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/Optional.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/Analysis/AssumptionCache.h" 19 #include "llvm/Analysis/ConstantFolding.h" 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/Analysis/TargetLibraryInfo.h" 22 #include "llvm/Analysis/ValueTracking.h" 23 #include "llvm/Analysis/ValueLattice.h" 24 #include "llvm/IR/AssemblyAnnotationWriter.h" 25 #include "llvm/IR/CFG.h" 26 #include "llvm/IR/ConstantRange.h" 27 #include "llvm/IR/Constants.h" 28 #include "llvm/IR/DataLayout.h" 29 #include "llvm/IR/Dominators.h" 30 #include "llvm/IR/Instructions.h" 31 #include "llvm/IR/IntrinsicInst.h" 32 #include "llvm/IR/Intrinsics.h" 33 #include "llvm/IR/LLVMContext.h" 34 #include "llvm/IR/PatternMatch.h" 35 #include "llvm/IR/ValueHandle.h" 36 #include "llvm/Support/Debug.h" 37 #include "llvm/Support/FormattedStream.h" 38 #include "llvm/Support/raw_ostream.h" 39 #include <map> 40 using namespace llvm; 41 using namespace PatternMatch; 42 43 #define DEBUG_TYPE "lazy-value-info" 44 45 // This is the number of worklist items we will process to try to discover an 46 // answer for a given value. 47 static const unsigned MaxProcessedPerValue = 500; 48 49 char LazyValueInfoWrapperPass::ID = 0; 50 INITIALIZE_PASS_BEGIN(LazyValueInfoWrapperPass, "lazy-value-info", 51 "Lazy Value Information Analysis", false, true) 52 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 53 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 54 INITIALIZE_PASS_END(LazyValueInfoWrapperPass, "lazy-value-info", 55 "Lazy Value Information Analysis", false, true) 56 57 namespace llvm { 58 FunctionPass *createLazyValueInfoPass() { return new LazyValueInfoWrapperPass(); } 59 } 60 61 AnalysisKey LazyValueAnalysis::Key; 62 63 /// Returns true if this lattice value represents at most one possible value. 64 /// This is as precise as any lattice value can get while still representing 65 /// reachable code. 66 static bool hasSingleValue(const ValueLatticeElement &Val) { 67 if (Val.isConstantRange() && 68 Val.getConstantRange().isSingleElement()) 69 // Integer constants are single element ranges 70 return true; 71 if (Val.isConstant()) 72 // Non integer constants 73 return true; 74 return false; 75 } 76 77 /// Combine two sets of facts about the same value into a single set of 78 /// facts. Note that this method is not suitable for merging facts along 79 /// different paths in a CFG; that's what the mergeIn function is for. This 80 /// is for merging facts gathered about the same value at the same location 81 /// through two independent means. 82 /// Notes: 83 /// * This method does not promise to return the most precise possible lattice 84 /// value implied by A and B. It is allowed to return any lattice element 85 /// which is at least as strong as *either* A or B (unless our facts 86 /// conflict, see below). 87 /// * Due to unreachable code, the intersection of two lattice values could be 88 /// contradictory. If this happens, we return some valid lattice value so as 89 /// not confuse the rest of LVI. Ideally, we'd always return Undefined, but 90 /// we do not make this guarantee. TODO: This would be a useful enhancement. 91 static ValueLatticeElement intersect(const ValueLatticeElement &A, 92 const ValueLatticeElement &B) { 93 // Undefined is the strongest state. It means the value is known to be along 94 // an unreachable path. 95 if (A.isUndefined()) 96 return A; 97 if (B.isUndefined()) 98 return B; 99 100 // If we gave up for one, but got a useable fact from the other, use it. 101 if (A.isOverdefined()) 102 return B; 103 if (B.isOverdefined()) 104 return A; 105 106 // Can't get any more precise than constants. 107 if (hasSingleValue(A)) 108 return A; 109 if (hasSingleValue(B)) 110 return B; 111 112 // Could be either constant range or not constant here. 113 if (!A.isConstantRange() || !B.isConstantRange()) { 114 // TODO: Arbitrary choice, could be improved 115 return A; 116 } 117 118 // Intersect two constant ranges 119 ConstantRange Range = 120 A.getConstantRange().intersectWith(B.getConstantRange()); 121 // Note: An empty range is implicitly converted to overdefined internally. 