| 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-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file defines the interface for lazy computation of value constraint |
| 10 | // 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/AssumeBundleQueries.h" |
| 18 | #include "llvm/Analysis/AssumptionCache.h" |
| 19 | #include "llvm/Analysis/ConstantFolding.h" |
| 20 | #include "llvm/Analysis/InstructionSimplify.h" |
| 21 | #include "llvm/Analysis/Passes.h" |
| 22 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 23 | #include "llvm/Analysis/ValueLattice.h" |
| 24 | #include "llvm/Analysis/ValueTracking.h" |
| 25 | #include "llvm/IR/AssemblyAnnotationWriter.h" |
| 26 | #include "llvm/IR/CFG.h" |
| 27 | #include "llvm/IR/ConstantRange.h" |
| 28 | #include "llvm/IR/Constants.h" |
| 29 | #include "llvm/IR/DataLayout.h" |
| 30 | #include "llvm/IR/Dominators.h" |
| 31 | #include "llvm/IR/InstrTypes.h" |
| 32 | #include "llvm/IR/Instructions.h" |
| 33 | #include "llvm/IR/IntrinsicInst.h" |
| 34 | #include "llvm/IR/Intrinsics.h" |
| 35 | #include "llvm/IR/LLVMContext.h" |
| 36 | #include "llvm/IR/Module.h" |
| 37 | #include "llvm/IR/PatternMatch.h" |
| 38 | #include "llvm/IR/ValueHandle.h" |
| 39 | #include "llvm/InitializePasses.h" |
| 40 | #include "llvm/Support/Debug.h" |
| 41 | #include "llvm/Support/FormattedStream.h" |
| 42 | #include "llvm/Support/KnownBits.h" |
| 43 | #include "llvm/Support/raw_ostream.h" |
| 44 | #include <optional> |
| 45 | using namespace llvm; |
| 46 | using namespace PatternMatch; |
| 47 | |
| 48 | #define DEBUG_TYPE "lazy-value-info" |
| 49 | |
| 50 | // This is the number of worklist items we will process to try to discover an |
| 51 | // answer for a given value. |
| 52 | static const unsigned MaxProcessedPerValue = 500; |
| 53 | |
| 54 | char LazyValueInfoWrapperPass::ID = 0; |
| 55 | LazyValueInfoWrapperPass::LazyValueInfoWrapperPass() : FunctionPass(ID) {} |
| 56 | INITIALIZE_PASS_BEGIN(LazyValueInfoWrapperPass, "lazy-value-info" , |
| 57 | "Lazy Value Information Analysis" , false, true) |
| 58 | INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) |
| 59 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) |
| 60 | INITIALIZE_PASS_END(LazyValueInfoWrapperPass, "lazy-value-info" , |
| 61 | "Lazy Value Information Analysis" , false, true) |
| 62 | |
| 63 | static cl::opt<bool> PerPredRanges( |
| 64 | "lvi-per-pred-ranges" , cl::Hidden, cl::init(Val: false), |
| 65 | cl::desc("Enable tracking of ranges for a value in a block for" |
| 66 | "each block predecessor (default = false)" )); |
| 67 | |
| 68 | namespace llvm { |
| 69 | FunctionPass *createLazyValueInfoPass() { |
| 70 | return new LazyValueInfoWrapperPass(); |
| 71 | } |
| 72 | } // namespace llvm |
| 73 | |
| 74 | AnalysisKey LazyValueAnalysis::Key; |
| 75 | |
| 76 | /// Returns true if this lattice value represents at most one possible value. |
| 77 | /// This is as precise as any lattice value can get while still representing |
| 78 | /// reachable code. |
| 79 | static bool hasSingleValue(const ValueLatticeElement &Val) { |
| 80 | if (Val.isConstantRange() && |
| 81 | Val.getConstantRange().isSingleElement()) |
| 82 | // Integer constants are single element ranges |
| 83 | return true; |
| 84 | if (Val.isConstant()) |
| 85 | // Non integer constants |
| 86 | return true; |
| 87 | return false; |
| 88 | } |
| 89 | |
| 90 | //===----------------------------------------------------------------------===// |
| 91 | // LazyValueInfoCache Decl |
| 92 | //===----------------------------------------------------------------------===// |
| 93 | |
| 94 | namespace { |
| 95 | /// A callback value handle updates the cache when values are erased. |
| 96 | class LazyValueInfoCache; |
| 97 | struct LVIValueHandle final : public CallbackVH { |
| 98 | LazyValueInfoCache *Parent; |
| 99 | |
| 100 | LVIValueHandle(Value *V, LazyValueInfoCache *P = nullptr) |
| 101 | : CallbackVH(V), Parent(P) { } |
| 102 | |
| 103 | void deleted() override; |
| 104 | void allUsesReplacedWith(Value *V) override { |
| 105 | deleted(); |
| 106 | } |
| 107 | }; |
| 108 | } // end anonymous namespace |
| 109 | |
| 110 | namespace { |
| 111 | using NonNullPointerSet = SmallDenseSet<AssertingVH<Value>, 2>; |
| 112 | using BBLatticeElementMap = |
| 113 | SmallDenseMap<PoisoningVH<BasicBlock>, ValueLatticeElement, 4>; |
| 114 | using PredecessorValueLatticeMap = |
| 115 | SmallDenseMap<AssertingVH<Value>, BBLatticeElementMap, 2>; |
| 116 | |
| 117 | /// This is the cache kept by LazyValueInfo which |
| 118 | /// maintains information about queries across the clients' queries. |
| 119 | class LazyValueInfoCache { |
| 120 | /// This is all of the cached information for one basic block. It contains |
| 121 | /// the per-value lattice elements, as well as a separate set for |
| 122 | /// overdefined values to reduce memory usage. Additionally pointers |
| 123 | /// dereferenced in the block are cached for nullability queries. |
| 124 | struct BlockCacheEntry { |
| 125 | SmallDenseMap<AssertingVH<Value>, ValueLatticeElement, 4> LatticeElements; |
| 126 | SmallDenseSet<AssertingVH<Value>, 4> OverDefined; |
| 127 | // std::nullopt indicates that the nonnull pointers for this basic block |
| 128 | // block have not been computed yet. |
| 129 | std::optional<NonNullPointerSet> NonNullPointers; |
| 130 | // This is an extension of the above LatticeElements, caching, for each |
| 131 | // Value, a ValueLatticeElement, for each predecessor of the BB tracked by |
| 132 | // this entry. |
| 133 | std::optional<PredecessorValueLatticeMap> PredecessorLatticeElements; |
| 134 | }; |
| 135 | |
| 136 | /// Cached information per basic block. |
| 137 | DenseMap<PoisoningVH<BasicBlock>, std::unique_ptr<BlockCacheEntry>> |
| 138 | BlockCache; |
| 139 | /// Set of value handles used to erase values from the cache on deletion. |
| 140 | DenseSet<LVIValueHandle, DenseMapInfo<Value *>> ValueHandles; |
| 141 | |
| 142 | const BlockCacheEntry *getBlockEntry(BasicBlock *BB) const { |
| 143 | auto It = BlockCache.find_as(Val: BB); |
| 144 | if (It == BlockCache.end()) |
| 145 | return nullptr; |
| 146 | return It->second.get(); |
| 147 | } |
| 148 | |
| 149 | BlockCacheEntry *getOrCreateBlockEntry(BasicBlock *BB) { |
| 150 | auto It = BlockCache.find_as(Val: BB); |
| 151 | if (It == BlockCache.end()) { |
| 152 | std::unique_ptr<BlockCacheEntry> BCE = |
| 153 | std::make_unique<BlockCacheEntry>(); |
| 154 | if (PerPredRanges) |
| 155 | BCE->PredecessorLatticeElements = |
| 156 | std::make_optional<PredecessorValueLatticeMap>(); |
| 157 | It = BlockCache.insert(KV: {BB, std::move(BCE)}).first; |
| 158 | } |
| 159 | |
| 160 | return It->second.get(); |
| 161 | } |
| 162 | |
| 163 | void addValueHandle(Value *Val) { |
| 164 | auto HandleIt = ValueHandles.find_as(Val); |
| 165 | if (HandleIt == ValueHandles.end()) |
| 166 | ValueHandles.insert(V: {Val, this}); |
| 167 | } |
| 168 | |
| 169 | public: |
| 170 | void insertResult(Value *Val, BasicBlock *BB, |
| 171 | const ValueLatticeElement &Result) { |
| 172 | BlockCacheEntry *Entry = getOrCreateBlockEntry(BB); |
| 173 | |
| 174 | // Insert over-defined values into their own cache to reduce memory |
| 175 | // overhead. |
| 176 | if (Result.isOverdefined()) |
| 177 | Entry->OverDefined.insert(V: Val); |
| 178 | else |
| 179 | Entry->LatticeElements.insert(KV: {Val, Result}); |
| 180 | |
| 181 | addValueHandle(Val); |
| 182 | } |
| 183 | |
| 184 | void insertPredecessorResults(Value *Val, BasicBlock *BB, |
| 185 | BBLatticeElementMap &PredLatticeElements) { |
| 186 | BlockCacheEntry *Entry = getOrCreateBlockEntry(BB); |
| 187 | |
| 188 | Entry->PredecessorLatticeElements->insert(KV: {Val, PredLatticeElements}); |
| 189 | |
| 190 | addValueHandle(Val); |
| 191 | } |
| 192 | |
| 193 | std::optional<BBLatticeElementMap> |
| 194 | getCachedPredecessorInfo(Value *V, BasicBlock *BB) const { |
| 195 | const BlockCacheEntry *Entry = getBlockEntry(BB); |
| 196 | if (!Entry) |
| 197 | return std::nullopt; |
| 198 | |
| 199 | auto LatticeIt = Entry->PredecessorLatticeElements->find_as(Val: V); |
| 200 | if (LatticeIt == Entry->PredecessorLatticeElements->end()) |
| 201 | return std::nullopt; |
| 202 | |
| 203 | return LatticeIt->second; |
| 204 | } |
| 205 | |
| 206 | std::optional<ValueLatticeElement> getCachedValueInfo(Value *V, |
| 207 | BasicBlock *BB) const { |
| 208 | const BlockCacheEntry *Entry = getBlockEntry(BB); |
| 209 | if (!Entry) |
| 210 | return std::nullopt; |
| 211 | |
| 212 | if (Entry->OverDefined.count(V)) |
| 213 | return ValueLatticeElement::getOverdefined(); |
| 214 | |
| 215 | auto LatticeIt = Entry->LatticeElements.find_as(Val: V); |
| 216 | if (LatticeIt == Entry->LatticeElements.end()) |
| 217 | return std::nullopt; |
| 218 | |
| 219 | return LatticeIt->second; |
| 220 | } |
| 221 | |
| 222 | bool |
| 223 | isNonNullAtEndOfBlock(Value *V, BasicBlock *BB, |
| 224 | function_ref<NonNullPointerSet(BasicBlock *)> InitFn) { |
| 225 | BlockCacheEntry *Entry = getOrCreateBlockEntry(BB); |
| 226 | if (!Entry->NonNullPointers) { |
| 227 | Entry->NonNullPointers = InitFn(BB); |
| 228 | for (Value *V : *Entry->NonNullPointers) |
| 229 | addValueHandle(Val: V); |
| 230 | } |
| 231 | |
| 232 | return Entry->NonNullPointers->count(V); |
| 233 | } |
| 234 | |
| 235 | /// clear - Empty the cache. |
| 236 | void clear() { |
| 237 | BlockCache.clear(); |
| 238 | ValueHandles.clear(); |
| 239 | } |
| 240 | |
| 241 | /// Inform the cache that a given value has been deleted. |
| 242 | void eraseValue(Value *V); |
| 243 | |
| 244 | /// This is part of the update interface to inform the cache |
| 245 | /// that a block has been deleted. |
| 246 | void eraseBlock(BasicBlock *BB); |
| 247 | |
| 248 | /// Updates the cache to remove any influence an overdefined value in |
| 249 | /// OldSucc might have (unless also overdefined in NewSucc). This just |
| 250 | /// flushes elements from the cache and does not add any. |
| 251 | void threadEdgeImpl(BasicBlock *OldSucc, BasicBlock *NewSucc); |
| 252 | }; |
| 253 | } // namespace |
| 254 | |
| 255 | void LazyValueInfoCache::eraseValue(Value *V) { |
| 256 | for (auto &Pair : BlockCache) { |
| 257 | Pair.second->LatticeElements.erase(Val: V); |
| 258 | Pair.second->OverDefined.erase(V); |
| 259 | if (Pair.second->NonNullPointers) |
| 260 | Pair.second->NonNullPointers->erase(V); |
| 261 | if (PerPredRanges) |
| 262 | Pair.second->PredecessorLatticeElements->erase(Val: V); |
| 263 | } |
| 264 | |
| 265 | auto HandleIt = ValueHandles.find_as(Val: V); |
| 266 | if (HandleIt != ValueHandles.end()) |
| 267 | ValueHandles.erase(I: HandleIt); |
| 268 | } |
| 269 | |
| 270 | void LVIValueHandle::deleted() { |
| 271 | // This erasure deallocates *this, so it MUST happen after we're done |
| 272 | // using any and all members of *this. |
| 273 | Parent->eraseValue(V: *this); |
| 274 | } |
| 275 | |
| 276 | void LazyValueInfoCache::eraseBlock(BasicBlock *BB) { |
| 277 | // Clear all when a BB is removed. |
| 278 | if (PerPredRanges) |
| 279 | for (auto &Pair : BlockCache) |
| 280 | Pair.second->PredecessorLatticeElements->clear(); |
| 281 | BlockCache.erase(Val: BB); |
| 282 | } |
| 283 | |
| 284 | void LazyValueInfoCache::threadEdgeImpl(BasicBlock *OldSucc, |
| 285 | BasicBlock *NewSucc) { |
| 286 | // When an edge in the graph has been threaded, values that we could not |
| 287 | // determine a value for before (i.e. were marked overdefined) may be |
| 288 | // possible to solve now. We do NOT try to proactively update these values. |
| 289 | // Instead, we clear their entries from the cache, and allow lazy updating to |
| 290 | // recompute them when needed. |
| 291 | |
| 292 | // The updating process is fairly simple: we need to drop cached info |
| 293 | // for all values that were marked overdefined in OldSucc, and for those same |
| 294 | // values in any successor of OldSucc (except NewSucc) in which they were |
| 295 | // also marked overdefined. |
| 296 | std::vector<BasicBlock*> worklist; |
| 297 | worklist.push_back(x: OldSucc); |
| 298 | |
| 299 | const BlockCacheEntry *Entry = getBlockEntry(BB: OldSucc); |
| 300 | if (!Entry || Entry->OverDefined.empty()) |
| 301 | return; // Nothing to process here. |
| 302 | SmallVector<Value *, 4> ValsToClear(Entry->OverDefined.begin(), |
| 303 | Entry->OverDefined.end()); |
| 304 | |
| 305 | // Use a worklist to perform a depth-first search of OldSucc's successors. |
| 306 | // NOTE: We do not need a visited list since any blocks we have already |
| 307 | // visited will have had their overdefined markers cleared already, and we |
| 308 | // thus won't loop to their successors. |
| 309 | while (!worklist.empty()) { |
| 310 | BasicBlock *ToUpdate = worklist.back(); |
| 311 | worklist.pop_back(); |
| 312 | |
| 313 | // Skip blocks only accessible through NewSucc. |
| 314 | if (ToUpdate == NewSucc) continue; |
| 315 | |
| 316 | // If a value was marked overdefined in OldSucc, and is here too... |
| 317 | auto OI = BlockCache.find_as(Val: ToUpdate); |
| 318 | if (OI == BlockCache.end() || OI->second->OverDefined.empty()) |
| 319 | continue; |
| 320 | auto &ValueSet = OI->second->OverDefined; |
| 321 | |
| 322 | bool changed = false; |
| 323 | for (Value *V : ValsToClear) { |
| 324 | if (!ValueSet.erase(V)) |
