343 if (ICI->isEquality() && LHS == Val) {1344 if (EdgePred == ICmpInst::ICMP_EQ)1345 return ValueLatticeElement::get(cast<Constant>(RHS));1346 else if (!isa<UndefValue>(RHS))1347 return ValueLatticeElement::getNot(cast<Constant>(RHS));1348 }1349 }1350 1351 Type *Ty = Val->getType();1352 if (!Ty->isIntegerTy())1353 return ValueLatticeElement::getOverdefined();1354 1355 unsigned BitWidth = Ty->getScalarSizeInBits();1356 APInt Offset(BitWidth, 0);1357 if (matchICmpOperand(Offset, LHS, Val, EdgePred))1358 return getValueFromSimpleICmpCondition(EdgePred, RHS, Offset, ICI,1359 UseBlockValue);1360 1361 CmpInst::Predicate SwappedPred = CmpInst::getSwappedPredicate(EdgePred);1362 if (matchICmpOperand(Offset, RHS, Val, SwappedPred))1363 return getValueFromSimpleICmpCondition(SwappedPred, LHS, Offset, ICI,1364 UseBlockValue);1365 1366 if (match(LHS, m_Intrinsic<Intrinsic::ctpop>(m_Specific(Val))))1367 return getValueFromICmpCtpop(EdgePred, RHS);1368 1369 const APInt *Mask, *C;1370 if (match(LHS, m_And(m_Specific(Val), m_APInt(Mask))) &&1371 match(RHS, m_APInt(C))) {1372 // If (Val & Mask) == C then all the masked bits are known and we can1373 // compute a value range based on that.1374 if (EdgePred == ICmpInst::ICMP_EQ) {1375 KnownBits Known;1376 Known.Zero = ~*C & *Mask;1377 Known.One = *C & *Mask;1378 return ValueLatticeElement::getRange(1379 ConstantRange::fromKnownBits(Known, /*IsSigned*/ false));1380 }1381 1382 if (EdgePred == ICmpInst::ICMP_NE)1383 return ValueLatticeElement::getRange(1384 ConstantRange::makeMaskNotEqualRange(*Mask, *C));1385 }1386 1387 // If (X urem Modulus) >= C, then X >= C.1388 // If trunc X >= C, then X >= C.1389 // TODO: An upper bound could be computed as well.1390 if (match(LHS, m_CombineOr(m_URem(m_Specific(Val), m_Value()),1391 m_Trunc(m_Specific(Val)))) &&1392 match(RHS, m_APInt(C))) {1393 // Use the icmp region so we don't have to deal with different predicates.1394 ConstantRange CR = ConstantRange::makeExactICmpRegion(EdgePred, *C);1395 if (!CR.isEmptySet())1396 return ValueLatticeElement::getRange(ConstantRange::getNonEmpty(1397 CR.getUnsignedMin().zext(BitWidth), APInt(BitWidth, 0)));1398 }1399 1400 // Recognize:1401 // icmp slt (ashr X, ShAmtC), C --> icmp slt X, C << ShAmtC1402 // Preconditions: (C << ShAmtC) >> ShAmtC == C1403 const APInt *ShAmtC;1404 if (CmpInst::isSigned(EdgePred) &&1405 match(LHS, m_AShr(m_Specific(Val), m_APInt(ShAmtC))) &&1406 match(RHS, m_APInt(C))) {1407 auto CR = getRangeViaSLT(1408 EdgePred, *C, [&](const APInt &RHS) -> std::optional<ConstantRange> {1409 APInt New = RHS << *ShAmtC;1410 if ((New.ashr(*ShAmtC)) != RHS)1411 return std::nullopt;1412 return ConstantRange::getNonEmpty(1413 APInt::getSignedMinValue(New.getBitWidth()), New);1414 });1415 if (CR)1416 return ValueLatticeElement::getRange(*CR);1417 }1418 1419 // a - b or ptrtoint(a) - ptrtoint(b) ==/!= 0 if a ==/!= b1420 Value *X, *Y;1421 if (ICI->isEquality() && match(Val, m_Sub(m_Value(X), m_Value(Y)))) {1422 // Peek through ptrtoints1423 match(X, m_PtrToIntSameSize(DL, m_Value(X)));1424 match(Y, m_PtrToIntSameSize(DL, m_Value(Y)));1425 if ((X == LHS && Y == RHS) || (X == RHS && Y == LHS)) {1426 Constant *NullVal = Constant::getNullValue(Val->getType());1427 if (EdgePred == ICmpInst::ICMP_EQ)1428 return ValueLatticeElement::get(NullVal);1429 return ValueLatticeElement::getNot(NullVal);1430 }1431 }1432 1433 return ValueLatticeElement::getOverdefined();1434}1435 1436ValueLatticeElement LazyValueInfoImpl::getValueFromTrunc(Value *Val,1437 TruncInst *Trunc,1438 bool IsTrueDest) {1439 assert(Trunc->getType()->isIntOrIntVectorTy(1));1440 1441 if (Trunc->getOperand(0) != Val)1442 return ValueLatticeElement::getOverdefined();1443 1444 Type *Ty = Val->getType();1445 1446 if (Trunc->hasNoUnsignedWrap()) {1447 if (IsTrueDest)1448 return ValueLatticeElement::get(ConstantInt::get(Ty, 1));1449 return ValueLatticeElement::get(Constant::getNullValue(Ty));1450 }1451 1452 if (IsTrueDest)1453 return ValueLatticeElement::getNot(Constant::getNullValue(Ty));1454 return ValueLatticeElement::getNot(Constant::getAllOnesValue(Ty));1455}1456 1457// Handle conditions of the form1458// extractvalue(op.with.overflow(%x, C), 1).1459static ValueLatticeElement getValueFromOverflowCondition(1460 Value *Val, WithOverflowInst *WO, bool IsTrueDest) {1461 // TODO: This only works with a constant RHS for now. We could also compute1462 // the range of the RHS, but this doesn't fit into the current structure of1463 // the edge value calculation.1464 const APInt *C;1465 if (WO->getLHS() != Val || !match(WO->getRHS(), m_APInt(C)))1466 return ValueLatticeElement::getOverdefined();1467 1468 // Calculate the