122 // TODO: We could instead use Undefined here since we've proven a conflict 123 // and thus know this path must be unreachable. 124 return ValueLatticeElement::getRange(std::move(Range)); 125 } 126 127 //===----------------------------------------------------------------------===// 128 // LazyValueInfoCache Decl 129 //===----------------------------------------------------------------------===// 130 131 namespace { 132 /// A callback value handle updates the cache when values are erased. 133 class LazyValueInfoCache; 134 struct LVIValueHandle final : public CallbackVH { 135 // Needs to access getValPtr(), which is protected. 136 friend struct DenseMapInfo<LVIValueHandle>; 137 138 LazyValueInfoCache *Parent; 139 140 LVIValueHandle(Value *V, LazyValueInfoCache *P) 141 : CallbackVH(V), Parent(P) { } 142 143 void deleted() override; 144 void allUsesReplacedWith(Value *V) override { 145 deleted(); 146 } 147 }; 148 } // end anonymous namespace 149 150 namespace { 151 /// This is the cache kept by LazyValueInfo which 152 /// maintains information about queries across the clients' queries. 153 class LazyValueInfoCache { 154 /// This is all of the cached block information for exactly one Value*. 155 /// The entries are sorted by the BasicBlock* of the 156 /// entries, allowing us to do a lookup with a binary search. 157 /// Over-defined lattice values are recorded in OverDefinedCache to reduce 158 /// memory overhead. 159 struct ValueCacheEntryTy { 160 ValueCacheEntryTy(Value *V, LazyValueInfoCache *P) : Handle(V, P) {} 161 LVIValueHandle Handle; 162 SmallDenseMap<PoisoningVH<BasicBlock>, ValueLatticeElement, 4> BlockVals; 163 }; 164 165 /// This tracks, on a per-block basis, the set of values that are 166 /// over-defined at the end of that block. 167 typedef DenseMap<PoisoningVH<BasicBlock>, SmallPtrSet<Value *, 4>> 168 OverDefinedCacheTy; 169 /// Keep track of all blocks that we have ever seen, so we 170 /// don't spend time removing unused blocks from our caches. 171 DenseSet<PoisoningVH<BasicBlock> > SeenBlocks; 172 173 /// This is all of the cached information for all values, 174 /// mapped from Value* to key information. 175 DenseMap<Value *, std::unique_ptr<ValueCacheEntryTy>> ValueCache; 176 OverDefinedCacheTy OverDefinedCache; 177 178 179 public: 180 void insertResult(Value *Val, BasicBlock *BB, 181 const ValueLatticeElement &Result) { 182 SeenBlocks.insert(BB); 183 184 // Insert over-defined values into their own cache to reduce memory 185 // overhead. 186 if (Result.isOverdefined()) 187 OverDefinedCache[BB].insert(Val); 188 else { 189 auto It = ValueCache.find_as(Val); 190 if (It == ValueCache.end()) { 191 ValueCache[Val] = make_unique<ValueCacheEntryTy>(Val, this); 192 It = ValueCache.find_as(Val); 193 assert(It != ValueCache.end() && "Val was just added to the map!"); 194 } 195 It->second->BlockVals[BB] = Result; 196 } 197 } 198 199 bool isOverdefined(Value *V, BasicBlock *BB) const { 200 auto ODI = OverDefinedCache.find(BB); 201 202 if (ODI == OverDefinedCache.end()) 203 return false; 204 205 return ODI->second.count(V); 206 } 207 208 bool hasCachedValueInfo(Value *V, BasicBlock *BB) const { 209 if (isOverdefined(V, BB)) 210 return true; 211 212 auto I = ValueCache.find_as(V); 213 if (I == ValueCache.end()) 214 return false; 215 216 return I->second->BlockVals.count(BB); 217 } 218 219 ValueLatticeElement getCachedValueInfo(Value *V, BasicBlock *BB) const { 220 if (isOverdefined(V, BB)) 221 return ValueLatticeElement::getOverdefined(); 222 223 auto I = ValueCache.find_as(V); 224 if (I == ValueCache.end()) 225 return ValueLatticeElement(); 226 auto BBI = I->second->BlockVals.find(BB); 227 if (BBI == I->second->BlockVals.end()) 228 return ValueLatticeElement(); 229 return BBI->second; 230 } 231 232 /// clear - Empty the cache. 233 void clear() { 234 SeenBlocks.clear(); 235 ValueCache.clear(); 236 OverDefinedCache.clear(); 237 } 238 239 /// Inform the cache that a given value has been deleted. 240 void eraseValue(Value *V); 241 242 /// This is part of the update interface to inform the cache 243 /// that a block has been deleted. 