| 325 | continue; |
| 326 | |
| 327 | // If we removed anything, then we potentially need to update |
| 328 | // blocks successors too. |
| 329 | changed = true; |
| 330 | } |
| 331 | |
| 332 | if (!changed) continue; |
| 333 | |
| 334 | llvm::append_range(C&: worklist, R: successors(BB: ToUpdate)); |
| 335 | } |
| 336 | } |
| 337 | |
| 338 | namespace llvm { |
| 339 | namespace { |
| 340 | /// An assembly annotator class to print LazyValueCache information in |
| 341 | /// comments. |
| 342 | class LazyValueInfoAnnotatedWriter : public AssemblyAnnotationWriter { |
| 343 | LazyValueInfoImpl *LVIImpl; |
| 344 | // While analyzing which blocks we can solve values for, we need the dominator |
| 345 | // information. |
| 346 | DominatorTree &DT; |
| 347 | |
| 348 | public: |
| 349 | LazyValueInfoAnnotatedWriter(LazyValueInfoImpl *L, DominatorTree &DTree) |
| 350 | : LVIImpl(L), DT(DTree) {} |
| 351 | |
| 352 | void emitBasicBlockStartAnnot(const BasicBlock *BB, |
| 353 | formatted_raw_ostream &OS) override; |
| 354 | |
| 355 | void emitInstructionAnnot(const Instruction *I, |
| 356 | formatted_raw_ostream &OS) override; |
| 357 | }; |
| 358 | } // namespace |
| 359 | // The actual implementation of the lazy analysis and update. |
| 360 | class LazyValueInfoImpl { |
| 361 | |
| 362 | /// Cached results from previous queries |
| 363 | LazyValueInfoCache TheCache; |
| 364 | |
| 365 | /// This stack holds the state of the value solver during a query. |
| 366 | /// It basically emulates the callstack of the naive |
| 367 | /// recursive value lookup process. |
| 368 | SmallVector<std::pair<BasicBlock*, Value*>, 8> BlockValueStack; |
| 369 | |
| 370 | /// Keeps track of which block-value pairs are in BlockValueStack. |
| 371 | DenseSet<std::pair<BasicBlock*, Value*> > BlockValueSet; |
| 372 | |
| 373 | /// Push BV onto BlockValueStack unless it's already in there. |
| 374 | /// Returns true on success. |
| 375 | bool pushBlockValue(const std::pair<BasicBlock *, Value *> &BV) { |
| 376 | if (!BlockValueSet.insert(V: BV).second) |
| 377 | return false; // It's already in the stack. |
| 378 | |
| 379 | LLVM_DEBUG(dbgs() << "PUSH: " << *BV.second << " in " |
| 380 | << BV.first->getName() << "\n" ); |
| 381 | BlockValueStack.push_back(Elt: BV); |
| 382 | return true; |
| 383 | } |
| 384 | |
| 385 | AssumptionCache *AC; ///< A pointer to the cache of @llvm.assume calls. |
| 386 | const DataLayout &DL; ///< A mandatory DataLayout |
| 387 | |
| 388 | /// Declaration of the llvm.experimental.guard() intrinsic, |
| 389 | /// if it exists in the module. |
| 390 | Function *GuardDecl; |
| 391 | |
| 392 | std::optional<ValueLatticeElement> getBlockValue(Value *Val, BasicBlock *BB, |
| 393 | Instruction *CxtI); |
| 394 | std::optional<ValueLatticeElement> getEdgeValue(Value *V, BasicBlock *F, |
| 395 | BasicBlock *T, |
| 396 | Instruction *CxtI = nullptr); |
| 397 | |
| 398 | // These methods process one work item and may add more. A false value |
| 399 | // returned means that the work item was not completely processed and must |
| 400 | // be revisited after going through the new items. |
| 401 | bool solveBlockValue(Value *Val, BasicBlock *BB); |
| 402 | std::optional<ValueLatticeElement> solveBlockValueImpl(Value *Val, |
| 403 | BasicBlock *BB); |
| 404 | std::optional<ValueLatticeElement> solveBlockValueNonLocal(Value *Val, |
| 405 | BasicBlock *BB); |
| 406 | std::optional<ValueLatticeElement> solveBlockValuePHINode(PHINode *PN, |
| 407 | BasicBlock *BB); |
| 408 | std::optional<ValueLatticeElement> solveBlockValueSelect(SelectInst *S, |
| 409 | BasicBlock *BB); |
| 410 | std::optional<ConstantRange> getRangeFor(Value *V, Instruction *CxtI, |
| 411 | BasicBlock *BB); |
| 412 | std::optional<ValueLatticeElement> solveBlockValueBinaryOpImpl( |
| 413 | Instruction *I, BasicBlock *BB, |
| 414 | std::function<ConstantRange(const ConstantRange &, const ConstantRange &)> |
| 415 | OpFn); |
| 416 | std::optional<ValueLatticeElement> |
| 417 | solveBlockValueBinaryOp(BinaryOperator *BBI, BasicBlock *BB); |
| 418 | std::optional<ValueLatticeElement> solveBlockValueCast(CastInst *CI, |
| 419 | BasicBlock *BB); |
| 420 | std::optional<ValueLatticeElement> |
| 421 | solveBlockValueOverflowIntrinsic(WithOverflowInst *WO, BasicBlock *BB); |
| 422 | std::optional<ValueLatticeElement> solveBlockValueIntrinsic(IntrinsicInst *II, |
| 423 | BasicBlock *BB); |
| 424 | std::optional<ValueLatticeElement> |
| 425 | solveBlockValueInsertElement(InsertElementInst *IEI, BasicBlock *BB); |
| 426 | std::optional<ValueLatticeElement> |
| 427 | solveBlockValueExtractValue(ExtractValueInst *EVI, BasicBlock *BB); |
| 428 | bool isNonNullAtEndOfBlock(Value *Val, BasicBlock *BB); |
| 429 | void intersectAssumeOrGuardBlockValueConstantRange(Value *Val, |
| 430 | ValueLatticeElement &BBLV, |
| 431 | Instruction *BBI); |
| 432 | |
| 433 | void solve(); |
| 434 | |
| 435 | // For the following methods, if UseBlockValue is true, the function may |
| 436 | // push additional values to the worklist and return nullopt. If |
| 437 | // UseBlockValue is false, it will never return nullopt. |
| 438 | |
| 439 | std::optional<ValueLatticeElement> |
| 440 | getValueFromSimpleICmpCondition(CmpInst::Predicate Pred, Value *RHS, |
| 441 | const APInt &Offset, Instruction *CxtI, |
| 442 | bool UseBlockValue); |
| 443 | |
| 444 | std::optional<ValueLatticeElement> |
| 445 | getValueFromICmpCondition(Value *Val, ICmpInst *ICI, bool isTrueDest, |
| 446 | bool UseBlockValue); |
| 447 | ValueLatticeElement getValueFromTrunc(Value *Val, TruncInst *Trunc, |
| 448 | bool IsTrueDest); |
| 449 | |
| 450 | std::optional<ValueLatticeElement> |
| 451 | getValueFromCondition(Value *Val, Value *Cond, bool IsTrueDest, |
| 452 | bool UseBlockValue, unsigned Depth = 0); |
| 453 | |
| 454 | std::optional<ValueLatticeElement> getEdgeValueLocal(Value *Val, |
| 455 | BasicBlock *BBFrom, |
| 456 | BasicBlock *BBTo, |
| 457 | bool UseBlockValue); |
| 458 | |
| 459 | public: |
| 460 | /// This is the query interface to determine the lattice value for the |
| 461 | /// specified Value* at the context instruction (if specified) or at the |
| 462 | /// start of the block. |
| 463 | ValueLatticeElement getValueInBlock(Value *V, BasicBlock *BB, |
| 464 | Instruction *CxtI = nullptr); |
| 465 | |
| 466 | /// This is the query interface to determine the lattice value for the |
| 467 | /// specified Value* at the specified instruction using only information |
| 468 | /// from assumes/guards and range metadata. Unlike getValueInBlock(), no |
| 469 | /// recursive query is performed. |
| 470 | ValueLatticeElement getValueAt(Value *V, Instruction *CxtI); |
| 471 | |
| 472 | /// This is the query interface to determine the lattice |
| 473 | /// value for the specified Value* that is true on the specified edge. |
| 474 | ValueLatticeElement getValueOnEdge(Value *V, BasicBlock *FromBB, |
| 475 | BasicBlock *ToBB, |
| 476 | Instruction *CxtI = nullptr); |
| 477 | |
| 478 | ValueLatticeElement getValueAtUse(const Use &U); |
| 479 | |
| 480 | /// Complete flush all previously computed values |
| 481 | void clear() { |
| 482 | TheCache.clear(); |
| 483 | } |
| 484 | |
| 485 | /// Printing the LazyValueInfo Analysis. |
| 486 | void printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) { |
| 487 | LazyValueInfoAnnotatedWriter Writer(this, DTree); |
| 488 | F.print(OS, AAW: &Writer); |
| 489 | } |
| 490 | |
| 491 | /// This is part of the update interface to remove information related to this |
| 492 | /// value from the cache. |
| 493 | void forgetValue(Value *V) { TheCache.eraseValue(V); } |
| 494 | |
| 495 | /// This is part of the update interface to inform the cache |
| 496 | /// that a block has been deleted. |
| 497 | void eraseBlock(BasicBlock *BB) { |
| 498 | TheCache.eraseBlock(BB); |
| 499 | } |
| 500 | |
| 501 | /// This is the update interface to inform the cache that an edge from |
| 502 | /// PredBB to OldSucc has been threaded to be from PredBB to NewSucc. |
| 503 | void threadEdge(BasicBlock *PredBB,BasicBlock *OldSucc,BasicBlock *NewSucc); |
| 504 | |
| 505 | LazyValueInfoImpl(AssumptionCache *AC, const DataLayout &DL, |
| 506 | Function *GuardDecl) |
| 507 | : AC(AC), DL(DL), GuardDecl(GuardDecl) {} |
| 508 | }; |
| 509 | } // namespace llvm |
| 510 | |
| 511 | void LazyValueInfoImpl::solve() { |
| 512 | SmallVector<std::pair<BasicBlock *, Value *>, 8> StartingStack = |
| 513 | BlockValueStack; |
| 514 | |
| 515 | unsigned processedCount = 0; |
| 516 | while (!BlockValueStack.empty()) { |
| 517 | processedCount++; |
| 518 | // Abort if we have to process too many values to get a result for this one. |
| 519 | // Because of the design of the overdefined cache currently being per-block |
| 520 | // to avoid naming-related issues (IE it wants to try to give different |
| 521 | // results for the same name in different blocks), overdefined results don't |
| 522 | // get cached globally, which in turn means we will often try to rediscover |
| 523 | // the same overdefined result again and again. Once something like |
| 524 | // PredicateInfo is used in LVI or CVP, we should be able to make the |
| 525 | // overdefined cache global, and remove this throttle. |
| 526 | if (processedCount > MaxProcessedPerValue) { |
| 527 | LLVM_DEBUG( |
| 528 | dbgs() << "Giving up on stack because we are getting too deep\n" ); |
| 529 | // Fill in the original values |
| 530 | while (!StartingStack.empty()) { |
| 531 | std::pair<BasicBlock *, Value *> &e = StartingStack.back(); |
| 532 | TheCache.insertResult(Val: e.second, BB: e.first, |
| 533 | Result: ValueLatticeElement::getOverdefined()); |
| 534 | StartingStack.pop_back(); |
| 535 | } |
| 536 | BlockValueSet.clear(); |
| 537 | BlockValueStack.clear(); |
| 538 | return; |
| 539 | } |
| 540 | std::pair<BasicBlock *, Value *> e = BlockValueStack.back(); |
| 541 | assert(BlockValueSet.count(e) && "Stack value should be in BlockValueSet!" ); |
| 542 | unsigned StackSize = BlockValueStack.size(); |
| 543 | (void) StackSize; |
| 544 | |
| 545 | if (solveBlockValue(Val: e.second, BB: e.first)) { |
| 546 | // The work item was completely processed. |
| 547 | assert(BlockValueStack.size() == StackSize && |
| 548 | BlockValueStack.back() == e && "Nothing should have been pushed!" ); |
| 549 | #ifndef NDEBUG |
| 550 | std::optional<ValueLatticeElement> BBLV = |
| 551 | TheCache.getCachedValueInfo(e.second, e.first); |
| 552 | assert(BBLV && "Result should be in cache!" ); |
| 553 | LLVM_DEBUG( |
| 554 | dbgs() << "POP " << *e.second << " in " << e.first->getName() << " = " |
| 555 | << *BBLV << "\n" ); |
| 556 | #endif |
| 557 | |
| 558 | BlockValueStack.pop_back(); |
| 559 | BlockValueSet.erase(V: e); |
| 560 | } else { |
| 561 | // More work needs to be done before revisiting. |
| 562 | assert(BlockValueStack.size() == StackSize + 1 && |
| 563 | "Exactly one element should have been pushed!" ); |
| 564 | } |
| 565 | } |
| 566 | } |
| 567 | |
| 568 | std::optional<ValueLatticeElement> |
| 569 | LazyValueInfoImpl::getBlockValue(Value *Val, BasicBlock *BB, |
| 570 | Instruction *CxtI) { |
| 571 | // If already a constant, there is nothing to compute. |
| 572 | if (Constant *VC = dyn_cast<Constant>(Val)) |
| 573 | return ValueLatticeElement::get(C: VC); |
| 574 | |
| 575 | if (std::optional<ValueLatticeElement> OptLatticeVal = |
| 576 | TheCache.getCachedValueInfo(V: Val, BB)) { |
| 577 | intersectAssumeOrGuardBlockValueConstantRange(Val, BBLV&: *OptLatticeVal, BBI: CxtI); |
| 578 | return OptLatticeVal; |
| 579 | } |
| 580 | |
| 581 | // We have hit a cycle, assume overdefined. |
| 582 | if (!pushBlockValue(BV: { BB, Val })) |
| 583 | return ValueLatticeElement::getOverdefined(); |
| 584 | |
| 585 | // Yet to be resolved. |
| 586 | return std::nullopt; |
| 587 | } |
| 588 | |
| 589 | static ValueLatticeElement getFromRangeMetadata(Instruction *BBI) { |
| 590 | switch (BBI->getOpcode()) { |
| 591 | default: |
| 592 | break; |
| 593 | case Instruction::Call: |
| 594 | case Instruction::Invoke: |
| 595 | if (std::optional<ConstantRange> Range = cast<CallBase>(Val: BBI)->getRange()) |
| 596 | return ValueLatticeElement::getRange(CR: *Range); |
| 597 | [[fallthrough]]; |
| 598 | case Instruction::Load: |
| 599 | if (MDNode *Ranges = BBI->getMetadata(KindID: LLVMContext::MD_range)) |
| 600 | if (isa<IntegerType>(Val: BBI->getType())) { |
| 601 | return ValueLatticeElement::getRange( |
| 602 | CR: getConstantRangeFromMetadata(RangeMD: *Ranges)); |
| 603 | } |
| 604 | break; |
| 605 | }; |
| 606 | // Nothing known - will be intersected with other facts |
| 607 | return ValueLatticeElement::getOverdefined(); |
| 608 | } |
| 609 | |
| 610 | bool LazyValueInfoImpl::solveBlockValue(Value *Val, BasicBlock *BB) { |
| 611 | assert(!isa<Constant>(Val) && "Value should not be constant" ); |
| 612 | assert(!TheCache.getCachedValueInfo(Val, BB) && |
| 613 | "Value should not be in cache" ); |
| 614 | |
| 615 | // Hold off inserting this value into the Cache in case we have to return |
| 616 | // false and come back later. |
| 617 | std::optional<ValueLatticeElement> Res = solveBlockValueImpl(Val, BB); |
| 618 | if (!Res) |
| 619 | // Work pushed, will revisit |
| 620 | return false; |
| 621 | |