possible values of %x for which no overflow occurs.1469 ConstantRange NWR = ConstantRange::makeExactNoWrapRegion(1470 WO->getBinaryOp(), *C, WO->getNoWrapKind());1471 1472 // If overflow is false, %x is constrained to NWR. If overflow is true, %x is1473 // constrained to it's inverse (all values that might cause overflow).1474 if (IsTrueDest)1475 NWR = NWR.inverse();1476 return ValueLatticeElement::getRange(NWR);1477}1478 1479std::optional<ValueLatticeElement>1480LazyValueInfoImpl::getValueFromCondition(Value *Val, Value *Cond,1481 bool IsTrueDest, bool UseBlockValue,1482 unsigned Depth) {1483 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cond))1484 return getValueFromICmpCondition(Val, ICI, IsTrueDest, UseBlockValue);1485 1486 if (auto *Trunc = dyn_cast<TruncInst>(Cond))1487 return getValueFromTrunc(Val, Trunc, IsTrueDest);1488 1489 if (auto *EVI = dyn_cast<ExtractValueInst>(Cond))1490 if (auto *WO = dyn_cast<WithOverflowInst>(EVI->getAggregateOperand()))1491 if (EVI->getNumIndices() == 1 && *EVI->idx_begin() == 1)1492 return getValueFromOverflowCondition(Val, WO, IsTrueDest);1493 1494 if (++Depth == MaxAnalysisRecursionDepth)1495 return ValueLatticeElement::getOverdefined();1496 1497 Value *N;1498 if (match(Cond, m_Not(m_Value(N))))1499 return getValueFromCondition(Val, N, !IsTrueDest, UseBlockValue, Depth);1500 1501 Value *L, *R;1502 bool IsAnd;1503 if (match(Cond, m_LogicalAnd(m_Value(L), m_Value(R))))1504 IsAnd = true;1505 else if (match(Cond, m_LogicalOr(m_Value(L), m_Value(R))))1506 IsAnd = false;1507 else1508 return ValueLatticeElement::getOverdefined();1509 1510 std::optional<ValueLatticeElement> LV =1511 getValueFromCondition(Val, L, IsTrueDest, UseBlockValue, Depth);1512 if (!LV)1513 return std::nullopt;1514 std::optional<ValueLatticeElement> RV =1515 getValueFromCondition(Val, R, IsTrueDest, UseBlockValue, Depth);1516 if (!RV)1517 return std::nullopt;1518 1519 // if (L && R) -> intersect L and R1520 // if (!(L || R)) -> intersect !L and !R1521 // if (L || R) -> union L and R1522 // if (!(L && R)) -> union !L and !R1523 if (IsTrueDest ^ IsAnd) {1524 LV->mergeIn(*RV);1525 return *LV;1526 }1527 1528 return LV->intersect(*RV);1529}1530 1531// Return true if Usr has Op as an operand, otherwise false.1532static bool usesOperand(User *Usr, Value *Op) {1533 return is_contained(Usr->operands(), Op);1534}1535 1536// Return true if the instruction type of Val is supported by1537// constantFoldUser(). Currently CastInst, BinaryOperator and FreezeInst only.1538// Call this before calling constantFoldUser() to find out if it's even worth1539// attempting to call it.1540static bool isOperationFoldable(User *Usr) {1541 return isa<CastInst>(Usr) || isa<BinaryOperator>(Usr) || isa<FreezeInst>(Usr);1542}1543 1544// Check if Usr can be simplified to an integer constant when the value of one1545// of its operands Op is an integer constant OpConstVal. If so, return it as an1546// lattice value range with a single element or otherwise return an overdefined1547// lattice value.1548static ValueLatticeElement constantFoldUser(User *Usr, Value *Op,1549 const APInt &OpConstVal,1550 const DataLayout &DL) {1551 assert(isOperationFoldable(Usr) && "Precondition");1552 Constant* OpConst = Constant::getIntegerValue(Op->getType(), OpConstVal);1553 // Check if Usr can be simplified to a constant.1554 if (auto *CI = dyn_cast<CastInst>(Usr)) {1555 assert(CI->getOperand(0) == Op && "Operand 0 isn't Op");1556 if (auto *C = dyn_cast_or_null<ConstantInt>(1557 simplifyCastInst(CI->getOpcode(), OpConst,1558 CI->getDestTy(), DL))) {1559 return ValueLatticeElement::getRange(ConstantRange(C->getValue()));1560 }1561 } else if (auto *BO = dyn_cast<BinaryOperator>(Usr)) {1562 bool Op0Match = BO->getOperand(0) == Op;1563 bool Op1Match = BO->getOperand(1) == Op;1564 assert((Op0Match || Op1Match) &&1565 "Operand 0 nor Operand 1 isn't a match");1566 Value *LHS = Op0Match ? OpConst : BO->getOperand(0);1567 Value *RHS = Op1Match ? OpConst : BO->getOperand(1);1568 if (auto *C = dyn_cast_or_null<ConstantInt>(1569 simplifyBinOp(BO->getOpcode(), LHS, RHS, DL))) {1570 return ValueLatticeElement::getRange(ConstantRange(C->getValue()));1571 }1572 } else if (isa<FreezeInst>(Usr)) {1573 assert(cast<FreezeInst>(Usr)->getOperand(0) == Op && "Operand 0 isn't Op");1574 return ValueLatticeElement::getRange(ConstantRange(OpConstVal));1575 }1576 return ValueLatticeElement::getOverdefined();1577}1578 1579/// Compute the value of Val on the edge BBFrom -> BBTo.1580std::optional<ValueLatticeElement>1581LazyValueInfoImpl::getEdgeValueLocal(Value *Val, BasicBlock *BBFrom,1582 BasicBlock *BBTo, bool UseBlockValue) {1583 // TODO: Handle