244 void eraseBlock(BasicBlock *BB); 245 246 /// Updates the cache to remove any influence an overdefined value in 247 /// OldSucc might have (unless also overdefined in NewSucc). This just 248 /// flushes elements from the cache and does not add any. 249 void threadEdgeImpl(BasicBlock *OldSucc,BasicBlock *NewSucc); 250 251 friend struct LVIValueHandle; 252 }; 253 } 254 255 void LazyValueInfoCache::eraseValue(Value *V) { 256 for (auto I = OverDefinedCache.begin(), E = OverDefinedCache.end(); I != E;) { 257 // Copy and increment the iterator immediately so we can erase behind 258 // ourselves. 259 auto Iter = I++; 260 SmallPtrSetImpl<Value *> &ValueSet = Iter->second; 261 ValueSet.erase(V); 262 if (ValueSet.empty()) 263 OverDefinedCache.erase(Iter); 264 } 265 266 ValueCache.erase(V); 267 } 268 269 void LVIValueHandle::deleted() { 270 // This erasure deallocates *this, so it MUST happen after we're done 271 // using any and all members of *this. 272 Parent->eraseValue(*this); 273 } 274 275 void LazyValueInfoCache::eraseBlock(BasicBlock *BB) { 276 // Shortcut if we have never seen this block. 277 DenseSet<PoisoningVH<BasicBlock> >::iterator I = SeenBlocks.find(BB); 278 if (I == SeenBlocks.end()) 279 return; 280 SeenBlocks.erase(I); 281 282 auto ODI = OverDefinedCache.find(BB); 283 if (ODI != OverDefinedCache.end()) 284 OverDefinedCache.erase(ODI); 285 286 for (auto &I : ValueCache) 287 I.second->BlockVals.erase(BB); 288 } 289 290 void LazyValueInfoCache::threadEdgeImpl(BasicBlock *OldSucc, 291 BasicBlock *NewSucc) { 292 // When an edge in the graph has been threaded, values that we could not 293 // determine a value for before (i.e. were marked overdefined) may be 294 // possible to solve now. We do NOT try to proactively update these values. 295 // Instead, we clear their entries from the cache, and allow lazy updating to 296 // recompute them when needed. 297 298 // The updating process is fairly simple: we need to drop cached info 299 // for all values that were marked overdefined in OldSucc, and for those same 300 // values in any successor of OldSucc (except NewSucc) in which they were 301 // also marked overdefined. 302 std::vector<BasicBlock*> worklist; 303 worklist.push_back(OldSucc); 304 305 auto I = OverDefinedCache.find(OldSucc); 306 if (I == OverDefinedCache.end()) 307 return; // Nothing to process here. 308 SmallVector<Value *, 4> ValsToClear(I->second.begin(), I->second.end()); 309 310 // Use a worklist to perform a depth-first search of OldSucc's successors. 311 // NOTE: We do not need a visited list since any blocks we have already 312 // visited will have had their overdefined markers cleared already, and we 313 // thus won't loop to their successors. 314 while (!worklist.empty()) { 315 BasicBlock *ToUpdate = worklist.back(); 316 worklist.pop_back(); 317 318 // Skip blocks only accessible through NewSucc. 319 if (ToUpdate == NewSucc) continue; 320 321 // If a value was marked overdefined in OldSucc, and is here too... 322 auto OI = OverDefinedCache.find(ToUpdate); 323 if (OI == OverDefinedCache.end()) 324 continue; 325 SmallPtrSetImpl<Value *> &ValueSet = OI->second; 326 327 bool changed = false; 328 for (Value *V : ValsToClear) { 329 if (!ValueSet.erase(V)) 330 continue; 331 332 // If we removed anything, then we potentially need to update 333 // blocks successors too. 334 changed = true; 335 336 if (ValueSet.empty()) { 337 OverDefinedCache.erase(OI); 338 break; 339 } 340 } 341 342 if (!changed) continue; 343 344 worklist.insert(worklist.end(), succ_begin(ToUpdate), succ_end(ToUpdate)); 345 } 346 } 347 348 349 namespace { 350 /// An assembly annotator class to print LazyValueCache information in 351 /// comments. 352 class LazyValueInfoImpl; 353 class LazyValueInfoAnnotatedWriter : public AssemblyAnnotationWriter { 354 LazyValueInfoImpl *LVIImpl; 355 // While analyzing which blocks we can solve values for, we need the dominator 356 // information. Since this is an optional parameter in LVI, we require this 357 // DomTreeAnalysis pass in the printer pass, and pass the dominator 358 // tree to the LazyValueInfoAnnotatedWriter. 359 DominatorTree &DT; 360 361 public: 362 LazyValueInfoAnnotatedWriter(LazyValueInfoImpl *L, DominatorTree &DTree) 363 : LVIImpl(L), DT(DTree) {} 364 365 virtual void emitBasicBlockStartAnnot(const BasicBlock *BB, 366 formatted_raw_ostream &OS); 367 368 virtual void emitInstructionAnnot(const Instruction *I, 369 formatted_raw_ostream &OS); 370 }; 371 } 372 namespace { 373 // The actual implementation of the lazy analysis and update. Note that the 374 // inheritance from LazyValueInfoCache is intended to be temporary while 375 // splitting the code and then transitioning to a has-a relationship. 