| 622 | TheCache.insertResult(Val, BB, Result: *Res); |
| 623 | return true; |
| 624 | } |
| 625 | |
| 626 | std::optional<ValueLatticeElement> |
| 627 | LazyValueInfoImpl::solveBlockValueImpl(Value *Val, BasicBlock *BB) { |
| 628 | Instruction *BBI = dyn_cast<Instruction>(Val); |
| 629 | if (!BBI || BBI->getParent() != BB) |
| 630 | return solveBlockValueNonLocal(Val, BB); |
| 631 | |
| 632 | if (PHINode *PN = dyn_cast<PHINode>(Val: BBI)) |
| 633 | return solveBlockValuePHINode(PN, BB); |
| 634 | |
| 635 | if (auto *SI = dyn_cast<SelectInst>(Val: BBI)) |
| 636 | return solveBlockValueSelect(S: SI, BB); |
| 637 | |
| 638 | // If this value is a nonnull pointer, record it's range and bailout. Note |
| 639 | // that for all other pointer typed values, we terminate the search at the |
| 640 | // definition. We could easily extend this to look through geps, bitcasts, |
| 641 | // and the like to prove non-nullness, but it's not clear that's worth it |
| 642 | // compile time wise. The context-insensitive value walk done inside |
| 643 | // isKnownNonZero gets most of the profitable cases at much less expense. |
| 644 | // This does mean that we have a sensitivity to where the defining |
| 645 | // instruction is placed, even if it could legally be hoisted much higher. |
| 646 | // That is unfortunate. |
| 647 | PointerType *PT = dyn_cast<PointerType>(Val: BBI->getType()); |
| 648 | if (PT && isKnownNonZero(V: BBI, Q: DL)) |
| 649 | return ValueLatticeElement::getNot(C: ConstantPointerNull::get(T: PT)); |
| 650 | |
| 651 | if (BBI->getType()->isIntOrIntVectorTy()) { |
| 652 | if (auto *CI = dyn_cast<CastInst>(Val: BBI)) |
| 653 | return solveBlockValueCast(CI, BB); |
| 654 | |
| 655 | if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Val: BBI)) |
| 656 | return solveBlockValueBinaryOp(BBI: BO, BB); |
| 657 | |
| 658 | if (auto *IEI = dyn_cast<InsertElementInst>(Val: BBI)) |
| 659 | return solveBlockValueInsertElement(IEI, BB); |
| 660 | |
| 661 | if (auto *EVI = dyn_cast<ExtractValueInst>(Val: BBI)) |
| 662 | return solveBlockValueExtractValue(EVI, BB); |
| 663 | |
| 664 | if (auto *II = dyn_cast<IntrinsicInst>(Val: BBI)) |
| 665 | return solveBlockValueIntrinsic(II, BB); |
| 666 | } |
| 667 | |
| 668 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() |
| 669 | << "' - unknown inst def found.\n" ); |
| 670 | return getFromRangeMetadata(BBI); |
| 671 | } |
| 672 | |
| 673 | static void AddNonNullPointer(Value *Ptr, NonNullPointerSet &PtrSet, |
| 674 | bool IsDereferenced = true) { |
| 675 | // TODO: Use NullPointerIsDefined instead. |
| 676 | if (Ptr->getType()->getPointerAddressSpace() == 0) |
| 677 | PtrSet.insert(V: IsDereferenced ? getUnderlyingObject(V: Ptr) |
| 678 | : Ptr->stripInBoundsOffsets()); |
| 679 | } |
| 680 | |
| 681 | static void AddNonNullPointersByInstruction( |
| 682 | Instruction *I, NonNullPointerSet &PtrSet) { |
| 683 | if (LoadInst *L = dyn_cast<LoadInst>(Val: I)) { |
| 684 | AddNonNullPointer(Ptr: L->getPointerOperand(), PtrSet); |
| 685 | } else if (StoreInst *S = dyn_cast<StoreInst>(Val: I)) { |
| 686 | AddNonNullPointer(Ptr: S->getPointerOperand(), PtrSet); |
| 687 | } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Val: I)) { |
| 688 | if (MI->isVolatile()) return; |
| 689 | |
| 690 | // FIXME: check whether it has a valuerange that excludes zero? |
| 691 | ConstantInt *Len = dyn_cast<ConstantInt>(Val: MI->getLength()); |
| 692 | if (!Len || Len->isZero()) return; |
| 693 | |
| 694 | AddNonNullPointer(Ptr: MI->getRawDest(), PtrSet); |
| 695 | if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(Val: MI)) |
| 696 | AddNonNullPointer(Ptr: MTI->getRawSource(), PtrSet); |
| 697 | } else if (auto *CB = dyn_cast<CallBase>(Val: I)) { |
| 698 | for (auto &U : CB->args()) { |
| 699 | if (U->getType()->isPointerTy() && |
| 700 | CB->paramHasNonNullAttr(ArgNo: CB->getArgOperandNo(U: &U), |
| 701 | /*AllowUndefOrPoison=*/false)) |
| 702 | AddNonNullPointer(Ptr: U.get(), PtrSet, /*IsDereferenced=*/false); |
| 703 | } |
| 704 | } |
| 705 | } |
| 706 | |
| 707 | bool LazyValueInfoImpl::isNonNullAtEndOfBlock(Value *Val, BasicBlock *BB) { |
| 708 | if (NullPointerIsDefined(F: BB->getParent(), |
| 709 | AS: Val->getType()->getPointerAddressSpace())) |
| 710 | return false; |
| 711 | |
| 712 | Val = Val->stripInBoundsOffsets(); |
| 713 | return TheCache.isNonNullAtEndOfBlock(V: Val, BB, InitFn: [](BasicBlock *BB) { |
| 714 | NonNullPointerSet NonNullPointers; |
| 715 | for (Instruction &I : *BB) |
| 716 | AddNonNullPointersByInstruction(I: &I, PtrSet&: NonNullPointers); |
| 717 | return NonNullPointers; |
| 718 | }); |
| 719 | } |
| 720 | |
| 721 | std::optional<ValueLatticeElement> |
| 722 | LazyValueInfoImpl::solveBlockValueNonLocal(Value *Val, BasicBlock *BB) { |
| 723 | ValueLatticeElement Result; // Start Undefined. |
| 724 | |
| 725 | // If this is the entry block, we must be asking about an argument. |
| 726 | if (BB->isEntryBlock()) { |
| 727 | assert(isa<Argument>(Val) && "Unknown live-in to the entry block" ); |
| 728 | if (std::optional<ConstantRange> Range = cast<Argument>(Val)->getRange()) |
| 729 | return ValueLatticeElement::getRange(CR: *Range); |
| 730 | return ValueLatticeElement::getOverdefined(); |
| 731 | } |
| 732 | |
| 733 | // Loop over all of our predecessors, merging what we know from them into |
| 734 | // result. If we encounter an unexplored predecessor, we eagerly explore it |
| 735 | // in a depth first manner. In practice, this has the effect of discovering |
| 736 | // paths we can't analyze eagerly without spending compile times analyzing |
| 737 | // other paths. This heuristic benefits from the fact that predecessors are |
| 738 | // frequently arranged such that dominating ones come first and we quickly |
| 739 | // find a path to function entry. TODO: We should consider explicitly |
| 740 | // canonicalizing to make this true rather than relying on this happy |
| 741 | // accident. |
| 742 | std::optional<BBLatticeElementMap> PredLatticeElements; |
| 743 | if (PerPredRanges) |
| 744 | PredLatticeElements = std::make_optional<BBLatticeElementMap>(); |
| 745 | for (BasicBlock *Pred : predecessors(BB)) { |
| 746 | // Skip self loops. |
| 747 | if (Pred == BB) |
| 748 | continue; |
| 749 | std::optional<ValueLatticeElement> EdgeResult = getEdgeValue(V: Val, F: Pred, T: BB); |
| 750 | if (!EdgeResult) |
| 751 | // Explore that input, then return here |
| 752 | return std::nullopt; |
| 753 | |
| 754 | Result.mergeIn(RHS: *EdgeResult); |
| 755 | |
| 756 | // If we hit overdefined, exit early. The BlockVals entry is already set |
| 757 | // to overdefined. |
| 758 | if (Result.isOverdefined()) { |
| 759 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() |
| 760 | << "' - overdefined because of pred '" |
| 761 | << Pred->getName() << "' (non local).\n" ); |
| 762 | return Result; |
| 763 | } |
| 764 | if (PerPredRanges) |
| 765 | PredLatticeElements->insert(KV: {Pred, *EdgeResult}); |
| 766 | } |
| 767 | |
| 768 | if (PerPredRanges) |
| 769 | TheCache.insertPredecessorResults(Val, BB, PredLatticeElements&: *PredLatticeElements); |
| 770 | |
| 771 | // Return the merged value, which is more precise than 'overdefined'. |
| 772 | assert(!Result.isOverdefined()); |
| 773 | return Result; |
| 774 | } |
| 775 | |
| 776 | std::optional<ValueLatticeElement> |
| 777 | LazyValueInfoImpl::solveBlockValuePHINode(PHINode *PN, BasicBlock *BB) { |
| 778 | ValueLatticeElement Result; // Start Undefined. |
| 779 | |
| 780 | // Loop over all of our predecessors, merging what we know from them into |
| 781 | // result. See the comment about the chosen traversal order in |
| 782 | // solveBlockValueNonLocal; the same reasoning applies here. |
| 783 | std::optional<BBLatticeElementMap> PredLatticeElements; |
| 784 | if (PerPredRanges) |
| 785 | PredLatticeElements = std::make_optional<BBLatticeElementMap>(); |
| 786 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { |
| 787 | BasicBlock *PhiBB = PN->getIncomingBlock(i); |
| 788 | Value *PhiVal = PN->getIncomingValue(i); |
| 789 | // Note that we can provide PN as the context value to getEdgeValue, even |
| 790 | // though the results will be cached, because PN is the value being used as |
| 791 | // the cache key in the caller. |
| 792 | std::optional<ValueLatticeElement> EdgeResult = |
| 793 | getEdgeValue(V: PhiVal, F: PhiBB, T: BB, CxtI: PN); |
| 794 | if (!EdgeResult) |
| 795 | // Explore that input, then return here |
| 796 | return std::nullopt; |
| 797 | |
| 798 | Result.mergeIn(RHS: *EdgeResult); |
| 799 | |
| 800 | // If we hit overdefined, exit early. The BlockVals entry is already set |
| 801 | // to overdefined. |
| 802 | if (Result.isOverdefined()) { |
| 803 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() |
| 804 | << "' - overdefined because of pred (local).\n" ); |
| 805 | |
| 806 | return Result; |
| 807 | } |
| 808 | |
| 809 | if (PerPredRanges) |
| 810 | PredLatticeElements->insert(KV: {PhiBB, *EdgeResult}); |
| 811 | } |
| 812 | |
| 813 | if (PerPredRanges) |
| 814 | TheCache.insertPredecessorResults(Val: PN, BB, PredLatticeElements&: *PredLatticeElements); |
| 815 | |
| 816 | // Return the merged value, which is more precise than 'overdefined'. |
| 817 | assert(!Result.isOverdefined() && "Possible PHI in entry block?" ); |
| 818 | return Result; |
| 819 | } |
| 820 | |
| 821 | // If we can determine a constraint on the value given conditions assumed by |
| 822 | // the program, intersect those constraints with BBLV |
| 823 | void LazyValueInfoImpl::intersectAssumeOrGuardBlockValueConstantRange( |
| 824 | Value *Val, ValueLatticeElement &BBLV, Instruction *BBI) { |
| 825 | BBI = BBI ? BBI : dyn_cast<Instruction>(Val); |
| 826 | if (!BBI) |
| 827 | return; |
| 828 | |
| 829 | BasicBlock *BB = BBI->getParent(); |
| 830 | for (auto &AssumeVH : AC->assumptionsFor(V: Val)) { |
| 831 | if (!AssumeVH) |
| 832 | continue; |
| 833 | |
| 834 | // Only check assumes in the block of the context instruction. Other |
| 835 | // assumes will have already been taken into account when the value was |
| 836 | // propagated from predecessor blocks. |
| 837 | auto *I = cast<AssumeInst>(Val&: AssumeVH); |
| 838 | |
| 839 | if (I->getParent() != BB || !isValidAssumeForContext(I, CxtI: BBI)) |
| 840 | continue; |
| 841 | |
| 842 | if (AssumeVH.Index != AssumptionCache::ExprResultIdx) { |
| 843 | if (RetainedKnowledge RK = getKnowledgeFromBundle( |
| 844 | Assume&: *I, BOI: I->bundle_op_info_begin()[AssumeVH.Index])) { |
| 845 | if (RK.WasOn != Val) |
| 846 | continue; |
| 847 | switch (RK.AttrKind) { |
| 848 | case Attribute::NonNull: |
| 849 | BBLV = BBLV.intersect(Other: ValueLatticeElement::getNot( |
| 850 | C: Constant::getNullValue(Ty: RK.WasOn->getType()))); |
| 851 | break; |
| 852 | |
| 853 | case Attribute::Dereferenceable: |
| 854 | if (auto *CI = dyn_cast<ConstantInt>(Val: RK.IRArgValue); |
| 855 | CI && !CI->isZero()) |
| 856 | BBLV = BBLV.intersect(Other: ValueLatticeElement::getNot( |
| 857 | C: Constant::getNullValue(Ty: RK.WasOn->getType()))); |
| 858 | break; |
| 859 | |
| 860 | default: |
| 861 | break; |
| 862 | } |
| 863 | } |
| 864 | } else { |
| 865 | BBLV = BBLV.intersect(Other: *getValueFromCondition(Val, Cond: I->getArgOperand(i: 0), |
| 866 | /*IsTrueDest*/ true, |
| 867 | /*UseBlockValue*/ false)); |
| 868 | } |
| 869 | } |
| 870 | |
| 871 | // If guards are not used in the module, don't spend time looking for them |
| 872 | if (GuardDecl && !GuardDecl->use_empty() && |
| 873 | BBI->getIterator() != BB->begin()) { |
| 874 | for (Instruction &I : |
| 875 | make_range(x: std::next(x: BBI->getIterator().getReverse()), y: BB->rend())) { |
| 876 | Value *Cond = nullptr; |
| 877 | if (match(V: &I, P: m_Intrinsic<Intrinsic::experimental_guard>(Op0: m_Value(V&: Cond)))) |
| 878 | BBLV = BBLV.intersect(Other: *getValueFromCondition(Val, Cond, |
| 879 | /*IsTrueDest*/ true, |
| 880 | /*UseBlockValue*/ false)); |
| 881 | } |
| 882 | } |
| 883 | |
| 884 | if (BBLV.isOverdefined()) { |
| 885 | // Check whether we're checking at the terminator, and the pointer has |
| 886 | // been dereferenced in this block. |
| 887 | PointerType *PTy = dyn_cast<PointerType>(Val: Val->getType()); |
| 888 | if (PTy && BB->getTerminator() == BBI && |
| 889 | isNonNullAtEndOfBlock(Val, BB)) |
| 890 | BBLV = ValueLatticeElement::getNot(C: ConstantPointerNull::get(T: PTy)); |
| 891 | } |
| 892 | } |
| 893 | |
| 894 | std::optional<ValueLatticeElement> |
| 895 | LazyValueInfoImpl::solveBlockValueSelect(SelectInst *SI, BasicBlock *BB) { |
| 896 | // Recurse on our inputs if needed |
| 897 | std::optional<ValueLatticeElement> OptTrueVal = |
| 898 | getBlockValue(Val: SI->getTrueValue(), BB, CxtI: SI); |
| 899 | if (!OptTrueVal) |
| 900 | return std::nullopt; |
| 901 | ValueLatticeElement &TrueVal = *OptTrueVal; |
| 902 | |
| 903 | std::optional<ValueLatticeElement> OptFalseVal = |
| 904 | getBlockValue(Val: SI->getFalseValue(), BB, CxtI: SI); |
| 905 | if (!OptFalseVal) |
| 906 | return std::nullopt; |
| 907 | ValueLatticeElement &FalseVal = *OptFalseVal; |
| 908 | |
| 909 | if (TrueVal.isConstantRange() || FalseVal.isConstantRange()) { |
| 910 | const ConstantRange &TrueCR = TrueVal.asConstantRange(Ty: SI->getType()); |
| 911 | const ConstantRange &FalseCR = FalseVal.asConstantRange(Ty: SI->getType()); |