more complex conditionals. If (v == 0 || v2 < 1) is false, we1584 // know that v != 0.1585 if (BranchInst *BI = dyn_cast<BranchInst>(BBFrom->getTerminator())) {1586 // If this is a conditional branch and only one successor goes to BBTo, then1587 // we may be able to infer something from the condition.1588 if (BI->isConditional() &&1589 BI->getSuccessor(0) != BI->getSuccessor(1)) {1590 bool isTrueDest = BI->getSuccessor(0) == BBTo;1591 assert(BI->getSuccessor(!isTrueDest) == BBTo &&1592 "BBTo isn't a successor of BBFrom");1593 Value *Condition = BI->getCondition();1594 1595 // If V is the condition of the branch itself, then we know exactly what1596 // it is.1597 // NB: The condition on a `br` can't be a vector type.1598 if (Condition == Val)1599 return ValueLatticeElement::get(ConstantInt::get(1600 Type::getInt1Ty(Val->getContext()), isTrueDest));1601 1602 // If the condition of the branch is an equality comparison, we may be1603 // able to infer the value.1604 std::optional<ValueLatticeElement> Result =1605 getValueFromCondition(Val, Condition, isTrueDest, UseBlockValue);1606 if (!Result)1607 return std::nullopt;1608 1609 if (!Result->isOverdefined())1610 return Result;1611 1612 if (User *Usr = dyn_cast<User>(Val)) {1613 assert(Result->isOverdefined() && "Result isn't overdefined");1614 // Check with isOperationFoldable() first to avoid linearly iterating1615 // over the operands unnecessarily which can be expensive for1616 // instructions with many operands.1617 if (isa<IntegerType>(Usr->getType()) && isOperationFoldable(Usr)) {1618 const DataLayout &DL = BBTo->getDataLayout();1619 if (usesOperand(Usr, Condition)) {1620 // If Val has Condition as an operand and Val can be folded into a1621 // constant with either Condition == true or Condition == false,1622 // propagate the constant.1623 // eg.1624 // ; %Val is true on the edge to %then.1625 // %Val = and i1 %Condition, true.1626 // br %Condition, label %then, label %else1627 APInt ConditionVal(1, isTrueDest ? 1 : 0);1628 Result = constantFoldUser(Usr, Condition, ConditionVal, DL);1629 } else if (isa<TruncInst, ZExtInst, SExtInst>(Usr)) {1630 ValueLatticeElement OpLatticeVal =1631 *getValueFromCondition(Usr->getOperand(0), Condition,1632 isTrueDest, /*UseBlockValue*/ false);1633 1634 if (OpLatticeVal.isConstantRange()) {1635 const unsigned ResultBitWidth =1636 Usr->getType()->getScalarSizeInBits();1637 if (auto *Trunc = dyn_cast<TruncInst>(Usr))1638 return ValueLatticeElement::getRange(1639 OpLatticeVal.getConstantRange().truncate(1640 ResultBitWidth, Trunc->getNoWrapKind()));1641 1642 return ValueLatticeElement::getRange(1643 OpLatticeVal.getConstantRange().castOp(1644 cast<CastInst>(Usr)->getOpcode(), ResultBitWidth));1645 }1646 if (OpLatticeVal.isConstant()) {1647 Constant *C = OpLatticeVal.getConstant();1648 if (auto *CastC = ConstantFoldCastOperand(1649 cast<CastInst>(Usr)->getOpcode(), C, Usr->getType(), DL))1650 return ValueLatticeElement::get(CastC);1651 }1652 return ValueLatticeElement::getOverdefined();1653 } else {1654 // If one of Val's operand has an inferred value, we may be able to1655 // infer the value of Val.1656 // eg.1657 // ; %Val is 94 on the edge to %then.1658 // %Val = add i8 %Op, 11659 // %Condition = icmp eq i8 %Op, 931660 // br i1 %Condition, label %then, label %else1661 for (unsigned i = 0; i < Usr->getNumOperands(); ++i) {1662 Value *Op = Usr->getOperand(i);1663 ValueLatticeElement OpLatticeVal = *getValueFromCondition(1664 Op, Condition, isTrueDest, /*UseBlockValue*/ false);1665 if (std::optional<APInt> OpConst =1666 OpLatticeVal.asConstantInteger()) {1667 Result = constantFoldUser(Usr, Op, *OpConst, DL);1668 break;1669 }1670 }1671 }1672 }1673 }1674 if (!Result->isOverdefined())1675 return Result;1676 }1677 }1678 1679 // If the edge was formed by a switch on the value, then we may know exactly1680 // what it is.1681 if (SwitchInst *SI = dyn_cast<SwitchInst>(BBFrom->getTerminator())) {1682 Value *Condition = SI->getCondition();1683 if (!isa<IntegerType>(Val->getType()))1684 return ValueLatticeElement::getOverdefined();1685 bool ValUsesConditionAndMayBeFoldable = false;1686 if (Condition != Val) {1687 // Check if Val has Condition as an operand.1688 if (User *Usr = dyn_cast<User>(Val))1689 ValUsesConditionAndMayBeFoldable = isOperationFoldable(Usr) &&1690 usesOperand(Usr, Condition);1691 if (!ValUsesConditionAndMayBeFoldable)1692 return ValueLatticeElement::getOverdefined();1693 }1694 assert((Condition == Val || ValUsesConditionAndMayBeFoldable) &&1695 "Condition != Val nor Val doesn't use Condition");1696 1697 bool DefaultCase = SI->getDefaultDest() == BBTo;1698 unsigned BitWidth = Val->getType()->getIntegerBitWidth();1699 ConstantRange EdgesVals(BitWidth, DefaultCase/*isFullSet*/);1700 1701 for (auto Case : SI->cases()) {1702 APInt CaseValue = Case.getCaseValue()->getValue();1703 ConstantRange EdgeVal(CaseValue);1704 if (ValUsesConditionAndMayBeFoldable) {1705 User *Usr = cast<User>(Val);1706 const DataLayout &DL = BBTo->getDataLayout();1707 ValueLatticeElement EdgeLatticeVal =1708 constantFoldUser(Usr, Condition, CaseValue, DL);1709 if (EdgeLatticeVal.isOverdefined())1710 return ValueLatticeElement::getOverdefined();1711 EdgeVal = EdgeLatticeVal.getConstantRange();1712 }1713 if (DefaultCase) {1714 // It is possible that the default destination is the destination of1715 // some cases. We cannot perform difference for those cases.1716 // We know Condition != CaseValue in BBTo. In some cases we can use1717 // this to infer Val == f(Condition) is != f(CaseValue). For now, we1718 // only do this when f is identity (i.e. Val == Condition), but we1719 // should be able to do this for any injective f.1720 if (Case.getCaseSuccessor() != BBTo && Condition == Val)1721 EdgesVals = EdgesVals.difference(EdgeVal);1722 } else if (Case.getCaseSuccessor() == BBTo)1723 EdgesVals = EdgesVals.unionWith(EdgeVal);1724 }1725 return ValueLatticeElement::getRange(std::move(EdgesVals));1726 }1727 return ValueLatticeElement::getOverdefined();1728}1729 1730/// Compute the value of Val on the edge BBFrom -> BBTo or the value at1731/// the basic block if the edge does not constrain Val.1732std::optional<ValueLatticeElement>1733LazyValueInfoImpl::getEdgeValue(Value *Val, BasicBlock *BBFrom,1734 BasicBlock *BBTo, Instruction *CxtI) {1735 // If already a constant, there is nothing to compute.1736 if (Constant *VC = dyn_cast<Constant>(Val))1737 return ValueLatticeElement::get(VC);1738 1739 std::optional<ValueLatticeElement> LocalResult =1740 getEdgeValueLocal(Val, BBFrom, BBTo, /*UseBlockValue*/ true);1741 if (!LocalResult)1742 return std::nullopt;1743 1744 if (hasSingleValue(*LocalResult))1745 // Can't get any more precise here1746 return LocalResult;1747 1748 std::optional<ValueLatticeElement> OptInBlock =1749 getBlockValue(Val, BBFrom, BBFrom->getTerminator());1750 if (!OptInBlock)1751 return std::nullopt;1752 ValueLatticeElement &InBlock = *OptInBlock;1753 1754 // We can use the context instruction (generically the ultimate instruction1755 // the calling pass is trying to simplify) here, even though the result of1756 // this function is generally cached when called from the solve* functions1757 // (and that cached result might be used with queries using a different1758 // context instruction), because when this function is called from the solve*1759 // functions, the context instruction is not provided. When called from1760 // LazyValueInfoImpl::getValueOnEdge, the context instruction is provided,1761 // but then the result is not cached.1762 intersectAssumeOrGuardBlockValueConstantRange(Val, InBlock, CxtI);1763 1764 return LocalResult->intersect(InBlock);1765}1766 1767ValueLatticeElement LazyValueInfoImpl::getValueInBlock(Value *V, BasicBlock *BB,1768 Instruction *CxtI) {1769 LLVM_DEBUG(dbgs() << "LVI Getting block end value " << *V << " at '"1770 << BB->getName() << "'\n");1771 1772 assert(BlockValueStack.empty() && BlockValueSet.empty());1773 std::optional<ValueLatticeElement> OptResult = getBlockValue(V, BB, CxtI);1774 if (!OptResult) {1775 solve();1776 OptResult = getBlockValue(V, BB, CxtI);1777 assert(OptResult && "Value not available after solving");1778 }1779 1780 LLVM_DEBUG(dbgs() << " Result = " << *OptResult << "\n");1781 return *OptResult;1782}1783 1784ValueLatticeElement LazyValueInfoImpl::getValueAt(Value *V, Instruction *CxtI) {1785 LLVM_DEBUG(dbgs() << "LVI Getting value " << *V << " at '" << CxtI->getName()1786 << "'\n");1787 1788 if (auto *C = dyn_cast<Constant>(V))1789 return ValueLatticeElement::get(C);1790 1791 ValueLatticeElement Result = ValueLatticeElement::getOverdefined();1792 if (auto *I = dyn_cast<Instruction>(V))1793 Result = getFromRangeMetadata(I);1794 intersectAssumeOrGuardBlockValueConstantRange(V, Result, CxtI);1795 1796 LLVM_DEBUG(dbgs() << " Result = " << Result << "\n");1797 return Result;1798}1799 1800ValueLatticeElement LazyValueInfoImpl::1801getValueOnEdge(Value *V, BasicBlock *FromBB, BasicBlock *ToBB,1802 Instruction *CxtI) {1803 LLVM_DEBUG(dbgs() << "LVI Getting edge value " << *V << " from '"1804 << FromBB->getName() << "' to '" << ToBB->getName()1805 << "'\n");1806 1807 std::optional<ValueLatticeElement> Result =1808 getEdgeValue(V, FromBB, ToBB, CxtI);1809 while (!Result) {1810 // As the worklist only explicitly tracks block values (but not edge