376 class LazyValueInfoImpl { 377 378 /// Cached results from previous queries 379 LazyValueInfoCache TheCache; 380 381 /// This stack holds the state of the value solver during a query. 382 /// It basically emulates the callstack of the naive 383 /// recursive value lookup process. 384 SmallVector<std::pair<BasicBlock*, Value*>, 8> BlockValueStack; 385 386 /// Keeps track of which block-value pairs are in BlockValueStack. 387 DenseSet<std::pair<BasicBlock*, Value*> > BlockValueSet; 388 389 /// Push BV onto BlockValueStack unless it's already in there. 390 /// Returns true on success. 391 bool pushBlockValue(const std::pair<BasicBlock *, Value *> &BV) { 392 if (!BlockValueSet.insert(BV).second) 393 return false; // It's already in the stack. 394 395 LLVM_DEBUG(dbgs() << "PUSH: " << *BV.second << " in " 396 << BV.first->getName() << "\n"); 397 BlockValueStack.push_back(BV); 398 return true; 399 } 400 401 AssumptionCache *AC; ///< A pointer to the cache of @llvm.assume calls. 402 const DataLayout &DL; ///< A mandatory DataLayout 403 DominatorTree *DT; ///< An optional DT pointer. 404 DominatorTree *DisabledDT; ///< Stores DT if it's disabled. 405 406 ValueLatticeElement getBlockValue(Value *Val, BasicBlock *BB); 407 bool getEdgeValue(Value *V, BasicBlock *F, BasicBlock *T, 408 ValueLatticeElement &Result, Instruction *CxtI = nullptr); 409 bool hasBlockValue(Value *Val, BasicBlock *BB); 410 411 // These methods process one work item and may add more. A false value 412 // returned means that the work item was not completely processed and must 413 // be revisited after going through the new items. 414 bool solveBlockValue(Value *Val, BasicBlock *BB); 415 bool solveBlockValueImpl(ValueLatticeElement &Res, Value *Val, 416 BasicBlock *BB); 417 bool solveBlockValueNonLocal(ValueLatticeElement &BBLV, Value *Val, 418 BasicBlock *BB); 419 bool solveBlockValuePHINode(ValueLatticeElement &BBLV, PHINode *PN, 420 BasicBlock *BB); 421 bool solveBlockValueSelect(ValueLatticeElement &BBLV, SelectInst *S, 422 BasicBlock *BB); 423 Optional<ConstantRange> getRangeForOperand(unsigned Op, Instruction *I, 424 BasicBlock *BB); 425 bool solveBlockValueBinaryOp(ValueLatticeElement &BBLV, BinaryOperator *BBI, 426 BasicBlock *BB); 427 bool solveBlockValueCast(ValueLatticeElement &BBLV, CastInst *CI, 428 BasicBlock *BB); 429 void intersectAssumeOrGuardBlockValueConstantRange(Value *Val, 430 ValueLatticeElement &BBLV, 431 Instruction *BBI); 432 433 void solve(); 434 435 public: 436 /// This is the query interface to determine the lattice 437 /// value for the specified Value* at the end of the specified block. 438 ValueLatticeElement getValueInBlock(Value *V, BasicBlock *BB, 439 Instruction *CxtI = nullptr); 440 441 /// This is the query interface to determine the lattice 442 /// value for the specified Value* at the specified instruction (generally 443 /// from an assume intrinsic). 444 ValueLatticeElement getValueAt(Value *V, Instruction *CxtI); 445 446 /// This is the query interface to determine the lattice 447 /// value for the specified Value* that is true on the specified edge. 448 ValueLatticeElement getValueOnEdge(Value *V, BasicBlock *FromBB, 449 BasicBlock *ToBB, 450 Instruction *CxtI = nullptr); 451 452 /// Complete flush all previously computed values 453 void clear() { 454 TheCache.clear(); 455 } 456 457 /// Printing the LazyValueInfo Analysis. 458 void printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) { 459 LazyValueInfoAnnotatedWriter Writer(this, DTree); 460 F.print(OS, &Writer); 461 } 462 463 /// This is part of the update interface to inform the cache 464 /// that a block has been deleted. 465 void eraseBlock(BasicBlock *BB) { 466 TheCache.eraseBlock(BB); 467 } 468 469 /// Disables use of the DominatorTree within LVI. 470 void disableDT() { 471 if (DT) { 472 assert(!DisabledDT && "Both DT and DisabledDT are not nullptr!"); 473 std::swap(DT, DisabledDT); 474 } 475 } 476 477 /// Enables use of the DominatorTree within LVI. Does nothing if the class 478 /// instance was initialized without a DT pointer. 