| 912 | Value *LHS = nullptr; |
| 913 | Value *RHS = nullptr; |
| 914 | SelectPatternResult SPR = matchSelectPattern(V: SI, LHS, RHS); |
| 915 | // Is this a min specifically of our two inputs? (Avoid the risk of |
| 916 | // ValueTracking getting smarter looking back past our immediate inputs.) |
| 917 | if (SelectPatternResult::isMinOrMax(SPF: SPR.Flavor) && |
| 918 | ((LHS == SI->getTrueValue() && RHS == SI->getFalseValue()) || |
| 919 | (RHS == SI->getTrueValue() && LHS == SI->getFalseValue()))) { |
| 920 | ConstantRange ResultCR = [&]() { |
| 921 | switch (SPR.Flavor) { |
| 922 | default: |
| 923 | llvm_unreachable("unexpected minmax type!" ); |
| 924 | case SPF_SMIN: /// Signed minimum |
| 925 | return TrueCR.smin(Other: FalseCR); |
| 926 | case SPF_UMIN: /// Unsigned minimum |
| 927 | return TrueCR.umin(Other: FalseCR); |
| 928 | case SPF_SMAX: /// Signed maximum |
| 929 | return TrueCR.smax(Other: FalseCR); |
| 930 | case SPF_UMAX: /// Unsigned maximum |
| 931 | return TrueCR.umax(Other: FalseCR); |
| 932 | }; |
| 933 | }(); |
| 934 | return ValueLatticeElement::getRange( |
| 935 | CR: ResultCR, MayIncludeUndef: TrueVal.isConstantRangeIncludingUndef() || |
| 936 | FalseVal.isConstantRangeIncludingUndef()); |
| 937 | } |
| 938 | |
| 939 | if (SPR.Flavor == SPF_ABS) { |
| 940 | if (LHS == SI->getTrueValue()) |
| 941 | return ValueLatticeElement::getRange( |
| 942 | CR: TrueCR.abs(), MayIncludeUndef: TrueVal.isConstantRangeIncludingUndef()); |
| 943 | if (LHS == SI->getFalseValue()) |
| 944 | return ValueLatticeElement::getRange( |
| 945 | CR: FalseCR.abs(), MayIncludeUndef: FalseVal.isConstantRangeIncludingUndef()); |
| 946 | } |
| 947 | |
| 948 | if (SPR.Flavor == SPF_NABS) { |
| 949 | ConstantRange Zero(APInt::getZero(numBits: TrueCR.getBitWidth())); |
| 950 | if (LHS == SI->getTrueValue()) |
| 951 | return ValueLatticeElement::getRange( |
| 952 | CR: Zero.sub(Other: TrueCR.abs()), MayIncludeUndef: FalseVal.isConstantRangeIncludingUndef()); |
| 953 | if (LHS == SI->getFalseValue()) |
| 954 | return ValueLatticeElement::getRange( |
| 955 | CR: Zero.sub(Other: FalseCR.abs()), MayIncludeUndef: FalseVal.isConstantRangeIncludingUndef()); |
| 956 | } |
| 957 | } |
| 958 | |
| 959 | // Can we constrain the facts about the true and false values by using the |
| 960 | // condition itself? This shows up with idioms like e.g. select(a > 5, a, 5). |
| 961 | // TODO: We could potentially refine an overdefined true value above. |
| 962 | Value *Cond = SI->getCondition(); |
| 963 | // If the value is undef, a different value may be chosen in |
| 964 | // the select condition. |
| 965 | if (isGuaranteedNotToBeUndef(V: Cond, AC)) { |
| 966 | TrueVal = |
| 967 | TrueVal.intersect(Other: *getValueFromCondition(Val: SI->getTrueValue(), Cond, |
| 968 | /*IsTrueDest*/ true, |
| 969 | /*UseBlockValue*/ false)); |
| 970 | FalseVal = |
| 971 | FalseVal.intersect(Other: *getValueFromCondition(Val: SI->getFalseValue(), Cond, |
| 972 | /*IsTrueDest*/ false, |
| 973 | /*UseBlockValue*/ false)); |
| 974 | } |
| 975 | |
| 976 | TrueVal.mergeIn(RHS: FalseVal); |
| 977 | return TrueVal; |
| 978 | } |
| 979 | |
| 980 | std::optional<ConstantRange> |
| 981 | LazyValueInfoImpl::getRangeFor(Value *V, Instruction *CxtI, BasicBlock *BB) { |
| 982 | std::optional<ValueLatticeElement> OptVal = getBlockValue(Val: V, BB, CxtI); |
| 983 | if (!OptVal) |
| 984 | return std::nullopt; |
| 985 | return OptVal->asConstantRange(Ty: V->getType()); |
| 986 | } |
| 987 | |
| 988 | std::optional<ValueLatticeElement> |
| 989 | LazyValueInfoImpl::solveBlockValueCast(CastInst *CI, BasicBlock *BB) { |
| 990 | // Filter out casts we don't know how to reason about before attempting to |
| 991 | // recurse on our operand. This can cut a long search short if we know we're |
| 992 | // not going to be able to get any useful information anways. |
| 993 | switch (CI->getOpcode()) { |
| 994 | case Instruction::Trunc: |
| 995 | case Instruction::SExt: |
| 996 | case Instruction::ZExt: |
| 997 | break; |
| 998 | default: |
| 999 | // Unhandled instructions are overdefined. |
| 1000 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() |
| 1001 | << "' - overdefined (unknown cast).\n" ); |
| 1002 | return ValueLatticeElement::getOverdefined(); |
| 1003 | } |
| 1004 | |
| 1005 | // Figure out the range of the LHS. If that fails, we still apply the |
| 1006 | // transfer rule on the full set since we may be able to locally infer |
| 1007 | // interesting facts. |
| 1008 | std::optional<ConstantRange> LHSRes = getRangeFor(V: CI->getOperand(i_nocapture: 0), CxtI: CI, BB); |
| 1009 | if (!LHSRes) |
| 1010 | // More work to do before applying this transfer rule. |
| 1011 | return std::nullopt; |
| 1012 | const ConstantRange &LHSRange = *LHSRes; |
| 1013 | |
| 1014 | const unsigned ResultBitWidth = CI->getType()->getScalarSizeInBits(); |
| 1015 | |
| 1016 | // NOTE: We're currently limited by the set of operations that ConstantRange |
| 1017 | // can evaluate symbolically. Enhancing that set will allows us to analyze |
| 1018 | // more definitions. |
| 1019 | ConstantRange Res = ConstantRange::getEmpty(BitWidth: ResultBitWidth); |
| 1020 | if (auto *Trunc = dyn_cast<TruncInst>(Val: CI)) |
| 1021 | Res = LHSRange.truncate(BitWidth: ResultBitWidth, NoWrapKind: Trunc->getNoWrapKind()); |
| 1022 | else |
| 1023 | Res = LHSRange.castOp(CastOp: CI->getOpcode(), BitWidth: ResultBitWidth); |
| 1024 | |
| 1025 | return ValueLatticeElement::getRange(CR: Res); |
| 1026 | } |
| 1027 | |
| 1028 | std::optional<ValueLatticeElement> |
| 1029 | LazyValueInfoImpl::solveBlockValueBinaryOpImpl( |
| 1030 | Instruction *I, BasicBlock *BB, |
| 1031 | std::function<ConstantRange(const ConstantRange &, const ConstantRange &)> |
| 1032 | OpFn) { |
| 1033 | Value *LHS = I->getOperand(i: 0); |
| 1034 | Value *RHS = I->getOperand(i: 1); |
| 1035 | |
| 1036 | auto ThreadBinOpOverSelect = |
| 1037 | [&](Value *X, const ConstantRange &CRX, SelectInst *Y, |
| 1038 | bool XIsLHS) -> std::optional<ValueLatticeElement> { |
| 1039 | Value *Cond = Y->getCondition(); |
| 1040 | // Only handle selects with constant values. |
| 1041 | Constant *TrueC = dyn_cast<Constant>(Val: Y->getTrueValue()); |
| 1042 | if (!TrueC) |
| 1043 | return std::nullopt; |
| 1044 | Constant *FalseC = dyn_cast<Constant>(Val: Y->getFalseValue()); |
| 1045 | if (!FalseC) |
| 1046 | return std::nullopt; |
| 1047 | if (!isGuaranteedNotToBeUndef(V: Cond, AC)) |
| 1048 | return std::nullopt; |
| 1049 | |
| 1050 | ConstantRange TrueX = |
| 1051 | CRX.intersectWith(CR: getValueFromCondition(Val: X, Cond, /*CondIsTrue=*/IsTrueDest: true, |
| 1052 | /*UseBlockValue=*/false) |
| 1053 | ->asConstantRange(Ty: X->getType())); |
| 1054 | ConstantRange FalseX = |
| 1055 | CRX.intersectWith(CR: getValueFromCondition(Val: X, Cond, /*CondIsTrue=*/IsTrueDest: false, |
| 1056 | /*UseBlockValue=*/false) |
| 1057 | ->asConstantRange(Ty: X->getType())); |
| 1058 | ConstantRange TrueY = TrueC->toConstantRange(); |
| 1059 | ConstantRange FalseY = FalseC->toConstantRange(); |
| 1060 | |
| 1061 | if (XIsLHS) |
| 1062 | return ValueLatticeElement::getRange( |
| 1063 | CR: OpFn(TrueX, TrueY).unionWith(CR: OpFn(FalseX, FalseY))); |
| 1064 | return ValueLatticeElement::getRange( |
| 1065 | CR: OpFn(TrueY, TrueX).unionWith(CR: OpFn(FalseY, FalseX))); |
| 1066 | }; |
| 1067 | |
| 1068 | // Figure out the ranges of the operands. If that fails, use a |
| 1069 | // conservative range, but apply the transfer rule anyways. This |
| 1070 | // lets us pick up facts from expressions like "and i32 (call i32 |
| 1071 | // @foo()), 32" |
| 1072 | std::optional<ConstantRange> LHSRes = getRangeFor(V: LHS, CxtI: I, BB); |
| 1073 | if (!LHSRes) |
| 1074 | return std::nullopt; |
| 1075 | |
| 1076 | // Try to thread binop over rhs select |
| 1077 | if (auto *SI = dyn_cast<SelectInst>(Val: RHS)) { |
| 1078 | if (auto Res = ThreadBinOpOverSelect(LHS, *LHSRes, SI, /*XIsLHS=*/true)) |
| 1079 | return *Res; |
| 1080 | } |
| 1081 | |
| 1082 | std::optional<ConstantRange> RHSRes = getRangeFor(V: RHS, CxtI: I, BB); |
| 1083 | if (!RHSRes) |
| 1084 | return std::nullopt; |
| 1085 | |
| 1086 | // Try to thread binop over lhs select |
| 1087 | if (auto *SI = dyn_cast<SelectInst>(Val: LHS)) { |
| 1088 | if (auto Res = ThreadBinOpOverSelect(RHS, *RHSRes, SI, /*XIsLHS=*/false)) |
| 1089 | return *Res; |
| 1090 | } |
| 1091 | |
| 1092 | const ConstantRange &LHSRange = *LHSRes; |
| 1093 | const ConstantRange &RHSRange = *RHSRes; |
| 1094 | |
| 1095 | std::optional<ValueLatticeElement> MergedResult = |
| 1096 | ValueLatticeElement::getRange(CR: OpFn(LHSRange, RHSRange)); |
| 1097 | |
| 1098 | if (!PerPredRanges) |
| 1099 | return MergedResult; |
| 1100 | |
| 1101 | std::optional<BBLatticeElementMap> PredLHS = |
| 1102 | TheCache.getCachedPredecessorInfo(V: LHS, BB); |
| 1103 | if (!PredLHS) |
| 1104 | return MergedResult; |
| 1105 | std::optional<BBLatticeElementMap> PredRHS = |
| 1106 | TheCache.getCachedPredecessorInfo(V: RHS, BB); |
| 1107 | if (!PredRHS) |
| 1108 | return MergedResult; |
| 1109 | |
| 1110 | const BBLatticeElementMap &LHSPredMap = *PredLHS; |
| 1111 | const BBLatticeElementMap &RHSPredMap = *PredRHS; |
| 1112 | |
| 1113 | BBLatticeElementMap PredLatticeElements; |
| 1114 | ValueLatticeElement OverallPredResult; |
| 1115 | for (auto *Pred : predecessors(BB)) { |
| 1116 | auto LHSIt = LHSPredMap.find_as(Val: Pred); |
| 1117 | if (LHSIt == LHSPredMap.end()) |
| 1118 | return MergedResult; |
| 1119 | const ValueLatticeElement &LHSFromPred = LHSIt->second; |
| 1120 | std::optional<ConstantRange> LHSFromPredRes = |
| 1121 | LHSFromPred.asConstantRange(Ty: LHS->getType()); |
| 1122 | if (!LHSFromPredRes) |
| 1123 | return MergedResult; |
| 1124 | |
| 1125 | auto RHSIt = RHSPredMap.find_as(Val: Pred); |
| 1126 | if (RHSIt == RHSPredMap.end()) |
| 1127 | return MergedResult; |
| 1128 | const ValueLatticeElement &RHSFromPred = RHSIt->second; |
| 1129 | std::optional<ConstantRange> RHSFromPredRes = |
| 1130 | RHSFromPred.asConstantRange(Ty: RHS->getType()); |
| 1131 | if (!RHSFromPredRes) |
| 1132 | return MergedResult; |
| 1133 | |
| 1134 | const ConstantRange &LHSFromPredRange = *LHSFromPredRes; |
| 1135 | const ConstantRange &RHSFromPredRange = *RHSFromPredRes; |
| 1136 | std::optional<ValueLatticeElement> PredResult = |
| 1137 | ValueLatticeElement::getRange(CR: OpFn(LHSFromPredRange, RHSFromPredRange)); |
| 1138 | if (!PredResult) |
| 1139 | return MergedResult; |
| 1140 | if (PredResult->isOverdefined()) { |
| 1141 | LLVM_DEBUG( |
| 1142 | dbgs() << " pred BB '" << Pred->getName() << "' for BB '" |
| 1143 | << BB->getName() |
| 1144 | << "' overdefined. Discarding all predecessor intervals.\n" ); |
| 1145 | return MergedResult; |
| 1146 | } |
| 1147 | PredLatticeElements.insert(KV: {Pred, *PredResult}); |
| 1148 | OverallPredResult.mergeIn(RHS: *PredResult); |
| 1149 | } |
| 1150 | |
| 1151 | // If this point is reached, all predecessors for both LHS and RHS have |
| 1152 | // constant ranges previously computed. Can cache result and use the |
| 1153 | // OverallPredResult; |
| 1154 | TheCache.insertPredecessorResults(Val: I, BB, PredLatticeElements); |
| 1155 | |
| 1156 | LLVM_DEBUG(dbgs() << " Using predecessor intervals, evaluated " << *I |
| 1157 | << " to: " << OverallPredResult << ".\n" ); |
| 1158 | |
| 1159 | if (!MergedResult) |
| 1160 | return OverallPredResult; |
| 1161 | |
| 1162 | LLVM_DEBUG(dbgs() << " Intersecting intervals for " << *I << ": " |
| 1163 | << OverallPredResult << " and " << MergedResult << ".\n" ); |
| 1164 | return MergedResult->intersect(Other: OverallPredResult); |
| 1165 | } |
| 1166 | |
| 1167 | std::optional<ValueLatticeElement> |
| 1168 | LazyValueInfoImpl::solveBlockValueBinaryOp(BinaryOperator *BO, BasicBlock *BB) { |
| 1169 | assert(BO->getOperand(0)->getType()->isSized() && |
| 1170 | "all operands to binary operators are sized" ); |
| 1171 | if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(Val: BO)) { |
| 1172 | unsigned NoWrapKind = OBO->getNoWrapKind(); |
| 1173 | return solveBlockValueBinaryOpImpl( |
| 1174 | I: BO, BB, |
| 1175 | OpFn: [BO, NoWrapKind](const ConstantRange &CR1, const ConstantRange &CR2) { |
| 1176 | return CR1.overflowingBinaryOp(BinOp: BO->getOpcode(), Other: CR2, NoWrapKind); |
| 1177 | }); |
| 1178 | } |
| 1179 | |
| 1180 | return solveBlockValueBinaryOpImpl( |
| 1181 | I: BO, BB, OpFn: [BO](const ConstantRange &CR1, const ConstantRange &CR2) { |
| 1182 | return CR1.binaryOp(BinOp: BO->getOpcode(), Other: CR2); |
| 1183 | }); |
| 1184 | } |
| 1185 | |
| 1186 | std::optional<ValueLatticeElement> |
| 1187 | LazyValueInfoImpl::solveBlockValueOverflowIntrinsic(WithOverflowInst *WO, |
| 1188 | BasicBlock *BB) { |
| 1189 | return solveBlockValueBinaryOpImpl( |
| 1190 | I: WO, BB, OpFn: [WO](const ConstantRange &CR1, const ConstantRange &CR2) { |
| 1191 | return CR1.binaryOp(BinOp: WO->getBinaryOp(), Other: CR2); |
| 1192 | }); |
| 1193 | } |
| 1194 | |
| 1195 | std::optional<ValueLatticeElement> |
| 1196 | LazyValueInfoImpl::solveBlockValueIntrinsic(IntrinsicInst *II, BasicBlock *BB) { |
| 1197 | ValueLatticeElement MetadataVal = getFromRangeMetadata(BBI: II); |
| 1198 | if (!ConstantRange::isIntrinsicSupported(IntrinsicID: II->getIntrinsicID())) { |