values)1811 // we may have to call solve() multiple times, as the edge value calculation1812 // may request additional block values.1813 solve();1814 Result = getEdgeValue(V, FromBB, ToBB, CxtI);1815 }1816 1817 LLVM_DEBUG(dbgs() << " Result = " << *Result << "\n");1818 return *Result;1819}1820 1821ValueLatticeElement LazyValueInfoImpl::getValueAtUse(const Use &U) {1822 Value *V = U.get();1823 auto *CxtI = cast<Instruction>(U.getUser());1824 ValueLatticeElement VL = getValueInBlock(V, CxtI->getParent(), CxtI);1825 1826 // Check whether the only (possibly transitive) use of the value is in a1827 // position where V can be constrained by a select or branch condition.1828 const Use *CurrU = &U;1829 // TODO: Increase limit?1830 const unsigned MaxUsesToInspect = 3;1831 for (unsigned I = 0; I < MaxUsesToInspect; ++I) {1832 std::optional<ValueLatticeElement> CondVal;1833 auto *CurrI = cast<Instruction>(CurrU->getUser());1834 if (auto *SI = dyn_cast<SelectInst>(CurrI)) {1835 // If the value is undef, a different value may be chosen in1836 // the select condition and at use.1837 if (!isGuaranteedNotToBeUndef(SI->getCondition(), AC))1838 break;1839 if (CurrU->getOperandNo() == 1)1840 CondVal =1841 *getValueFromCondition(V, SI->getCondition(), /*IsTrueDest*/ true,1842 /*UseBlockValue*/ false);1843 else if (CurrU->getOperandNo() == 2)1844 CondVal =1845 *getValueFromCondition(V, SI->getCondition(), /*IsTrueDest*/ false,1846 /*UseBlockValue*/ false);1847 } else if (auto *PHI = dyn_cast<PHINode>(CurrI)) {1848 // TODO: Use non-local query?1849 CondVal = *getEdgeValueLocal(V, PHI->getIncomingBlock(*CurrU),1850 PHI->getParent(), /*UseBlockValue*/ false);1851 }1852 if (CondVal)1853 VL = VL.intersect(*CondVal);1854 1855 // Only follow one-use chain, to allow direct intersection of conditions.1856 // If there are multiple uses, we would have to intersect with the union of1857 // all conditions at different uses.1858 // Stop walking if we hit a non-speculatable instruction. Even if the1859 // result is only used under a specific condition, executing the1860 // instruction itself may cause side effects or UB already.1861 // This also disallows looking through phi nodes: If the phi node is part1862 // of a cycle, we might end up reasoning about values from different cycle1863 // iterations (PR60629).1864 if (!CurrI->hasOneUse() ||1865 !isSafeToSpeculativelyExecuteWithVariableReplaced(1866 CurrI, /*IgnoreUBImplyingAttrs=*/false))1867 break;1868 CurrU = &*CurrI->use_begin();1869 }1870 return VL;1871}1872 1873void LazyValueInfoImpl::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc,1874 BasicBlock *NewSucc) {1875 TheCache.threadEdgeImpl(OldSucc, NewSucc);1876}1877 1878//===----------------------------------------------------------------------===//1879// LazyValueInfo Impl1880//===----------------------------------------------------------------------===//1881 1882bool LazyValueInfoWrapperPass::runOnFunction(Function &F) {1883 Info.AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);1884 1885 if (auto *Impl = Info.getImpl())1886 Impl->clear();1887 1888 // Fully lazy.1889 return false;1890}1891 1892void LazyValueInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {1893 AU.setPreservesAll();1894 AU.addRequired<AssumptionCacheTracker>();1895 AU.addRequired<TargetLibraryInfoWrapperPass>();1896}1897 1898LazyValueInfo &LazyValueInfoWrapperPass::getLVI() { return Info; }1899 1900/// This lazily constructs the LazyValueInfoImpl.1901LazyValueInfoImpl &LazyValueInfo::getOrCreateImpl(const Module *M) {1902 if (!PImpl) {1903 assert(M && "getCache() called with a null Module");1904 const DataLayout &DL = M->getDataLayout();1905 Function *GuardDecl =1906 Intrinsic::getDeclarationIfExists(M, Intrinsic::experimental_guard);1907 PImpl = new LazyValueInfoImpl(AC, DL, GuardDecl);1908 }1909 return *PImpl;1910}1911 1912LazyValueInfoImpl *LazyValueInfo::getImpl() { return PImpl; }1913 1914LazyValueInfo::~LazyValueInfo() { releaseMemory(); }1915 1916void LazyValueInfo::releaseMemory() {1917 // If the cache was allocated, free it.1918 if (auto *Impl = getImpl()) {1919 delete &*Impl;1920 PImpl = nullptr;1921 }1922}1923 1924bool LazyValueInfo::invalidate(Function &F, const PreservedAnalyses &PA,1925 FunctionAnalysisManager::Invalidator &Inv) {1926 // We need to invalidate if we have either failed to preserve this analyses1927 // result directly or if any of its dependencies have been invalidated.1928 auto PAC = PA.getChecker<LazyValueAnalysis>();1929 if (!