479 void enableDT() { 480 if (DisabledDT) { 481 assert(!DT && "Both DT and DisabledDT are not nullptr!"); 482 std::swap(DT, DisabledDT); 483 } 484 } 485 486 /// This is the update interface to inform the cache that an edge from 487 /// PredBB to OldSucc has been threaded to be from PredBB to NewSucc. 488 void threadEdge(BasicBlock *PredBB,BasicBlock *OldSucc,BasicBlock *NewSucc); 489 490 LazyValueInfoImpl(AssumptionCache *AC, const DataLayout &DL, 491 DominatorTree *DT = nullptr) 492 : AC(AC), DL(DL), DT(DT), DisabledDT(nullptr) {} 493 }; 494 } // end anonymous namespace 495 496 497 void LazyValueInfoImpl::solve() { 498 SmallVector<std::pair<BasicBlock *, Value *>, 8> StartingStack( 499 BlockValueStack.begin(), BlockValueStack.end()); 500 501 unsigned processedCount = 0; 502 while (!BlockValueStack.empty()) { 503 processedCount++; 504 // Abort if we have to process too many values to get a result for this one. 505 // Because of the design of the overdefined cache currently being per-block 506 // to avoid naming-related issues (IE it wants to try to give different 507 // results for the same name in different blocks), overdefined results don't 508 // get cached globally, which in turn means we will often try to rediscover 509 // the same overdefined result again and again. Once something like 510 // PredicateInfo is used in LVI or CVP, we should be able to make the 511 // overdefined cache global, and remove this throttle. 512 if (processedCount > MaxProcessedPerValue) { 513 LLVM_DEBUG( 514 dbgs() << "Giving up on stack because we are getting too deep\n"); 515 // Fill in the original values 516 while (!StartingStack.empty()) { 517 std::pair<BasicBlock *, Value *> &e = StartingStack.back(); 518 TheCache.insertResult(e.second, e.first, 519 ValueLatticeElement::getOverdefined()); 520 StartingStack.pop_back(); 521 } 522 BlockValueSet.clear(); 523 BlockValueStack.clear(); 524 return; 525 } 526 std::pair<BasicBlock *, Value *> e = BlockValueStack.back(); 527 assert(BlockValueSet.count(e) && "Stack value should be in BlockValueSet!"); 528 529 if (solveBlockValue(e.second, e.first)) { 530 // The work item was completely processed. 531 assert(BlockValueStack.back() == e && "Nothing should have been pushed!"); 532 assert(TheCache.hasCachedValueInfo(e.second, e.first) && 533 "Result should be in cache!"); 534 535 LLVM_DEBUG( 536 dbgs() << "POP " << *e.second << " in " << e.first->getName() << " = " 537 << TheCache.getCachedValueInfo(e.second, e.first) << "\n"); 538 539 BlockValueStack.pop_back(); 540 BlockValueSet.erase(e); 541 } else { 542 // More work needs to be done before revisiting. 543 assert(BlockValueStack.back() != e && "Stack should have been pushed!"); 544 } 545 } 546 } 547 548 bool LazyValueInfoImpl::hasBlockValue(Value *Val, BasicBlock *BB) { 549 // If already a constant, there is nothing to compute. 550 if (isa<Constant>(Val)) 551 return true; 552 553 return TheCache.hasCachedValueInfo(Val, BB); 554 } 555 556 ValueLatticeElement LazyValueInfoImpl::getBlockValue(Value *Val, 557 BasicBlock *BB) { 558 // If already a constant, there is nothing to compute. 559 if (Constant *VC = dyn_cast<Constant>(Val)) 560 return ValueLatticeElement::get(VC); 561 562 return TheCache.getCachedValueInfo(Val, BB); 563 } 564 565 static ValueLatticeElement getFromRangeMetadata(Instruction *BBI) { 566 switch (BBI->getOpcode()) { 567 default: break; 568 case Instruction::Load: 569 case Instruction::Call: 570 case Instruction::Invoke: 571 if (MDNode *Ranges = BBI->getMetadata(LLVMContext::MD_range)) 572 if (isa<IntegerType>(BBI->getType())) { 573 return ValueLatticeElement::getRange( 574 getConstantRangeFromMetadata(*Ranges)); 575 } 576 break; 577 }; 578 // Nothing known - will be intersected with other facts 579 return ValueLatticeElement::getOverdefined(); 580 } 581 582 bool LazyValueInfoImpl::solveBlockValue(Value *Val, BasicBlock *BB) { 583 if (isa<Constant>(Val)) 584 return true; 585 586 if (TheCache.hasCachedValueInfo(Val, BB)) { 587 // If we have a cached value, use that. 588 LLVM_DEBUG(dbgs() << " reuse BB '" << BB->getName() << "' val=" 589 << TheCache.getCachedValueInfo(Val, BB) << '\n'); 590 591 // Since we're reusing a cached value, we don't need to update the 592 // OverDefinedCache. The cache will have been properly updated whenever the 593 // cached value was inserted. 