| 1199 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() |
| 1200 | << "' - unknown intrinsic.\n" ); |
| 1201 | return MetadataVal; |
| 1202 | } |
| 1203 | |
| 1204 | SmallVector<ConstantRange, 2> OpRanges; |
| 1205 | for (Value *Op : II->args()) { |
| 1206 | std::optional<ConstantRange> Range = getRangeFor(V: Op, CxtI: II, BB); |
| 1207 | if (!Range) |
| 1208 | return std::nullopt; |
| 1209 | OpRanges.push_back(Elt: *Range); |
| 1210 | } |
| 1211 | |
| 1212 | return ValueLatticeElement::getRange( |
| 1213 | CR: ConstantRange::intrinsic(IntrinsicID: II->getIntrinsicID(), Ops: OpRanges)) |
| 1214 | .intersect(Other: MetadataVal); |
| 1215 | } |
| 1216 | |
| 1217 | std::optional<ValueLatticeElement> |
| 1218 | LazyValueInfoImpl::solveBlockValueInsertElement(InsertElementInst *IEI, |
| 1219 | BasicBlock *BB) { |
| 1220 | std::optional<ValueLatticeElement> OptEltVal = |
| 1221 | getBlockValue(Val: IEI->getOperand(i_nocapture: 1), BB, CxtI: IEI); |
| 1222 | if (!OptEltVal) |
| 1223 | return std::nullopt; |
| 1224 | ValueLatticeElement &Res = *OptEltVal; |
| 1225 | |
| 1226 | std::optional<ValueLatticeElement> OptVecVal = |
| 1227 | getBlockValue(Val: IEI->getOperand(i_nocapture: 0), BB, CxtI: IEI); |
| 1228 | if (!OptVecVal) |
| 1229 | return std::nullopt; |
| 1230 | |
| 1231 | // Bail out if the inserted element is a constant expression. Unlike other |
| 1232 | // ValueLattice types, these are not considered an implicit splat when a |
| 1233 | // vector type is used. |
| 1234 | // We could call ConstantFoldInsertElementInstruction here to handle these. |
| 1235 | if (OptEltVal->isConstant()) |
| 1236 | return ValueLatticeElement::getOverdefined(); |
| 1237 | |
| 1238 | Res.mergeIn(RHS: *OptVecVal); |
| 1239 | return Res; |
| 1240 | } |
| 1241 | |
| 1242 | std::optional<ValueLatticeElement> |
| 1243 | LazyValueInfoImpl::(ExtractValueInst *EVI, |
| 1244 | BasicBlock *BB) { |
| 1245 | if (auto *WO = dyn_cast<WithOverflowInst>(Val: EVI->getAggregateOperand())) |
| 1246 | if (EVI->getNumIndices() == 1 && *EVI->idx_begin() == 0) |
| 1247 | return solveBlockValueOverflowIntrinsic(WO, BB); |
| 1248 | |
| 1249 | // Handle extractvalue of insertvalue to allow further simplification |
| 1250 | // based on replaced with.overflow intrinsics. |
| 1251 | if (Value *V = simplifyExtractValueInst( |
| 1252 | Agg: EVI->getAggregateOperand(), Idxs: EVI->getIndices(), |
| 1253 | Q: EVI->getDataLayout())) |
| 1254 | return getBlockValue(Val: V, BB, CxtI: EVI); |
| 1255 | |
| 1256 | LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() |
| 1257 | << "' - overdefined (unknown extractvalue).\n" ); |
| 1258 | return ValueLatticeElement::getOverdefined(); |
| 1259 | } |
| 1260 | |
| 1261 | static bool matchICmpOperand(APInt &Offset, Value *LHS, Value *Val, |
| 1262 | ICmpInst::Predicate Pred) { |
| 1263 | if (LHS == Val) |
| 1264 | return true; |
| 1265 | |
| 1266 | // Handle range checking idiom produced by InstCombine. We will subtract the |
| 1267 | // offset from the allowed range for RHS in this case. |
| 1268 | const APInt *C; |
| 1269 | if (match(V: LHS, P: m_AddLike(L: m_Specific(V: Val), R: m_APInt(Res&: C)))) { |
| 1270 | Offset = *C; |
| 1271 | return true; |
| 1272 | } |
| 1273 | |
| 1274 | // Handle the symmetric case. This appears in saturation patterns like |
| 1275 | // (x == 16) ? 16 : (x + 1). |
| 1276 | if (match(V: Val, P: m_AddLike(L: m_Specific(V: LHS), R: m_APInt(Res&: C)))) { |
| 1277 | Offset = -*C; |
| 1278 | return true; |
| 1279 | } |
| 1280 | |
| 1281 | // If (x | y) < C, then (x < C) && (y < C). |
| 1282 | if (match(V: LHS, P: m_c_Or(L: m_Specific(V: Val), R: m_Value())) && |
| 1283 | (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE)) |
| 1284 | return true; |
| 1285 | |
| 1286 | // If (x & y) > C, then (x > C) && (y > C). |
| 1287 | if (match(V: LHS, P: m_c_And(L: m_Specific(V: Val), R: m_Value())) && |
| 1288 | (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE)) |
| 1289 | return true; |
| 1290 | |
| 1291 | return false; |
| 1292 | } |
| 1293 | |
| 1294 | /// Get value range for a "(Val + Offset) Pred RHS" condition. |
| 1295 | std::optional<ValueLatticeElement> |
| 1296 | LazyValueInfoImpl::getValueFromSimpleICmpCondition(CmpInst::Predicate Pred, |
| 1297 | Value *RHS, |
| 1298 | const APInt &Offset, |
| 1299 | Instruction *CxtI, |
| 1300 | bool UseBlockValue) { |
| 1301 | ConstantRange RHSRange(RHS->getType()->getScalarSizeInBits(), |
| 1302 | /*isFullSet=*/true); |
| 1303 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Val: RHS)) { |
| 1304 | RHSRange = ConstantRange(CI->getValue()); |
| 1305 | } else if (UseBlockValue) { |
| 1306 | std::optional<ValueLatticeElement> R = |
| 1307 | getBlockValue(Val: RHS, BB: CxtI->getParent(), CxtI); |
| 1308 | if (!R) |
| 1309 | return std::nullopt; |
| 1310 | RHSRange = R->asConstantRange(Ty: RHS->getType()); |
| 1311 | } |
| 1312 | |
| 1313 | ConstantRange TrueValues = |
| 1314 | ConstantRange::makeAllowedICmpRegion(Pred, Other: RHSRange); |
| 1315 | return ValueLatticeElement::getRange(CR: TrueValues.subtract(CI: Offset)); |
| 1316 | } |
| 1317 | |
| 1318 | static std::optional<ConstantRange> |
| 1319 | getRangeViaSLT(CmpInst::Predicate Pred, APInt RHS, |
| 1320 | function_ref<std::optional<ConstantRange>(const APInt &)> Fn) { |
| 1321 | bool Invert = false; |
| 1322 | if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) { |
| 1323 | Pred = ICmpInst::getInversePredicate(pred: Pred); |
| 1324 | Invert = true; |
| 1325 | } |
| 1326 | if (Pred == ICmpInst::ICMP_SLE) { |
| 1327 | Pred = ICmpInst::ICMP_SLT; |
| 1328 | if (RHS.isMaxSignedValue()) |
| 1329 | return std::nullopt; // Could also return full/empty here, if we wanted. |
| 1330 | ++RHS; |
| 1331 | } |
| 1332 | assert(Pred == ICmpInst::ICMP_SLT && "Must be signed predicate" ); |
| 1333 | if (auto CR = Fn(RHS)) |
| 1334 | return Invert ? CR->inverse() : CR; |
| 1335 | return std::nullopt; |
| 1336 | } |
| 1337 | |
| 1338 | /// Get value range for a "ctpop(Val) Pred RHS" condition. |
| 1339 | static ValueLatticeElement getValueFromICmpCtpop(ICmpInst::Predicate Pred, |
| 1340 | Value *RHS) { |
| 1341 | unsigned BitWidth = RHS->getType()->getScalarSizeInBits(); |
| 1342 | |
| 1343 | auto *RHSConst = dyn_cast<ConstantInt>(Val: RHS); |
| 1344 | if (!RHSConst) |
| 1345 | return ValueLatticeElement::getOverdefined(); |
| 1346 | |
| 1347 | ConstantRange ResValRange = |
| 1348 | ConstantRange::makeExactICmpRegion(Pred, Other: RHSConst->getValue()); |
| 1349 | |
| 1350 | unsigned ResMin = ResValRange.getUnsignedMin().getLimitedValue(Limit: BitWidth); |
| 1351 | unsigned ResMax = ResValRange.getUnsignedMax().getLimitedValue(Limit: BitWidth); |
| 1352 | |
| 1353 | APInt ValMin = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: ResMin); |
| 1354 | APInt ValMax = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: ResMax); |
| 1355 | return ValueLatticeElement::getRange( |
| 1356 | CR: ConstantRange::getNonEmpty(Lower: std::move(ValMin), Upper: ValMax + 1)); |
| 1357 | } |
| 1358 | |
| 1359 | std::optional<ValueLatticeElement> LazyValueInfoImpl::getValueFromICmpCondition( |
| 1360 | Value *Val, ICmpInst *ICI, bool isTrueDest, bool UseBlockValue) { |
| 1361 | Value *LHS = ICI->getOperand(i_nocapture: 0); |
| 1362 | Value *RHS = ICI->getOperand(i_nocapture: 1); |
| 1363 | |
| 1364 | // Get the predicate that must hold along the considered edge. |
| 1365 | CmpInst::Predicate EdgePred = |
| 1366 | isTrueDest ? ICI->getPredicate() : ICI->getInversePredicate(); |
| 1367 | |
| 1368 | if (isa<Constant>(Val: RHS)) { |
| 1369 | if (ICI->isEquality() && LHS == Val) { |
| 1370 | if (EdgePred == ICmpInst::ICMP_EQ) |
| 1371 | return ValueLatticeElement::get(C: cast<Constant>(Val: RHS)); |
| 1372 | else if (!isa<UndefValue>(Val: RHS)) |
| 1373 | return ValueLatticeElement::getNot(C: cast<Constant>(Val: RHS)); |
| 1374 | } |
| 1375 | } |
| 1376 | |
| 1377 | Type *Ty = Val->getType(); |
| 1378 | if (!Ty->isIntegerTy()) |
| 1379 | return ValueLatticeElement::getOverdefined(); |
| 1380 | |
| 1381 | unsigned BitWidth = Ty->getScalarSizeInBits(); |
| 1382 | APInt Offset(BitWidth, 0); |
| 1383 | if (matchICmpOperand(Offset, LHS, Val, Pred: EdgePred)) |
| 1384 | return getValueFromSimpleICmpCondition(Pred: EdgePred, RHS, Offset, CxtI: ICI, |
| 1385 | UseBlockValue); |
| 1386 | |
| 1387 | CmpInst::Predicate SwappedPred = CmpInst::getSwappedPredicate(pred: EdgePred); |
| 1388 | if (matchICmpOperand(Offset, LHS: RHS, Val, Pred: SwappedPred)) |
| 1389 | return getValueFromSimpleICmpCondition(Pred: SwappedPred, RHS: LHS, Offset, CxtI: ICI, |
| 1390 | UseBlockValue); |
| 1391 | |
| 1392 | if (match(V: LHS, P: m_Intrinsic<Intrinsic::ctpop>(Op0: m_Specific(V: Val)))) |
| 1393 | return getValueFromICmpCtpop(Pred: EdgePred, RHS); |
| 1394 | |
| 1395 | const APInt *Mask, *C; |
| 1396 | if (match(V: LHS, P: m_And(L: m_Specific(V: Val), R: m_APInt(Res&: Mask))) && |
| 1397 | match(V: RHS, P: m_APInt(Res&: C))) { |
| 1398 | // If (Val & Mask) == C then all the masked bits are known and we can |
| 1399 | // compute a value range based on that. |
| 1400 | if (EdgePred == ICmpInst::ICMP_EQ) { |
| 1401 | KnownBits Known; |
| 1402 | Known.Zero = ~*C & *Mask; |
| 1403 | Known.One = *C & *Mask; |
| 1404 | return ValueLatticeElement::getRange( |
| 1405 | CR: ConstantRange::fromKnownBits(Known, /*IsSigned*/ false)); |
| 1406 | } |
| 1407 | |
| 1408 | if (EdgePred == ICmpInst::ICMP_NE) |
| 1409 | return ValueLatticeElement::getRange( |
| 1410 | CR: ConstantRange::makeMaskNotEqualRange(Mask: *Mask, C: *C)); |
| 1411 | } |
| 1412 | |
| 1413 | // If (X urem Modulus) >= C, then X >= C. |
| 1414 | // If trunc X >= C, then X >= C. |
| 1415 | // TODO: An upper bound could be computed as well. |
| 1416 | if (match(V: LHS, P: m_CombineOr(L: m_URem(L: m_Specific(V: Val), R: m_Value()), |
| 1417 | R: m_Trunc(Op: m_Specific(V: Val)))) && |
| 1418 | match(V: RHS, P: m_APInt(Res&: C))) { |
| 1419 | // Use the icmp region so we don't have to deal with different predicates. |
| 1420 | ConstantRange CR = ConstantRange::makeExactICmpRegion(Pred: EdgePred, Other: *C); |
| 1421 | if (!CR.isEmptySet()) |
| 1422 | return ValueLatticeElement::getRange(CR: ConstantRange::getNonEmpty( |
| 1423 | Lower: CR.getUnsignedMin().zext(width: BitWidth), Upper: APInt(BitWidth, 0))); |
| 1424 | } |
| 1425 | |
| 1426 | // Recognize: |
| 1427 | // icmp slt (ashr X, ShAmtC), C --> icmp slt X, C << ShAmtC |
| 1428 | // Preconditions: (C << ShAmtC) >> ShAmtC == C |
| 1429 | const APInt *ShAmtC; |
| 1430 | if (CmpInst::isSigned(predicate: EdgePred) && |
| 1431 | match(V: LHS, P: m_AShr(L: m_Specific(V: Val), R: m_APInt(Res&: ShAmtC))) && |
| 1432 | match(V: RHS, P: m_APInt(Res&: C))) { |
| 1433 | auto CR = getRangeViaSLT( |
| 1434 | Pred: EdgePred, RHS: *C, Fn: [&](const APInt &RHS) -> std::optional<ConstantRange> { |
| 1435 | APInt New = RHS << *ShAmtC; |
| 1436 | if ((New.ashr(ShiftAmt: *ShAmtC)) != RHS) |
| 1437 | return std::nullopt; |
| 1438 | return ConstantRange::getNonEmpty( |
| 1439 | Lower: APInt::getSignedMinValue(numBits: New.getBitWidth()), Upper: New); |
| 1440 | }); |
| 1441 | if (CR) |
| 1442 | return ValueLatticeElement::getRange(CR: *CR); |
| 1443 | } |
| 1444 | |
| 1445 | // a - b or ptrtoint(a) - ptrtoint(b) ==/!= 0 if a ==/!= b |
| 1446 | Value *X, *Y; |
| 1447 | if (ICI->isEquality() && match(V: Val, P: m_Sub(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
| 1448 | // Peek through ptrtoints |
| 1449 | match(V: X, P: m_PtrToIntSameSize(DL, Op: m_Value(V&: X))); |
| 1450 | match(V: Y, P: m_PtrToIntSameSize(DL, Op: m_Value(V&: Y))); |
| 1451 | if ((X == LHS && Y == RHS) || (X == RHS && Y == LHS)) { |
| 1452 | Constant *NullVal = Constant::getNullValue(Ty: Val->getType()); |
| 1453 | if (EdgePred == ICmpInst::ICMP_EQ) |
| 1454 | return ValueLatticeElement::get(C: NullVal); |
| 1455 | return ValueLatticeElement::getNot(C: NullVal); |
| 1456 | } |
| 1457 | } |
| 1458 | |
| 1459 | return ValueLatticeElement::getOverdefined(); |
| 1460 | } |
| 1461 | |
| 1462 | ValueLatticeElement LazyValueInfoImpl::getValueFromTrunc(Value *Val, |
| 1463 | TruncInst *Trunc, |
| 1464 | bool IsTrueDest) { |
| 1465 | assert(Trunc->getType()->isIntOrIntVectorTy(1)); |
| 1466 | |
| 1467 | if (Trunc->getOperand(i_nocapture: 0) != Val) |
| 1468 | return ValueLatticeElement::getOverdefined(); |
| 1469 | |
| 1470 | Type *Ty = Val->getType(); |
| 1471 | |
| 1472 | if (Trunc->hasNoUnsignedWrap()) { |
| 1473 | if (IsTrueDest) |
| 1474 | return ValueLatticeElement::get(C: ConstantInt::get(Ty, V: 1)); |
| 1475 | return ValueLatticeElement::get(C: Constant::getNullValue(Ty)); |
| 1476 | } |
| 1477 | |
| 1478 | if (IsTrueDest) |
| 1479 | return ValueLatticeElement::getNot(C: Constant::getNullValue(Ty)); |
| 1480 | return ValueLatticeElement::getNot(C: Constant::getAllOnesValue(Ty)); |
| 1481 | } |
| 1482 | |
| 1483 | // Handle conditions of the form |
| 1484 | // extractvalue(op.with.overflow(%x, C), 1). |
| 1485 | static ValueLatticeElement getValueFromOverflowCondition( |
| 1486 | Value *Val, WithOverflowInst *WO, bool IsTrueDest) { |
| 1487 | // TODO: This only works with a constant RHS for now. We could also compute |
| 1488 | // the range of the RHS, but this doesn't fit into the current structure of |