(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>()))1930 return true;1931 1932 return false;1933}1934 1935void LazyValueInfoWrapperPass::releaseMemory() { Info.releaseMemory(); }1936 1937LazyValueInfo LazyValueAnalysis::run(Function &F,1938 FunctionAnalysisManager &FAM) {1939 auto &AC = FAM.getResult<AssumptionAnalysis>(F);1940 1941 return LazyValueInfo(&AC, &F.getDataLayout());1942}1943 1944/// Returns true if we can statically tell that this value will never be a1945/// "useful" constant. In practice, this means we've got something like an1946/// alloca or a malloc call for which a comparison against a constant can1947/// only be guarding dead code. Note that we are potentially giving up some1948/// precision in dead code (a constant result) in favour of avoiding a1949/// expensive search for a easily answered common query.1950static bool isKnownNonConstant(Value *V) {1951 V = V->stripPointerCasts();1952 // The return val of alloc cannot be a Constant.1953 if (isa<AllocaInst>(V))1954 return true;1955 return false;1956}1957 1958Constant *LazyValueInfo::getConstant(Value *V, Instruction *CxtI) {1959 // Bail out early if V is known not to be a Constant.1960 if (isKnownNonConstant(V))1961 return nullptr;1962 1963 BasicBlock *BB = CxtI->getParent();1964 ValueLatticeElement Result =1965 getOrCreateImpl(BB->getModule()).getValueInBlock(V, BB, CxtI);1966 1967 if (Result.isConstant())1968 return Result.getConstant();1969 if (Result.isConstantRange()) {1970 const ConstantRange &CR = Result.getConstantRange();1971 if (const APInt *SingleVal = CR.getSingleElement())1972 return ConstantInt::get(V->getType(), *SingleVal);1973 }1974 return nullptr;1975}1976 1977ConstantRange LazyValueInfo::getConstantRange(Value *V, Instruction *CxtI,1978 bool UndefAllowed) {1979 BasicBlock *BB = CxtI->getParent();1980 ValueLatticeElement Result =1981 getOrCreateImpl(BB->getModule()).getValueInBlock(V, BB, CxtI);1982 return Result.asConstantRange(V->getType(), UndefAllowed);1983}1984 1985ConstantRange LazyValueInfo::getConstantRangeAtUse(const Use &U,1986 bool UndefAllowed) {1987 auto *Inst = cast<Instruction>(U.getUser());1988 ValueLatticeElement Result =1989 getOrCreateImpl(Inst->getModule()).getValueAtUse(U);1990 return Result.asConstantRange(U->getType(), UndefAllowed);1991}1992 1993/// Determine whether the specified value is known to be a1994/// constant on the specified edge. Return null if not.1995Constant *LazyValueInfo::getConstantOnEdge(Value *V, BasicBlock *FromBB,1996 BasicBlock *ToBB,1997 Instruction *CxtI) {1998 Module *M = FromBB->getModule();1999 ValueLatticeElement Result =2000 getOrCreateImpl(M).getValueOnEdge(V, FromBB, ToBB, CxtI);2001 2002 if (Result.isConstant())2003 return Result.getConstant();2004 if (Result.isConstantRange()) {2005 const ConstantRange &CR = Result.getConstantRange();2006 if (const APInt *SingleVal = CR.getSingleElement())2007 return ConstantInt::get(V->getType(), *SingleVal);2008 }2009 return nullptr;2010}2011 2012ConstantRange LazyValueInfo::getConstantRangeOnEdge(Value *V,2013 BasicBlock *FromBB,2014 BasicBlock *ToBB,2015 Instruction *CxtI) {2016 Module *M = FromBB->getModule();2017 ValueLatticeElement Result =2018 getOrCreateImpl(M).getValueOnEdge(V, FromBB, ToBB, CxtI);2019 // TODO: Should undef be allowed here?2020 return Result.asConstantRange(V->getType(), /*UndefAllowed*/ true);2021}2022 2023static Constant *getPredicateResult(CmpInst::Predicate Pred, Constant *C,2024 const ValueLatticeElement &Val,2025 const DataLayout &DL) {2026 // If we know the value is a constant, evaluate the conditional.2027 if (Val.isConstant())2028 return ConstantFoldCompareInstOperands(Pred, Val.getConstant(), C, DL);2029 2030 Type *ResTy = CmpInst::makeCmpResultType(C->getType());2031 if (Val.isConstantRange()) {2032 const ConstantRange &CR = Val.getConstantRange();2033 ConstantRange RHS = C->toConstantRange();2034 if (CR.icmp(Pred, RHS))2035 return ConstantInt::getTrue(ResTy);2036 if (CR.icmp(CmpInst::getInversePredicate(Pred), RHS))2037 return ConstantInt::getFalse(ResTy);2038 return nullptr;2039 }2040 2041 if (Val.isNotConstant()) {2042 // If this is an equality comparison, we can try to fold it knowing that2043 // "V != C1".2044 if (Pred == ICmpInst::ICMP_EQ) {2045 // !C1 == C -> false iff C1 == C.2046 Constant *Res = ConstantFoldCompareInstOperands(2047 ICmpInst::ICMP_NE, Val.getNotConstant(), C, DL);2048 if (Res && Res->isNullValue())2049 return ConstantInt::getFalse(ResTy);2050 } else if (Pred == ICmpInst::ICMP_NE) {2051 // !C1 != C -> true iff C1 == C.2052 Constant *Res = ConstantFoldCompareInstOperands(2053 ICmpInst::ICMP_NE, Val.getNotConstant(), C, DL);2054 if (Res && Res->isNullValue())2055 return ConstantInt::getTrue(ResTy);2056 }2057 return nullptr;2058 }2059 2060 return nullptr;2061}2062 2063/// Determine whether the