594 return true; 595 } 596 597 // Hold off inserting this value into the Cache in case we have to return 598 // false and come back later. 599 ValueLatticeElement Res; 600 if (!solveBlockValueImpl(Res, Val, BB)) 601 // Work pushed, will revisit 602 return false; 603 604 TheCache.insertResult(Val, BB, Res); 605 return true; 606 } 607 608 bool LazyValueInfoImpl::solveBlockValueImpl(ValueLatticeElement &Res, 609 Value *Val, BasicBlock *BB) { 610 611 Instruction *BBI = dyn_cast<Instruction>(Val); 612 if (!BBI || BBI->getParent() != BB) 613 return solveBlockValueNonLocal(Res, Val, BB); 614 615 if (PHINode *PN = dyn_cast<PHINode>(BBI)) 616 return solveBlockValuePHINode(Res, PN, BB); 617 618 if (auto *SI = dyn_cast<SelectInst>(BBI)) 619 return solveBlockValueSelect(Res, SI, BB); 620 621 // If this value is a nonnull pointer, record it's range and bailout. Note 622 // that for all other pointer typed values, we terminate the search at the 623 // definition. We could easily extend this to look through geps, bitcasts, 624 // and the like to prove non-nullness, but it's not clear that's worth it 625 // compile time wise. The context-insensitive value walk done inside 626 // isKnownNonZero gets most of the profitable cases at much less expense. 627 // This does mean that we have a sensitivity to where the defining 628 // instruction is placed, even if it could legally be hoisted much higher. 629 // That is unfortunate. 630 PointerType *PT = dyn_cast<PointerType>(BBI->getType()); 631 if (PT && isKnownNonZero(BBI, DL)) { 632 Res = ValueLatticeElement::getNot(ConstantPointerNull::get(PT)); 633 return true; 634 } 635 if (BBI->getType()->isIntegerTy()) { 636 if (auto *CI = dyn_cast<CastInst>(BBI)) 637 return solveBlockValueCast(Res, CI, BB); 638 639 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(BBI)) 640 return solveBlockValueBinaryOp(Res, BO, BB); 641 } 642 643 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() 644 << "' - unknown inst def found.\n"); 645 Res = getFromRangeMetadata(BBI); 646 return true; 647 } 648 649 static bool InstructionDereferencesPointer(Instruction *I, Value *Ptr) { 650 if (LoadInst *L = dyn_cast<LoadInst>(I)) { 651 return L->getPointerAddressSpace() == 0 && 652 GetUnderlyingObject(L->getPointerOperand(), 653 L->getModule()->getDataLayout()) == Ptr; 654 } 655 if (StoreInst *S = dyn_cast<StoreInst>(I)) { 656 return S->getPointerAddressSpace() == 0 && 657 GetUnderlyingObject(S->getPointerOperand(), 658 S->getModule()->getDataLayout()) == Ptr; 659 } 660 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) { 661 if (MI->isVolatile()) return false; 662 663 // FIXME: check whether it has a valuerange that excludes zero? 664 ConstantInt *Len = dyn_cast<ConstantInt>(MI->getLength()); 665 if (!Len || Len->isZero()) return false; 666 667 if (MI->getDestAddressSpace() == 0) 668 if (GetUnderlyingObject(MI->getRawDest(), 669 MI->getModule()->getDataLayout()) == Ptr) 670 return true; 671 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) 672 if (MTI->getSourceAddressSpace() == 0) 673 if (GetUnderlyingObject(MTI->getRawSource(), 674 MTI->getModule()->getDataLayout()) == Ptr) 675 return true; 676 } 677 return false; 678 } 679 680 /// Return true if the allocation associated with Val is ever dereferenced 681 /// within the given basic block. This establishes the fact Val is not null, 682 /// but does not imply that the memory at Val is dereferenceable. (Val may 683 /// point off the end of the dereferenceable part of the object.) 684 static bool isObjectDereferencedInBlock(Value *Val, BasicBlock *BB) { 685 assert(Val->getType()->isPointerTy()); 686 687 const DataLayout &DL = BB->getModule()->getDataLayout(); 688 Value *UnderlyingVal = GetUnderlyingObject(Val, DL); 689 // If 'GetUnderlyingObject' didn't converge, skip it. It won't converge 690 // inside InstructionDereferencesPointer either. 