| 1489 | // the edge value calculation. |
| 1490 | const APInt *C; |
| 1491 | if (WO->getLHS() != Val || !match(V: WO->getRHS(), P: m_APInt(Res&: C))) |
| 1492 | return ValueLatticeElement::getOverdefined(); |
| 1493 | |
| 1494 | // Calculate the possible values of %x for which no overflow occurs. |
| 1495 | ConstantRange NWR = ConstantRange::makeExactNoWrapRegion( |
| 1496 | BinOp: WO->getBinaryOp(), Other: *C, NoWrapKind: WO->getNoWrapKind()); |
| 1497 | |
| 1498 | // If overflow is false, %x is constrained to NWR. If overflow is true, %x is |
| 1499 | // constrained to it's inverse (all values that might cause overflow). |
| 1500 | if (IsTrueDest) |
| 1501 | NWR = NWR.inverse(); |
| 1502 | return ValueLatticeElement::getRange(CR: NWR); |
| 1503 | } |
| 1504 | |
| 1505 | std::optional<ValueLatticeElement> |
| 1506 | LazyValueInfoImpl::getValueFromCondition(Value *Val, Value *Cond, |
| 1507 | bool IsTrueDest, bool UseBlockValue, |
| 1508 | unsigned Depth) { |
| 1509 | if (ICmpInst *ICI = dyn_cast<ICmpInst>(Val: Cond)) |
| 1510 | return getValueFromICmpCondition(Val, ICI, isTrueDest: IsTrueDest, UseBlockValue); |
| 1511 | |
| 1512 | if (auto *Trunc = dyn_cast<TruncInst>(Val: Cond)) |
| 1513 | return getValueFromTrunc(Val, Trunc, IsTrueDest); |
| 1514 | |
| 1515 | if (auto *EVI = dyn_cast<ExtractValueInst>(Val: Cond)) |
| 1516 | if (auto *WO = dyn_cast<WithOverflowInst>(Val: EVI->getAggregateOperand())) |
| 1517 | if (EVI->getNumIndices() == 1 && *EVI->idx_begin() == 1) |
| 1518 | return getValueFromOverflowCondition(Val, WO, IsTrueDest); |
| 1519 | |
| 1520 | if (++Depth == MaxAnalysisRecursionDepth) |
| 1521 | return ValueLatticeElement::getOverdefined(); |
| 1522 | |
| 1523 | Value *N; |
| 1524 | if (match(V: Cond, P: m_Not(V: m_Value(V&: N)))) |
| 1525 | return getValueFromCondition(Val, Cond: N, IsTrueDest: !IsTrueDest, UseBlockValue, Depth); |
| 1526 | |
| 1527 | Value *L, *R; |
| 1528 | bool IsAnd; |
| 1529 | if (match(V: Cond, P: m_LogicalAnd(L: m_Value(V&: L), R: m_Value(V&: R)))) |
| 1530 | IsAnd = true; |
| 1531 | else if (match(V: Cond, P: m_LogicalOr(L: m_Value(V&: L), R: m_Value(V&: R)))) |
| 1532 | IsAnd = false; |
| 1533 | else |
| 1534 | return ValueLatticeElement::getOverdefined(); |
| 1535 | |
| 1536 | std::optional<ValueLatticeElement> LV = |
| 1537 | getValueFromCondition(Val, Cond: L, IsTrueDest, UseBlockValue, Depth); |
| 1538 | if (!LV) |
| 1539 | return std::nullopt; |
| 1540 | std::optional<ValueLatticeElement> RV = |
| 1541 | getValueFromCondition(Val, Cond: R, IsTrueDest, UseBlockValue, Depth); |
| 1542 | if (!RV) |
| 1543 | return std::nullopt; |
| 1544 | |
| 1545 | // if (L && R) -> intersect L and R |
| 1546 | // if (!(L || R)) -> intersect !L and !R |
| 1547 | // if (L || R) -> union L and R |
| 1548 | // if (!(L && R)) -> union !L and !R |
| 1549 | if (IsTrueDest ^ IsAnd) { |
| 1550 | LV->mergeIn(RHS: *RV); |
| 1551 | return *LV; |
| 1552 | } |
| 1553 | |
| 1554 | return LV->intersect(Other: *RV); |
| 1555 | } |
| 1556 | |
| 1557 | // Return true if Usr has Op as an operand, otherwise false. |
| 1558 | static bool usesOperand(User *Usr, Value *Op) { |
| 1559 | return is_contained(Range: Usr->operands(), Element: Op); |
| 1560 | } |
| 1561 | |
| 1562 | // Return true if the instruction type of Val is supported by |
| 1563 | // constantFoldUser(). Currently CastInst, BinaryOperator and FreezeInst only. |
| 1564 | // Call this before calling constantFoldUser() to find out if it's even worth |
| 1565 | // attempting to call it. |
| 1566 | static bool isOperationFoldable(User *Usr) { |
| 1567 | return isa<CastInst>(Val: Usr) || isa<BinaryOperator>(Val: Usr) || isa<FreezeInst>(Val: Usr); |
| 1568 | } |
| 1569 | |
| 1570 | // Check if Usr can be simplified to an integer constant when the value of one |
| 1571 | // of its operands Op is an integer constant OpConstVal. If so, return it as an |
| 1572 | // lattice value range with a single element or otherwise return an overdefined |
| 1573 | // lattice value. |
| 1574 | static ValueLatticeElement constantFoldUser(User *Usr, Value *Op, |
| 1575 | const APInt &OpConstVal, |
| 1576 | const DataLayout &DL) { |
| 1577 | assert(isOperationFoldable(Usr) && "Precondition" ); |
| 1578 | Constant* OpConst = Constant::getIntegerValue(Ty: Op->getType(), V: OpConstVal); |
| 1579 | // Check if Usr can be simplified to a constant. |
| 1580 | if (auto *CI = dyn_cast<CastInst>(Val: Usr)) { |
| 1581 | assert(CI->getOperand(0) == Op && "Operand 0 isn't Op" ); |
| 1582 | if (auto *C = dyn_cast_or_null<ConstantInt>( |
| 1583 | Val: simplifyCastInst(CastOpc: CI->getOpcode(), Op: OpConst, |
| 1584 | Ty: CI->getDestTy(), Q: DL))) { |
| 1585 | return ValueLatticeElement::getRange(CR: ConstantRange(C->getValue())); |
| 1586 | } |
| 1587 | } else if (auto *BO = dyn_cast<BinaryOperator>(Val: Usr)) { |
| 1588 | bool Op0Match = BO->getOperand(i_nocapture: 0) == Op; |
| 1589 | bool Op1Match = BO->getOperand(i_nocapture: 1) == Op; |
| 1590 | assert((Op0Match || Op1Match) && |
| 1591 | "Operand 0 nor Operand 1 isn't a match" ); |
| 1592 | Value *LHS = Op0Match ? OpConst : BO->getOperand(i_nocapture: 0); |
| 1593 | Value *RHS = Op1Match ? OpConst : BO->getOperand(i_nocapture: 1); |
| 1594 | if (auto *C = dyn_cast_or_null<ConstantInt>( |
| 1595 | Val: simplifyBinOp(Opcode: BO->getOpcode(), LHS, RHS, Q: DL))) { |
| 1596 | return ValueLatticeElement::getRange(CR: ConstantRange(C->getValue())); |
| 1597 | } |
| 1598 | } else if (isa<FreezeInst>(Val: Usr)) { |
| 1599 | assert(cast<FreezeInst>(Usr)->getOperand(0) == Op && "Operand 0 isn't Op" ); |
| 1600 | return ValueLatticeElement::getRange(CR: ConstantRange(OpConstVal)); |
| 1601 | } |
| 1602 | return ValueLatticeElement::getOverdefined(); |
| 1603 | } |
| 1604 | |
| 1605 | /// Compute the value of Val on the edge BBFrom -> BBTo. |
| 1606 | std::optional<ValueLatticeElement> |
| 1607 | LazyValueInfoImpl::getEdgeValueLocal(Value *Val, BasicBlock *BBFrom, |
| 1608 | BasicBlock *BBTo, bool UseBlockValue) { |
| 1609 | // TODO: Handle more complex conditionals. If (v == 0 || v2 < 1) is false, we |
| 1610 | // know that v != 0. |
| 1611 | if (BranchInst *BI = dyn_cast<BranchInst>(Val: BBFrom->getTerminator())) { |
| 1612 | // If this is a conditional branch and only one successor goes to BBTo, then |
| 1613 | // we may be able to infer something from the condition. |
| 1614 | if (BI->isConditional() && |
| 1615 | BI->getSuccessor(i: 0) != BI->getSuccessor(i: 1)) { |
| 1616 | bool isTrueDest = BI->getSuccessor(i: 0) == BBTo; |
| 1617 | assert(BI->getSuccessor(!isTrueDest) == BBTo && |
| 1618 | "BBTo isn't a successor of BBFrom" ); |
| 1619 | Value *Condition = BI->getCondition(); |
| 1620 | |
| 1621 | // If V is the condition of the branch itself, then we know exactly what |
| 1622 | // it is. |
| 1623 | // NB: The condition on a `br` can't be a vector type. |
| 1624 | if (Condition == Val) |
| 1625 | return ValueLatticeElement::get(C: ConstantInt::get( |
| 1626 | Ty: Type::getInt1Ty(C&: Val->getContext()), V: isTrueDest)); |
| 1627 | |
| 1628 | // If the condition of the branch is an equality comparison, we may be |
| 1629 | // able to infer the value. |
| 1630 | std::optional<ValueLatticeElement> Result = |
| 1631 | getValueFromCondition(Val, Cond: Condition, IsTrueDest: isTrueDest, UseBlockValue); |
| 1632 | if (!Result) |
| 1633 | return std::nullopt; |
| 1634 | |
| 1635 | if (!Result->isOverdefined()) |
| 1636 | return Result; |
| 1637 | |
| 1638 | if (User *Usr = dyn_cast<User>(Val)) { |
| 1639 | assert(Result->isOverdefined() && "Result isn't overdefined" ); |
| 1640 | // Check with isOperationFoldable() first to avoid linearly iterating |
| 1641 | // over the operands unnecessarily which can be expensive for |
| 1642 | // instructions with many operands. |
| 1643 | if (isa<IntegerType>(Val: Usr->getType()) && isOperationFoldable(Usr)) { |
| 1644 | const DataLayout &DL = BBTo->getDataLayout(); |
| 1645 | if (usesOperand(Usr, Op: Condition)) { |
| 1646 | // If Val has Condition as an operand and Val can be folded into a |
| 1647 | // constant with either Condition == true or Condition == false, |
| 1648 | // propagate the constant. |
| 1649 | // eg. |
| 1650 | // ; %Val is true on the edge to %then. |
| 1651 | // %Val = and i1 %Condition, true. |
| 1652 | // br %Condition, label %then, label %else |
| 1653 | APInt ConditionVal(1, isTrueDest ? 1 : 0); |
| 1654 | Result = constantFoldUser(Usr, Op: Condition, OpConstVal: ConditionVal, DL); |
| 1655 | } else if (isa<TruncInst, ZExtInst, SExtInst>(Val: Usr)) { |
| 1656 | ValueLatticeElement OpLatticeVal = |
| 1657 | *getValueFromCondition(Val: Usr->getOperand(i: 0), Cond: Condition, |
| 1658 | IsTrueDest: isTrueDest, /*UseBlockValue*/ false); |
| 1659 | |
| 1660 | if (OpLatticeVal.isConstantRange()) { |
| 1661 | const unsigned ResultBitWidth = |
| 1662 | Usr->getType()->getScalarSizeInBits(); |
| 1663 | if (auto *Trunc = dyn_cast<TruncInst>(Val: Usr)) |
| 1664 | return ValueLatticeElement::getRange( |
| 1665 | CR: OpLatticeVal.getConstantRange().truncate( |
| 1666 | BitWidth: ResultBitWidth, NoWrapKind: Trunc->getNoWrapKind())); |
| 1667 | |
| 1668 | return ValueLatticeElement::getRange( |
| 1669 | CR: OpLatticeVal.getConstantRange().castOp( |
| 1670 | CastOp: cast<CastInst>(Val: Usr)->getOpcode(), BitWidth: ResultBitWidth)); |
| 1671 | } |
| 1672 | if (OpLatticeVal.isConstant()) { |
| 1673 | Constant *C = OpLatticeVal.getConstant(); |
| 1674 | if (auto *CastC = ConstantFoldCastOperand( |
| 1675 | Opcode: cast<CastInst>(Val: Usr)->getOpcode(), C, DestTy: Usr->getType(), DL)) |
| 1676 | return ValueLatticeElement::get(C: CastC); |
| 1677 | } |
| 1678 | return ValueLatticeElement::getOverdefined(); |
| 1679 | } else { |
| 1680 | // If one of Val's operand has an inferred value, we may be able to |
| 1681 | // infer the value of Val. |
| 1682 | // eg. |
| 1683 | // ; %Val is 94 on the edge to %then. |
| 1684 | // %Val = add i8 %Op, 1 |
| 1685 | // %Condition = icmp eq i8 %Op, 93 |
| 1686 | // br i1 %Condition, label %then, label %else |
| 1687 | for (unsigned i = 0; i < Usr->getNumOperands(); ++i) { |
| 1688 | Value *Op = Usr->getOperand(i); |
| 1689 | ValueLatticeElement OpLatticeVal = *getValueFromCondition( |
| 1690 | Val: Op, Cond: Condition, IsTrueDest: isTrueDest, /*UseBlockValue*/ false); |
| 1691 | if (std::optional<APInt> OpConst = |
| 1692 | OpLatticeVal.asConstantInteger()) { |
| 1693 | Result = constantFoldUser(Usr, Op, OpConstVal: *OpConst, DL); |
| 1694 | break; |
| 1695 | } |
| 1696 | } |
| 1697 | } |
| 1698 | } |
| 1699 | } |
| 1700 | if (!Result->isOverdefined()) |
| 1701 | return Result; |
| 1702 | } |
| 1703 | } |
| 1704 | |
| 1705 | // If the edge was formed by a switch on the value, then we may know exactly |
| 1706 | // what it is. |
| 1707 | if (SwitchInst *SI = dyn_cast<SwitchInst>(Val: BBFrom->getTerminator())) { |
| 1708 | Value *Condition = SI->getCondition(); |
| 1709 | if (!isa<IntegerType>(Val: Val->getType())) |
| 1710 | return ValueLatticeElement::getOverdefined(); |
| 1711 | bool ValUsesConditionAndMayBeFoldable = false; |
| 1712 | if (Condition != Val) { |
| 1713 | // Check if Val has Condition as an operand. |
| 1714 | if (User *Usr = dyn_cast<User>(Val)) |
| 1715 | ValUsesConditionAndMayBeFoldable = isOperationFoldable(Usr) && |
| 1716 | usesOperand(Usr, Op: Condition); |
| 1717 | if (!ValUsesConditionAndMayBeFoldable) |
| 1718 | return ValueLatticeElement::getOverdefined(); |
| 1719 | } |
| 1720 | assert((Condition == Val || ValUsesConditionAndMayBeFoldable) && |
| 1721 | "Condition != Val nor Val doesn't use Condition" ); |
| 1722 | |
| 1723 | bool DefaultCase = SI->getDefaultDest() == BBTo; |
| 1724 | unsigned BitWidth = Val->getType()->getIntegerBitWidth(); |
| 1725 | ConstantRange EdgesVals(BitWidth, DefaultCase/*isFullSet*/); |
| 1726 | |
| 1727 | for (auto Case : SI->cases()) { |
| 1728 | APInt CaseValue = Case.getCaseValue()->getValue(); |
| 1729 | ConstantRange EdgeVal(CaseValue); |
| 1730 | if (ValUsesConditionAndMayBeFoldable) { |
| 1731 | User *Usr = cast<User>(Val); |
| 1732 | const DataLayout &DL = BBTo->getDataLayout(); |
| 1733 | ValueLatticeElement EdgeLatticeVal = |
| 1734 | constantFoldUser(Usr, Op: Condition, OpConstVal: CaseValue, DL); |
| 1735 | if (EdgeLatticeVal.isOverdefined()) |
| 1736 | return ValueLatticeElement::getOverdefined(); |
| 1737 | EdgeVal = EdgeLatticeVal.getConstantRange(); |
| 1738 | } |
| 1739 | if (DefaultCase) { |
| 1740 | // It is possible that the default destination is the destination of |
| 1741 | // some cases. We cannot perform difference for those cases. |
| 1742 | // We know Condition != CaseValue in BBTo. In some cases we can use |
| 1743 | // this to infer Val == f(Condition) is != f(CaseValue). For now, we |
| 1744 | // only do this when f is identity (i.e. Val == Condition), but we |
| 1745 | // should be able to do this for any injective f. |
| 1746 | if (Case.getCaseSuccessor() != BBTo && Condition == Val) |
| 1747 | EdgesVals = EdgesVals.difference(CR: EdgeVal); |
| 1748 | } else if (Case.getCaseSuccessor() == BBTo) |
| 1749 | EdgesVals = EdgesVals.unionWith(CR: EdgeVal); |
| 1750 | } |
| 1751 | return ValueLatticeElement::getRange(CR: std::move(EdgesVals)); |
| 1752 | } |
| 1753 | return ValueLatticeElement::getOverdefined(); |