specified value comparison with a constant is known to2064/// be true or false on the specified CFG edge. Pred is a CmpInst predicate.2065Constant *LazyValueInfo::getPredicateOnEdge(CmpInst::Predicate Pred, Value *V,2066 Constant *C, BasicBlock *FromBB,2067 BasicBlock *ToBB,2068 Instruction *CxtI) {2069 Module *M = FromBB->getModule();2070 ValueLatticeElement Result =2071 getOrCreateImpl(M).getValueOnEdge(V, FromBB, ToBB, CxtI);2072 2073 return getPredicateResult(Pred, C, Result, M->getDataLayout());2074}2075 2076Constant *LazyValueInfo::getPredicateAt(CmpInst::Predicate Pred, Value *V,2077 Constant *C, Instruction *CxtI,2078 bool UseBlockValue) {2079 // Is or is not NonNull are common predicates being queried. If2080 // isKnownNonZero can tell us the result of the predicate, we can2081 // return it quickly. But this is only a fastpath, and falling2082 // through would still be correct.2083 Module *M = CxtI->getModule();2084 const DataLayout &DL = M->getDataLayout();2085 if (V->getType()->isPointerTy() && C->isNullValue() &&2086 isKnownNonZero(V->stripPointerCastsSameRepresentation(), DL)) {2087 Type *ResTy = CmpInst::makeCmpResultType(C->getType());2088 if (Pred == ICmpInst::ICMP_EQ)2089 return ConstantInt::getFalse(ResTy);2090 else if (Pred == ICmpInst::ICMP_NE)2091 return ConstantInt::getTrue(ResTy);2092 }2093 2094 auto &Impl = getOrCreateImpl(M);2095 ValueLatticeElement Result =2096 UseBlockValue ? Impl.getValueInBlock(V, CxtI->getParent(), CxtI)2097 : Impl.getValueAt(V, CxtI);2098 Constant *Ret = getPredicateResult(Pred, C, Result, DL);2099 if (Ret)2100 return Ret;2101 2102 // Note: The following bit of code is somewhat distinct from the rest of LVI;2103 // LVI as a whole tries to compute a lattice value which is conservatively2104 // correct at a given location. In this case, we have a predicate which we2105 // weren't able to prove about the merged result, and we're pushing that2106 // predicate back along each incoming edge to see if we can prove it2107 // separately for each input. As a motivating example, consider:2108 // bb1:2109 // %v1 = ... ; constantrange<1, 5>2110 // br label %merge2111 // bb2:2112 // %v2 = ... ; constantrange<10, 20>2113 // br label %merge2114 // merge:2115 // %phi = phi [%v1, %v2] ; constantrange<1,20>2116 // %pred = icmp eq i32 %phi, 82117 // We can't tell from the lattice value for '%phi' that '%pred' is false2118 // along each path, but by checking the predicate over each input separately,2119 // we can.2120 // We limit the search to one step backwards from the current BB and value.2121 // We could consider extending this to search further backwards through the2122 // CFG and/or value graph, but there are non-obvious compile time vs quality2123 // tradeoffs.2124 BasicBlock *BB = CxtI->getParent();2125 2126 // Function entry or an unreachable block. Bail to avoid confusing2127 // analysis below.2128 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);2129 if (PI == PE)2130 return nullptr;2131 2132 // If V is a PHI node in the same block as the context, we need to ask2133 // questions about the predicate as applied to the incoming value along2134 // each edge. This is useful for eliminating cases where the predicate is2135 // known along all incoming edges.2136 if (auto *PHI = dyn_cast<PHINode>(V))2137 if (PHI->getParent() == BB) {2138 Constant *Baseline = nullptr;2139 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i < e; i++) {2140 Value *Incoming = PHI->getIncomingValue(i);2141 BasicBlock *PredBB = PHI->getIncomingBlock(i);2142 // Note that PredBB may be BB itself.2143 Constant *Result =2144 getPredicateOnEdge(Pred, Incoming, C, PredBB, BB, CxtI);2145 2146 // Keep going as long as we've seen a consistent known result for2147 // all inputs.2148 Baseline = (i == 0) ? Result /* First iteration */2149 : (Baseline == Result ? Baseline2150 : nullptr); /* All others */2151 if (!Baseline)2152 break;2153 }2154 if (Baseline)2155 return Baseline;2156 }2157 2158 // For a comparison where the V is outside this block, it's possible2159 // that we've branched on it before. Look to see if the value is known2160 // on all incoming edges.2161 if (!isa<Instruction>(V) || cast<Instruction>(V)->getParent() != BB) {2162 // For predecessor edge, determine if the comparison is true or false2163 // on that edge. If they're all true or all false, we can conclude2164 // the value of the comparison in this block.2165 Constant *Baseline = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI);2166 if (Baseline) {2167 // Check that all remaining incoming values match the first one.2168 while (++PI != PE) {2169 Constant *Ret = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI);2170 