691 if (UnderlyingVal == GetUnderlyingObject(UnderlyingVal, DL, 1)) 692 for (Instruction &I : *BB) 693 if (InstructionDereferencesPointer(&I, UnderlyingVal)) 694 return true; 695 return false; 696 } 697 698 bool LazyValueInfoImpl::solveBlockValueNonLocal(ValueLatticeElement &BBLV, 699 Value *Val, BasicBlock *BB) { 700 ValueLatticeElement Result; // Start Undefined. 701 702 // If this is the entry block, we must be asking about an argument. The 703 // value is overdefined. 704 if (BB == &BB->getParent()->getEntryBlock()) { 705 assert(isa<Argument>(Val) && "Unknown live-in to the entry block"); 706 // Before giving up, see if we can prove the pointer non-null local to 707 // this particular block. 708 PointerType *PTy = dyn_cast<PointerType>(Val->getType()); 709 if (PTy && 710 (isKnownNonZero(Val, DL) || 711 (isObjectDereferencedInBlock(Val, BB) && 712 !NullPointerIsDefined(BB->getParent(), PTy->getAddressSpace())))) { 713 Result = ValueLatticeElement::getNot(ConstantPointerNull::get(PTy)); 714 } else { 715 Result = ValueLatticeElement::getOverdefined(); 716 } 717 BBLV = Result; 718 return true; 719 } 720 721 // Loop over all of our predecessors, merging what we know from them into 722 // result. If we encounter an unexplored predecessor, we eagerly explore it 723 // in a depth first manner. In practice, this has the effect of discovering 724 // paths we can't analyze eagerly without spending compile times analyzing 725 // other paths. This heuristic benefits from the fact that predecessors are 726 // frequently arranged such that dominating ones come first and we quickly 727 // find a path to function entry. TODO: We should consider explicitly 728 // canonicalizing to make this true rather than relying on this happy 729 // accident. 730 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 731 ValueLatticeElement EdgeResult; 732 if (!getEdgeValue(Val, *PI, BB, EdgeResult)) 733 // Explore that input, then return here 734 return false; 735 736 Result.mergeIn(EdgeResult, DL); 737 738 // If we hit overdefined, exit early. The BlockVals entry is already set 739 // to overdefined. 740 if (Result.isOverdefined()) { 741 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() 742 << "' - overdefined because of pred (non local).\n"); 743 // Before giving up, see if we can prove the pointer non-null local to 744 // this particular block. 745 PointerType *PTy = dyn_cast<PointerType>(Val->getType()); 746 if (PTy && isObjectDereferencedInBlock(Val, BB) && 747 !NullPointerIsDefined(BB->getParent(), PTy->getAddressSpace())) { 748 Result = ValueLatticeElement::getNot(ConstantPointerNull::get(PTy)); 749 } 750 751 BBLV = Result; 752 return true; 753 } 754 } 755 756 // Return the merged value, which is more precise than 'overdefined'. 757 assert(!Result.isOverdefined()); 758 BBLV = Result; 759 return true; 760 } 761 762 bool LazyValueInfoImpl::solveBlockValuePHINode(ValueLatticeElement &BBLV, 763 PHINode *PN, BasicBlock *BB) { 764 ValueLatticeElement Result; // Start Undefined. 765 766 // Loop over all of our predecessors, merging what we know from them into 767 // result. See the comment about the chosen traversal order in 768 // solveBlockValueNonLocal; the same reasoning applies here. 769 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 770 BasicBlock *PhiBB = PN->getIncomingBlock(i); 771 Value *PhiVal = PN->getIncomingValue(i); 772 ValueLatticeElement EdgeResult; 773 // Note that we can provide PN as the context value to getEdgeValue, even 774 // though the results will be cached, because PN is the value being used as 775 // the cache key in the caller. 776 if (!getEdgeValue(PhiVal, PhiBB, BB, EdgeResult, PN)) 777 // Explore that input, then return here 778 return false; 779 780 Result.mergeIn(EdgeResult, DL); 781 782 // If we hit overdefined, exit early. The BlockVals entry is already set 783 // to overdefined. 784 if (Result.isOverdefined()) { 785 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() 786 << "' - overdefined because of pred (local).\n"); 787 788 BBLV = Result; 789 return true; 790 } 791 } 792 793 // Return the merged value, which is more precise than 'overdefined'. 