| 1754 | } |
| 1755 | |
| 1756 | /// Compute the value of Val on the edge BBFrom -> BBTo or the value at |
| 1757 | /// the basic block if the edge does not constrain Val. |
| 1758 | std::optional<ValueLatticeElement> |
| 1759 | LazyValueInfoImpl::getEdgeValue(Value *Val, BasicBlock *BBFrom, |
| 1760 | BasicBlock *BBTo, Instruction *CxtI) { |
| 1761 | // If already a constant, there is nothing to compute. |
| 1762 | if (Constant *VC = dyn_cast<Constant>(Val)) |
| 1763 | return ValueLatticeElement::get(C: VC); |
| 1764 | |
| 1765 | std::optional<ValueLatticeElement> LocalResult = |
| 1766 | getEdgeValueLocal(Val, BBFrom, BBTo, /*UseBlockValue*/ true); |
| 1767 | if (!LocalResult) |
| 1768 | return std::nullopt; |
| 1769 | |
| 1770 | if (hasSingleValue(Val: *LocalResult)) |
| 1771 | // Can't get any more precise here |
| 1772 | return LocalResult; |
| 1773 | |
| 1774 | std::optional<ValueLatticeElement> OptInBlock = |
| 1775 | getBlockValue(Val, BB: BBFrom, CxtI: BBFrom->getTerminator()); |
| 1776 | if (!OptInBlock) |
| 1777 | return std::nullopt; |
| 1778 | ValueLatticeElement &InBlock = *OptInBlock; |
| 1779 | |
| 1780 | // We can use the context instruction (generically the ultimate instruction |
| 1781 | // the calling pass is trying to simplify) here, even though the result of |
| 1782 | // this function is generally cached when called from the solve* functions |
| 1783 | // (and that cached result might be used with queries using a different |
| 1784 | // context instruction), because when this function is called from the solve* |
| 1785 | // functions, the context instruction is not provided. When called from |
| 1786 | // LazyValueInfoImpl::getValueOnEdge, the context instruction is provided, |
| 1787 | // but then the result is not cached. |
| 1788 | intersectAssumeOrGuardBlockValueConstantRange(Val, BBLV&: InBlock, BBI: CxtI); |
| 1789 | |
| 1790 | return LocalResult->intersect(Other: InBlock); |
| 1791 | } |
| 1792 | |
| 1793 | ValueLatticeElement LazyValueInfoImpl::getValueInBlock(Value *V, BasicBlock *BB, |
| 1794 | Instruction *CxtI) { |
| 1795 | LLVM_DEBUG(dbgs() << "LVI Getting block end value " << *V << " at '" |
| 1796 | << BB->getName() << "'\n" ); |
| 1797 | |
| 1798 | assert(BlockValueStack.empty() && BlockValueSet.empty()); |
| 1799 | std::optional<ValueLatticeElement> OptResult = getBlockValue(Val: V, BB, CxtI); |
| 1800 | if (!OptResult) { |
| 1801 | solve(); |
| 1802 | OptResult = getBlockValue(Val: V, BB, CxtI); |
| 1803 | assert(OptResult && "Value not available after solving" ); |
| 1804 | } |
| 1805 | |
| 1806 | LLVM_DEBUG(dbgs() << " Result = " << *OptResult << "\n" ); |
| 1807 | return *OptResult; |
| 1808 | } |
| 1809 | |
| 1810 | ValueLatticeElement LazyValueInfoImpl::getValueAt(Value *V, Instruction *CxtI) { |
| 1811 | LLVM_DEBUG(dbgs() << "LVI Getting value " << *V << " at '" << CxtI->getName() |
| 1812 | << "'\n" ); |
| 1813 | |
| 1814 | if (auto *C = dyn_cast<Constant>(Val: V)) |
| 1815 | return ValueLatticeElement::get(C); |
| 1816 | |
| 1817 | ValueLatticeElement Result = ValueLatticeElement::getOverdefined(); |
| 1818 | if (auto *I = dyn_cast<Instruction>(Val: V)) |
| 1819 | Result = getFromRangeMetadata(BBI: I); |
| 1820 | intersectAssumeOrGuardBlockValueConstantRange(Val: V, BBLV&: Result, BBI: CxtI); |
| 1821 | |
| 1822 | LLVM_DEBUG(dbgs() << " Result = " << Result << "\n" ); |
| 1823 | return Result; |
| 1824 | } |
| 1825 | |
| 1826 | ValueLatticeElement LazyValueInfoImpl:: |
| 1827 | getValueOnEdge(Value *V, BasicBlock *FromBB, BasicBlock *ToBB, |
| 1828 | Instruction *CxtI) { |
| 1829 | LLVM_DEBUG(dbgs() << "LVI Getting edge value " << *V << " from '" |
| 1830 | << FromBB->getName() << "' to '" << ToBB->getName() |
| 1831 | << "'\n" ); |
| 1832 | |
| 1833 | std::optional<ValueLatticeElement> Result = |
| 1834 | getEdgeValue(Val: V, BBFrom: FromBB, BBTo: ToBB, CxtI); |
| 1835 | while (!Result) { |
| 1836 | // As the worklist only explicitly tracks block values (but not edge values) |
| 1837 | // we may have to call solve() multiple times, as the edge value calculation |
| 1838 | // may request additional block values. |
| 1839 | solve(); |
| 1840 | Result = getEdgeValue(Val: V, BBFrom: FromBB, BBTo: ToBB, CxtI); |
| 1841 | } |
| 1842 | |
| 1843 | LLVM_DEBUG(dbgs() << " Result = " << *Result << "\n" ); |
| 1844 | return *Result; |
| 1845 | } |
| 1846 | |
| 1847 | ValueLatticeElement LazyValueInfoImpl::getValueAtUse(const Use &U) { |
| 1848 | Value *V = U.get(); |
| 1849 | auto *CxtI = cast<Instruction>(Val: U.getUser()); |
| 1850 | ValueLatticeElement VL = getValueInBlock(V, BB: CxtI->getParent(), CxtI); |
| 1851 | |
| 1852 | // Check whether the only (possibly transitive) use of the value is in a |
| 1853 | // position where V can be constrained by a select or branch condition. |
| 1854 | const Use *CurrU = &U; |
| 1855 | // TODO: Increase limit? |
| 1856 | const unsigned MaxUsesToInspect = 3; |
| 1857 | for (unsigned I = 0; I < MaxUsesToInspect; ++I) { |
| 1858 | std::optional<ValueLatticeElement> CondVal; |
| 1859 | auto *CurrI = cast<Instruction>(Val: CurrU->getUser()); |
| 1860 | if (auto *SI = dyn_cast<SelectInst>(Val: CurrI)) { |
| 1861 | // If the value is undef, a different value may be chosen in |
| 1862 | // the select condition and at use. |
| 1863 | if (!isGuaranteedNotToBeUndef(V: SI->getCondition(), AC)) |
| 1864 | break; |
| 1865 | if (CurrU->getOperandNo() == 1) |
| 1866 | CondVal = |
| 1867 | *getValueFromCondition(Val: V, Cond: SI->getCondition(), /*IsTrueDest*/ true, |
| 1868 | /*UseBlockValue*/ false); |
| 1869 | else if (CurrU->getOperandNo() == 2) |
| 1870 | CondVal = |
| 1871 | *getValueFromCondition(Val: V, Cond: SI->getCondition(), /*IsTrueDest*/ false, |
| 1872 | /*UseBlockValue*/ false); |
| 1873 | } else if (auto *PHI = dyn_cast<PHINode>(Val: CurrI)) { |
| 1874 | // TODO: Use non-local query? |
| 1875 | CondVal = *getEdgeValueLocal(Val: V, BBFrom: PHI->getIncomingBlock(U: *CurrU), |
| 1876 | BBTo: PHI->getParent(), /*UseBlockValue*/ false); |
| 1877 | } |
| 1878 | if (CondVal) |
| 1879 | VL = VL.intersect(Other: *CondVal); |
| 1880 | |
| 1881 | // Only follow one-use chain, to allow direct intersection of conditions. |
| 1882 | // If there are multiple uses, we would have to intersect with the union of |
| 1883 | // all conditions at different uses. |
| 1884 | // Stop walking if we hit a non-speculatable instruction. Even if the |
| 1885 | // result is only used under a specific condition, executing the |
| 1886 | // instruction itself may cause side effects or UB already. |
| 1887 | // This also disallows looking through phi nodes: If the phi node is part |
| 1888 | // of a cycle, we might end up reasoning about values from different cycle |
| 1889 | // iterations (PR60629). |
| 1890 | if (!CurrI->hasOneUse() || |
| 1891 | !isSafeToSpeculativelyExecuteWithVariableReplaced( |
| 1892 | I: CurrI, /*IgnoreUBImplyingAttrs=*/false)) |
| 1893 | break; |
| 1894 | CurrU = &*CurrI->use_begin(); |
| 1895 | } |
| 1896 | return VL; |
| 1897 | } |
| 1898 | |
| 1899 | void LazyValueInfoImpl::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc, |
| 1900 | BasicBlock *NewSucc) { |
| 1901 | TheCache.threadEdgeImpl(OldSucc, NewSucc); |
| 1902 | } |
| 1903 | |
| 1904 | //===----------------------------------------------------------------------===// |
| 1905 | // LazyValueInfo Impl |
| 1906 | //===----------------------------------------------------------------------===// |
| 1907 | |
| 1908 | bool LazyValueInfoWrapperPass::runOnFunction(Function &F) { |
| 1909 | Info.AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); |
| 1910 | |
| 1911 | if (auto *Impl = Info.getImpl()) |
| 1912 | Impl->clear(); |
| 1913 | |
| 1914 | // Fully lazy. |
| 1915 | return false; |
| 1916 | } |
| 1917 | |
| 1918 | void LazyValueInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { |
| 1919 | AU.setPreservesAll(); |
| 1920 | AU.addRequired<AssumptionCacheTracker>(); |
| 1921 | AU.addRequired<TargetLibraryInfoWrapperPass>(); |
| 1922 | } |
| 1923 | |
| 1924 | LazyValueInfo &LazyValueInfoWrapperPass::getLVI() { return Info; } |
| 1925 | |
| 1926 | /// This lazily constructs the LazyValueInfoImpl. |
| 1927 | LazyValueInfoImpl &LazyValueInfo::getOrCreateImpl(const Module *M) { |
| 1928 | if (!PImpl) { |
| 1929 | assert(M && "getCache() called with a null Module" ); |
| 1930 | const DataLayout &DL = M->getDataLayout(); |
| 1931 | Function *GuardDecl = |
| 1932 | Intrinsic::getDeclarationIfExists(M, id: Intrinsic::experimental_guard); |
| 1933 | PImpl = new LazyValueInfoImpl(AC, DL, GuardDecl); |
| 1934 | } |
| 1935 | return *PImpl; |
| 1936 | } |
| 1937 | |
| 1938 | LazyValueInfoImpl *LazyValueInfo::getImpl() { return PImpl; } |
| 1939 | |
| 1940 | LazyValueInfo::~LazyValueInfo() { releaseMemory(); } |
| 1941 | |
| 1942 | void LazyValueInfo::releaseMemory() { |
| 1943 | // If the cache was allocated, free it. |
| 1944 | if (auto *Impl = getImpl()) { |
| 1945 | delete &*Impl; |
| 1946 | PImpl = nullptr; |
| 1947 | } |
| 1948 | } |
| 1949 | |
| 1950 | bool LazyValueInfo::invalidate(Function &F, const PreservedAnalyses &PA, |
| 1951 | FunctionAnalysisManager::Invalidator &Inv) { |
| 1952 | // We need to invalidate if we have either failed to preserve this analyses |
| 1953 | // result directly or if any of its dependencies have been invalidated. |
| 1954 | auto PAC = PA.getChecker<LazyValueAnalysis>(); |
| 1955 | if (!(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>())) |
| 1956 | return true; |
| 1957 | |
| 1958 | return false; |
| 1959 | } |
| 1960 | |
| 1961 | void LazyValueInfoWrapperPass::releaseMemory() { Info.releaseMemory(); } |
| 1962 | |
| 1963 | LazyValueInfo LazyValueAnalysis::run(Function &F, |
| 1964 | FunctionAnalysisManager &FAM) { |
| 1965 | auto &AC = FAM.getResult<AssumptionAnalysis>(IR&: F); |
| 1966 | |
| 1967 | return LazyValueInfo(&AC, &F.getDataLayout()); |
| 1968 | } |
| 1969 | |
| 1970 | /// Returns true if we can statically tell that this value will never be a |
| 1971 | /// "useful" constant. In practice, this means we've got something like an |
| 1972 | /// alloca or a malloc call for which a comparison against a constant can |
| 1973 | /// only be guarding dead code. Note that we are potentially giving up some |
| 1974 | /// precision in dead code (a constant result) in favour of avoiding a |
| 1975 | /// expensive search for a easily answered common query. |
| 1976 | static bool isKnownNonConstant(Value *V) { |
| 1977 | V = V->stripPointerCasts(); |
| 1978 | // The return val of alloc cannot be a Constant. |
| 1979 | if (isa<AllocaInst>(Val: V)) |
| 1980 | return true; |
| 1981 | return false; |
| 1982 | } |
| 1983 | |
| 1984 | Constant *LazyValueInfo::getConstant(Value *V, Instruction *CxtI) { |
| 1985 | // Bail out early if V is known not to be a Constant. |
| 1986 | if (isKnownNonConstant(V)) |
| 1987 | return nullptr; |
| 1988 | |
| 1989 | BasicBlock *BB = CxtI->getParent(); |
| 1990 | ValueLatticeElement Result = |
| 1991 | getOrCreateImpl(M: BB->getModule()).getValueInBlock(V, BB, CxtI); |
| 1992 | |
| 1993 | if (Result.isConstant()) |
| 1994 | return Result.getConstant(); |
| 1995 | if (Result.isConstantRange()) { |
| 1996 | const ConstantRange &CR = Result.getConstantRange(); |
| 1997 | if (const APInt *SingleVal = CR.getSingleElement()) |
| 1998 | return ConstantInt::get(Ty: V->getType(), V: *SingleVal); |
| 1999 | } |
| 2000 | return nullptr; |
| 2001 | } |
| 2002 | |
| 2003 | ConstantRange LazyValueInfo::getConstantRange(Value *V, Instruction *CxtI, |
| 2004 | bool UndefAllowed) { |
| 2005 | BasicBlock *BB = CxtI->getParent(); |
| 2006 | ValueLatticeElement Result = |
| 2007 | getOrCreateImpl(M: BB->getModule()).getValueInBlock(V, BB, CxtI); |
| 2008 | return Result.asConstantRange(Ty: V->getType(), UndefAllowed); |
| 2009 | } |
| 2010 | |
| 2011 | ConstantRange LazyValueInfo::getConstantRangeAtUse(const Use &U, |
| 2012 | bool UndefAllowed) { |
| 2013 | auto *Inst = cast<Instruction>(Val: U.getUser()); |
| 2014 | ValueLatticeElement Result = |
| 2015 | getOrCreateImpl(M: Inst->getModule()).getValueAtUse(U); |
| 2016 | return Result.asConstantRange(Ty: U->getType(), UndefAllowed); |
| 2017 | } |
| 2018 | |
| 2019 | /// Determine whether the specified value is known to be a |
| 2020 | /// constant on the specified edge. Return null if not. |
| 2021 | Constant *LazyValueInfo::getConstantOnEdge(Value *V, BasicBlock *FromBB, |
| 2022 | BasicBlock *ToBB, |
| 2023 | Instruction *CxtI) { |
| 2024 | Module *M = FromBB->getModule(); |
| 2025 | ValueLatticeElement Result = |
| 2026 | getOrCreateImpl(M).getValueOnEdge(V, FromBB, ToBB, CxtI); |
| 2027 | |
| 2028 | if (Result.isConstant()) |
| 2029 | return Result.getConstant(); |
| 2030 | if (Result.isConstantRange()) { |
| 2031 | const ConstantRange &CR = Result.getConstantRange(); |
| 2032 | if (const APInt *SingleVal = CR.getSingleElement()) |
| 2033 | return ConstantInt::get(Ty: V->getType(), V: *SingleVal); |
| 2034 | } |
| 2035 | return nullptr; |
| 2036 | } |
| 2037 | |
| 2038 | ConstantRange LazyValueInfo::getConstantRangeOnEdge(Value *V, |
| 2039 | BasicBlock *FromBB, |
| 2040 | BasicBlock *ToBB, |
| 2041 | Instruction *CxtI) { |
| 2042 | Module *M = FromBB->getModule(); |