if (Ret != Baseline)2171 break;2172 }2173 // If we terminated early, then one of the values didn't match.2174 if (PI == PE) {2175 return Baseline;2176 }2177 }2178 }2179 2180 return nullptr;2181}2182 2183Constant *LazyValueInfo::getPredicateAt(CmpInst::Predicate Pred, Value *LHS,2184 Value *RHS, Instruction *CxtI,2185 bool UseBlockValue) {2186 if (auto *C = dyn_cast<Constant>(RHS))2187 return getPredicateAt(Pred, LHS, C, CxtI, UseBlockValue);2188 if (auto *C = dyn_cast<Constant>(LHS))2189 return getPredicateAt(CmpInst::getSwappedPredicate(Pred), RHS, C, CxtI,2190 UseBlockValue);2191 2192 // Got two non-Constant values. Try to determine the comparison results based2193 // on the block values of the two operands, e.g. because they have2194 // non-overlapping ranges.2195 if (UseBlockValue) {2196 Module *M = CxtI->getModule();2197 ValueLatticeElement L =2198 getOrCreateImpl(M).getValueInBlock(LHS, CxtI->getParent(), CxtI);2199 if (L.isOverdefined())2200 return nullptr;2201 2202 ValueLatticeElement R =2203 getOrCreateImpl(M).getValueInBlock(RHS, CxtI->getParent(), CxtI);2204 Type *Ty = CmpInst::makeCmpResultType(LHS->getType());2205 return L.getCompare(Pred, Ty, R, M->getDataLayout());2206 }2207 return nullptr;2208}2209 2210void LazyValueInfo::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc,2211 BasicBlock *NewSucc) {2212 if (auto *Impl = getImpl())2213 Impl->threadEdge(PredBB, OldSucc, NewSucc);2214}2215 2216void LazyValueInfo::forgetValue(Value *V) {2217 if (auto *Impl = getImpl())2218 Impl->forgetValue(V);2219}2220 2221void LazyValueInfo::eraseBlock(BasicBlock *BB) {2222 if (auto *Impl = getImpl())2223 Impl->eraseBlock(BB);2224}2225 2226void LazyValueInfo::clear() {2227 if (auto *Impl = getImpl())2228 Impl->clear();2229}2230 2231void LazyValueInfo::printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) {2232 if (auto *Impl = getImpl())2233 Impl->printLVI(F, DTree, OS);2234}2235 2236// Print the LVI for the function arguments at the start of each basic block.2237void LazyValueInfoAnnotatedWriter::emitBasicBlockStartAnnot(2238 const BasicBlock *BB, formatted_raw_ostream &OS) {2239 // Find if there are latticevalues defined for arguments of the function.2240 auto *F = BB->getParent();2241 for (const auto &Arg : F->args()) {2242 ValueLatticeElement Result = LVIImpl->getValueInBlock(2243 const_cast<Argument *>(&Arg), const_cast<BasicBlock *>(BB));2244 if (Result.isUnknown())2245 continue;2246 OS << "; LatticeVal for: '" << Arg << "' is: " << Result << "\n";2247 }2248}2249 2250// This function prints the LVI analysis for the instruction I at the beginning2251// of various basic blocks. It relies on calculated values that are stored in2252// the LazyValueInfoCache, and in the absence of cached values, recalculate the2253// LazyValueInfo for `I`, and print that info.2254void LazyValueInfoAnnotatedWriter::emitInstructionAnnot(2255 const Instruction *I, formatted_raw_ostream &OS) {2256 2257 auto *ParentBB = I->getParent();2258 SmallPtrSet<const BasicBlock*, 16> BlocksContainingLVI;2259 // We can generate (solve) LVI values only for blocks that are dominated by2260 // the I's parent. However, to avoid generating LVI for all dominating blocks,2261 // that contain redundant/uninteresting information, we print LVI for2262 // blocks that may use this LVI information (such as immediate successor2263 // blocks, and blocks that contain uses of `I`).2264 auto printResult = [&](const BasicBlock *BB) {2265 if (!BlocksContainingLVI.insert(BB).second)2266 return;2267 ValueLatticeElement Result = LVIImpl->getValueInBlock(2268 const_cast<Instruction *>(I), const_cast<BasicBlock *>(BB));2269 OS << "; LatticeVal for: '" << *I << "' in BB: '";2270 BB->printAsOperand(OS, false);2271 OS << "' is: " << Result << "\n";2272 };2273 2274 printResult(ParentBB);2275 // Print the LVI analysis results for the immediate successor blocks, that2276 // are dominated by `ParentBB`.2277 for (const auto *BBSucc : successors(ParentBB))2278 if (DT.dominates(ParentBB, BBSucc))2279 printResult(BBSucc);2280 2281 // Print LVI in blocks where `I` is used.2282 for (const auto *U : I->users())2283 if (auto *UseI = dyn_cast<Instruction>(U))2284 if (!isa<PHINode>(UseI) || DT.dominates(ParentBB, UseI->getParent()))2285 printResult(UseI->getParent());2286 2287}2288 2289PreservedAnalyses LazyValueInfoPrinterPass::run(Function &F,2290 FunctionAnalysisManager &AM) {2291 OS << "LVI for function '" << F.getName() << "':\n";2292 auto &LVI = AM.getResult<LazyValueAnalysis>(F);2293 auto &DTree = AM.getResult<DominatorTreeAnalysis>(F);2294 LVI.printLVI(F, DTree, OS);2295 return PreservedAnalyses::all();2296}2297