| 2043 | ValueLatticeElement Result = |
| 2044 | getOrCreateImpl(M).getValueOnEdge(V, FromBB, ToBB, CxtI); |
| 2045 | // TODO: Should undef be allowed here? |
| 2046 | return Result.asConstantRange(Ty: V->getType(), /*UndefAllowed*/ true); |
| 2047 | } |
| 2048 | |
| 2049 | static Constant *getPredicateResult(CmpInst::Predicate Pred, Constant *C, |
| 2050 | const ValueLatticeElement &Val, |
| 2051 | const DataLayout &DL) { |
| 2052 | // If we know the value is a constant, evaluate the conditional. |
| 2053 | if (Val.isConstant()) |
| 2054 | return ConstantFoldCompareInstOperands(Predicate: Pred, LHS: Val.getConstant(), RHS: C, DL); |
| 2055 | |
| 2056 | Type *ResTy = CmpInst::makeCmpResultType(opnd_type: C->getType()); |
| 2057 | if (Val.isConstantRange()) { |
| 2058 | const ConstantRange &CR = Val.getConstantRange(); |
| 2059 | ConstantRange RHS = C->toConstantRange(); |
| 2060 | if (CR.icmp(Pred, Other: RHS)) |
| 2061 | return ConstantInt::getTrue(Ty: ResTy); |
| 2062 | if (CR.icmp(Pred: CmpInst::getInversePredicate(pred: Pred), Other: RHS)) |
| 2063 | return ConstantInt::getFalse(Ty: ResTy); |
| 2064 | return nullptr; |
| 2065 | } |
| 2066 | |
| 2067 | if (Val.isNotConstant()) { |
| 2068 | // If this is an equality comparison, we can try to fold it knowing that |
| 2069 | // "V != C1". |
| 2070 | if (Pred == ICmpInst::ICMP_EQ) { |
| 2071 | // !C1 == C -> false iff C1 == C. |
| 2072 | Constant *Res = ConstantFoldCompareInstOperands( |
| 2073 | Predicate: ICmpInst::ICMP_NE, LHS: Val.getNotConstant(), RHS: C, DL); |
| 2074 | if (Res && Res->isNullValue()) |
| 2075 | return ConstantInt::getFalse(Ty: ResTy); |
| 2076 | } else if (Pred == ICmpInst::ICMP_NE) { |
| 2077 | // !C1 != C -> true iff C1 == C. |
| 2078 | Constant *Res = ConstantFoldCompareInstOperands( |
| 2079 | Predicate: ICmpInst::ICMP_NE, LHS: Val.getNotConstant(), RHS: C, DL); |
| 2080 | if (Res && Res->isNullValue()) |
| 2081 | return ConstantInt::getTrue(Ty: ResTy); |
| 2082 | } |
| 2083 | return nullptr; |
| 2084 | } |
| 2085 | |
| 2086 | return nullptr; |
| 2087 | } |
| 2088 | |
| 2089 | /// Determine whether the specified value comparison with a constant is known to |
| 2090 | /// be true or false on the specified CFG edge. Pred is a CmpInst predicate. |
| 2091 | Constant *LazyValueInfo::getPredicateOnEdge(CmpInst::Predicate Pred, Value *V, |
| 2092 | Constant *C, BasicBlock *FromBB, |
| 2093 | BasicBlock *ToBB, |
| 2094 | Instruction *CxtI) { |
| 2095 | Module *M = FromBB->getModule(); |
| 2096 | ValueLatticeElement Result = |
| 2097 | getOrCreateImpl(M).getValueOnEdge(V, FromBB, ToBB, CxtI); |
| 2098 | |
| 2099 | return getPredicateResult(Pred, C, Val: Result, DL: M->getDataLayout()); |
| 2100 | } |
| 2101 | |
| 2102 | Constant *LazyValueInfo::getPredicateAt(CmpInst::Predicate Pred, Value *V, |
| 2103 | Constant *C, Instruction *CxtI, |
| 2104 | bool UseBlockValue) { |
| 2105 | // Is or is not NonNull are common predicates being queried. If |
| 2106 | // isKnownNonZero can tell us the result of the predicate, we can |
| 2107 | // return it quickly. But this is only a fastpath, and falling |
| 2108 | // through would still be correct. |
| 2109 | Module *M = CxtI->getModule(); |
| 2110 | const DataLayout &DL = M->getDataLayout(); |
| 2111 | if (V->getType()->isPointerTy() && C->isNullValue() && |
| 2112 | isKnownNonZero(V: V->stripPointerCastsSameRepresentation(), Q: DL)) { |
| 2113 | Type *ResTy = CmpInst::makeCmpResultType(opnd_type: C->getType()); |
| 2114 | if (Pred == ICmpInst::ICMP_EQ) |
| 2115 | return ConstantInt::getFalse(Ty: ResTy); |
| 2116 | else if (Pred == ICmpInst::ICMP_NE) |
| 2117 | return ConstantInt::getTrue(Ty: ResTy); |
| 2118 | } |
| 2119 | |
| 2120 | auto &Impl = getOrCreateImpl(M); |
| 2121 | ValueLatticeElement Result = |
| 2122 | UseBlockValue ? Impl.getValueInBlock(V, BB: CxtI->getParent(), CxtI) |
| 2123 | : Impl.getValueAt(V, CxtI); |
| 2124 | Constant *Ret = getPredicateResult(Pred, C, Val: Result, DL); |
| 2125 | if (Ret) |
| 2126 | return Ret; |
| 2127 | |
| 2128 | // Note: The following bit of code is somewhat distinct from the rest of LVI; |
| 2129 | // LVI as a whole tries to compute a lattice value which is conservatively |
| 2130 | // correct at a given location. In this case, we have a predicate which we |
| 2131 | // weren't able to prove about the merged result, and we're pushing that |
| 2132 | // predicate back along each incoming edge to see if we can prove it |
| 2133 | // separately for each input. As a motivating example, consider: |
| 2134 | // bb1: |
| 2135 | // %v1 = ... ; constantrange<1, 5> |
| 2136 | // br label %merge |
| 2137 | // bb2: |
| 2138 | // %v2 = ... ; constantrange<10, 20> |
| 2139 | // br label %merge |
| 2140 | // merge: |
| 2141 | // %phi = phi [%v1, %v2] ; constantrange<1,20> |
| 2142 | // %pred = icmp eq i32 %phi, 8 |
| 2143 | // We can't tell from the lattice value for '%phi' that '%pred' is false |
| 2144 | // along each path, but by checking the predicate over each input separately, |
| 2145 | // we can. |
| 2146 | // We limit the search to one step backwards from the current BB and value. |
| 2147 | // We could consider extending this to search further backwards through the |
| 2148 | // CFG and/or value graph, but there are non-obvious compile time vs quality |
| 2149 | // tradeoffs. |
| 2150 | BasicBlock *BB = CxtI->getParent(); |
| 2151 | |
| 2152 | // Function entry or an unreachable block. Bail to avoid confusing |
| 2153 | // analysis below. |
| 2154 | pred_iterator PI = pred_begin(BB), PE = pred_end(BB); |
| 2155 | if (PI == PE) |
| 2156 | return nullptr; |
| 2157 | |
| 2158 | // If V is a PHI node in the same block as the context, we need to ask |
| 2159 | // questions about the predicate as applied to the incoming value along |
| 2160 | // each edge. This is useful for eliminating cases where the predicate is |
| 2161 | // known along all incoming edges. |
| 2162 | if (auto *PHI = dyn_cast<PHINode>(Val: V)) |
| 2163 | if (PHI->getParent() == BB) { |
| 2164 | Constant *Baseline = nullptr; |
| 2165 | for (unsigned i = 0, e = PHI->getNumIncomingValues(); i < e; i++) { |
| 2166 | Value *Incoming = PHI->getIncomingValue(i); |
| 2167 | BasicBlock *PredBB = PHI->getIncomingBlock(i); |
| 2168 | // Note that PredBB may be BB itself. |
| 2169 | Constant *Result = |
| 2170 | getPredicateOnEdge(Pred, V: Incoming, C, FromBB: PredBB, ToBB: BB, CxtI); |
| 2171 | |
| 2172 | // Keep going as long as we've seen a consistent known result for |
| 2173 | // all inputs. |
| 2174 | Baseline = (i == 0) ? Result /* First iteration */ |
| 2175 | : (Baseline == Result ? Baseline |
| 2176 | : nullptr); /* All others */ |
| 2177 | if (!Baseline) |
| 2178 | break; |
| 2179 | } |
| 2180 | if (Baseline) |
| 2181 | return Baseline; |
| 2182 | } |
| 2183 | |
| 2184 | // For a comparison where the V is outside this block, it's possible |
| 2185 | // that we've branched on it before. Look to see if the value is known |
| 2186 | // on all incoming edges. |
| 2187 | if (!isa<Instruction>(Val: V) || cast<Instruction>(Val: V)->getParent() != BB) { |
| 2188 | // For predecessor edge, determine if the comparison is true or false |
| 2189 | // on that edge. If they're all true or all false, we can conclude |
| 2190 | // the value of the comparison in this block. |
| 2191 | Constant *Baseline = getPredicateOnEdge(Pred, V, C, FromBB: *PI, ToBB: BB, CxtI); |
| 2192 | if (Baseline) { |
| 2193 | // Check that all remaining incoming values match the first one. |
| 2194 | while (++PI != PE) { |
| 2195 | Constant *Ret = getPredicateOnEdge(Pred, V, C, FromBB: *PI, ToBB: BB, CxtI); |
| 2196 | if (Ret != Baseline) |
| 2197 | break; |
| 2198 | } |
| 2199 | // If we terminated early, then one of the values didn't match. |
| 2200 | if (PI == PE) { |
| 2201 | return Baseline; |
| 2202 | } |
| 2203 | } |
| 2204 | } |
| 2205 | |
| 2206 | return nullptr; |
| 2207 | } |
| 2208 | |
| 2209 | Constant *LazyValueInfo::getPredicateAt(CmpInst::Predicate Pred, Value *LHS, |
| 2210 | Value *RHS, Instruction *CxtI, |
| 2211 | bool UseBlockValue) { |
| 2212 | if (auto *C = dyn_cast<Constant>(Val: RHS)) |
| 2213 | return getPredicateAt(Pred, V: LHS, C, CxtI, UseBlockValue); |
| 2214 | if (auto *C = dyn_cast<Constant>(Val: LHS)) |
| 2215 | return getPredicateAt(Pred: CmpInst::getSwappedPredicate(pred: Pred), V: RHS, C, CxtI, |
| 2216 | UseBlockValue); |
| 2217 | |
| 2218 | // Got two non-Constant values. Try to determine the comparison results based |
| 2219 | // on the block values of the two operands, e.g. because they have |
| 2220 | // non-overlapping ranges. |
| 2221 | if (UseBlockValue) { |
| 2222 | Module *M = CxtI->getModule(); |
| 2223 | ValueLatticeElement L = |
| 2224 | getOrCreateImpl(M).getValueInBlock(V: LHS, BB: CxtI->getParent(), CxtI); |
| 2225 | if (L.isOverdefined()) |
| 2226 | return nullptr; |
| 2227 | |
| 2228 | ValueLatticeElement R = |
| 2229 | getOrCreateImpl(M).getValueInBlock(V: RHS, BB: CxtI->getParent(), CxtI); |
| 2230 | Type *Ty = CmpInst::makeCmpResultType(opnd_type: LHS->getType()); |
| 2231 | return L.getCompare(Pred, Ty, Other: R, DL: M->getDataLayout()); |
| 2232 | } |
| 2233 | return nullptr; |
| 2234 | } |
| 2235 | |
| 2236 | void LazyValueInfo::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc, |
| 2237 | BasicBlock *NewSucc) { |
| 2238 | if (auto *Impl = getImpl()) |
| 2239 | Impl->threadEdge(PredBB, OldSucc, NewSucc); |
| 2240 | } |
| 2241 | |
| 2242 | void LazyValueInfo::forgetValue(Value *V) { |
| 2243 | if (auto *Impl = getImpl()) |
| 2244 | Impl->forgetValue(V); |
| 2245 | } |
| 2246 | |
| 2247 | void LazyValueInfo::eraseBlock(BasicBlock *BB) { |
| 2248 | if (auto *Impl = getImpl()) |
| 2249 | Impl->eraseBlock(BB); |
| 2250 | } |
| 2251 | |
| 2252 | void LazyValueInfo::clear() { |
| 2253 | if (auto *Impl = getImpl()) |
| 2254 | Impl->clear(); |
| 2255 | } |
| 2256 | |
| 2257 | void LazyValueInfo::printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) { |
| 2258 | if (auto *Impl = getImpl()) |
| 2259 | Impl->printLVI(F, DTree, OS); |
| 2260 | } |
| 2261 | |
| 2262 | // Print the LVI for the function arguments at the start of each basic block. |
| 2263 | void LazyValueInfoAnnotatedWriter::emitBasicBlockStartAnnot( |
| 2264 | const BasicBlock *BB, formatted_raw_ostream &OS) { |
| 2265 | // Find if there are latticevalues defined for arguments of the function. |
| 2266 | auto *F = BB->getParent(); |
| 2267 | for (const auto &Arg : F->args()) { |
| 2268 | ValueLatticeElement Result = LVIImpl->getValueInBlock( |
| 2269 | V: const_cast<Argument *>(&Arg), BB: const_cast<BasicBlock *>(BB)); |
| 2270 | if (Result.isUnknown()) |
| 2271 | continue; |
| 2272 | OS << "; LatticeVal for: '" << Arg << "' is: " << Result << "\n" ; |
| 2273 | } |
| 2274 | } |
| 2275 | |
| 2276 | // This function prints the LVI analysis for the instruction I at the beginning |
| 2277 | // of various basic blocks. It relies on calculated values that are stored in |
| 2278 | // the LazyValueInfoCache, and in the absence of cached values, recalculate the |
| 2279 | // LazyValueInfo for `I`, and print that info. |
| 2280 | void LazyValueInfoAnnotatedWriter::emitInstructionAnnot( |
| 2281 | const Instruction *I, formatted_raw_ostream &OS) { |
| 2282 | |
| 2283 | auto *ParentBB = I->getParent(); |
| 2284 | SmallPtrSet<const BasicBlock*, 16> BlocksContainingLVI; |
| 2285 | // We can generate (solve) LVI values only for blocks that are dominated by |
| 2286 | // the I's parent. However, to avoid generating LVI for all dominating blocks, |
| 2287 | // that contain redundant/uninteresting information, we print LVI for |
| 2288 | // blocks that may use this LVI information (such as immediate successor |
| 2289 | // blocks, and blocks that contain uses of `I`). |
| 2290 | auto printResult = [&](const BasicBlock *BB) { |
| 2291 | if (!BlocksContainingLVI.insert(Ptr: BB).second) |
| 2292 | return; |
| 2293 | ValueLatticeElement Result = LVIImpl->getValueInBlock( |
| 2294 | V: const_cast<Instruction *>(I), BB: const_cast<BasicBlock *>(BB)); |
| 2295 | OS << "; LatticeVal for: '" << *I << "' in BB: '" ; |
| 2296 | BB->printAsOperand(O&: OS, PrintType: false); |
| 2297 | OS << "' is: " << Result << "\n" ; |
| 2298 | }; |
| 2299 | |
| 2300 | printResult(ParentBB); |
| 2301 | // Print the LVI analysis results for the immediate successor blocks, that |
| 2302 | // are dominated by `ParentBB`. |
| 2303 | for (const auto *BBSucc : successors(BB: ParentBB)) |
| 2304 | if (DT.dominates(A: ParentBB, B: BBSucc)) |
| 2305 | printResult(BBSucc); |
| 2306 | |
| 2307 | // Print LVI in blocks where `I` is used. |
| 2308 | for (const auto *U : I->users()) |
| 2309 | if (auto *UseI = dyn_cast<Instruction>(Val: U)) |
| 2310 | if (!isa<PHINode>(Val: UseI) || DT.dominates(A: ParentBB, B: UseI->getParent())) |
| 2311 | printResult(UseI->getParent()); |
| 2312 | |
| 2313 | } |
| 2314 | |
| 2315 | PreservedAnalyses LazyValueInfoPrinterPass::run(Function &F, |
| 2316 | FunctionAnalysisManager &AM) { |
| 2317 | OS << "LVI for function '" << F.getName() << "':\n" ; |
| 2318 | auto &LVI = AM.getResult<LazyValueAnalysis>(IR&: F); |
| 2319 | auto &DTree = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 2320 | LVI.printLVI(F, DTree, OS); |
| 2321 | return PreservedAnalyses::all(); |
| 2322 | } |
| 2323 | |