794 assert(!Result.isOverdefined() && "Possible PHI in entry block?"); 795 BBLV = Result; 796 return true; 797 } 798 799 static ValueLatticeElement getValueFromCondition(Value *Val, Value *Cond, 800 bool isTrueDest = true); 801 802 // If we can determine a constraint on the value given conditions assumed by 803 // the program, intersect those constraints with BBLV 804 void LazyValueInfoImpl::intersectAssumeOrGuardBlockValueConstantRange( 805 Value *Val, ValueLatticeElement &BBLV, Instruction *BBI) { 806 BBI = BBI ? BBI : dyn_cast<Instruction>(Val); 807 if (!BBI) 808 return; 809 810 for (auto &AssumeVH : AC->assumptionsFor(Val)) { 811 if (!AssumeVH) 812 continue; 813 auto *I = cast<CallInst>(AssumeVH); 814 if (!isValidAssumeForContext(I, BBI, DT)) 815 continue; 816 817 BBLV = intersect(BBLV, getValueFromCondition(Val, I->getArgOperand(0))); 818 } 819 820 // If guards are not used in the module, don't spend time looking for them 821 auto *GuardDecl = BBI->getModule()->getFunction( 822 Intrinsic::getName(Intrinsic::experimental_guard)); 823 if (!GuardDecl || GuardDecl->use_empty()) 824 return; 825 826 if (BBI->getIterator() == BBI->getParent()->begin()) 827 return; 828 for (Instruction &I : make_range(std::next(BBI->getIterator().getReverse()), 829 BBI->getParent()->rend())) { 830 Value *Cond = nullptr; 831 if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>(m_Value(Cond)))) 832 BBLV = intersect(BBLV, getValueFromCondition(Val, Cond)); 833 } 834 } 835 836 bool LazyValueInfoImpl::solveBlockValueSelect(ValueLatticeElement &BBLV, 837 SelectInst *SI, BasicBlock *BB) { 838 839 // Recurse on our inputs if needed 840 if (!hasBlockValue(SI->getTrueValue(), BB)) { 841 if (pushBlockValue(std::make_pair(BB, SI->getTrueValue()))) 842 return false; 843 BBLV = ValueLatticeElement::getOverdefined(); 844 return true; 845 } 846 ValueLatticeElement TrueVal = getBlockValue(SI->getTrueValue(), BB); 847 // If we hit overdefined, don't ask more queries. We want to avoid poisoning 848 // extra slots in the table if we can. 849 if (TrueVal.isOverdefined()) { 850 BBLV = ValueLatticeElement::getOverdefined(); 851 return true; 852 } 853 854 if (!hasBlockValue(SI->getFalseValue(), BB)) { 855 if (pushBlockValue(std::make_pair(BB, SI->getFalseValue()))) 856 return false; 857 BBLV = ValueLatticeElement::getOverdefined(); 858 return true; 859 } 860 ValueLatticeElement FalseVal = getBlockValue(SI->getFalseValue(), BB); 861 // If we hit overdefined, don't ask more queries. We want to avoid poisoning 862 // extra slots in the table if we can. 863 if (FalseVal.isOverdefined()) { 864 BBLV = ValueLatticeElement::getOverdefined(); 865 return true; 866 } 867 868 if (TrueVal.isConstantRange() && FalseVal.isConstantRange()) { 869 const ConstantRange &TrueCR = TrueVal.getConstantRange(); 870 const ConstantRange &FalseCR = FalseVal.getConstantRange(); 871 Value *LHS = nullptr; 872 Value *RHS = nullptr; 873 SelectPatternResult SPR = matchSelectPattern(SI, LHS, RHS); 874 // Is this a min specifically of our two inputs? (Avoid the risk of 875 // ValueTracking getting smarter looking back past our immediate inputs.) 876 if (SelectPatternResult::isMinOrMax(SPR.Flavor) && 877 LHS == SI->getTrueValue() && RHS == SI->getFalseValue()) { 878 ConstantRange ResultCR = [&]() { 879 switch (SPR.Flavor) { 880 default: 881 llvm_unreachable("unexpected minmax type!"); 882 case SPF_SMIN: /// Signed minimum 883 return TrueCR.smin(FalseCR); 884 case SPF_UMIN: /// Unsigned minimum 885 return TrueCR.umin(FalseCR); 886 case SPF_SMAX: /// Signed maximum 887 return TrueCR.smax(FalseCR); 888 case SPF_UMAX: /// Unsigned maximum 889 return TrueCR.umax(FalseCR); 890 }; 891 }(); 892 BBLV = ValueLatticeElement::getRange(ResultCR); 893 return true; 894 } 895 896 // TODO: ABS, NABS from the SelectPatternResult 897 } 898 899 // Can we constrain the facts about the true and false values by using the 900 // condition itself? This shows up with idioms like e.g. select(a > 5, a, 5). 901 // TODO: We could potentially refine an overdefined true value above. 902 Value *Cond = SI->getCondition(); 903 TrueVal = intersect(TrueVal, 904 getValueFromCondition(SI->getTrueValue(), Cond, true)); 905 FalseVal = intersect(FalseVal, 906 getValueFromCondition(SI->getFalseValue(), Cond, false)); 907 908 // Handle clamp idioms such as: 909 // %24 = constantrange<0, 17> 910 // %39 = icmp eq i32 %24, 0 911 // %40 = add i32 %24, -1 912 // %siv.next = select i1 %39, i32 16, i32 %40 913 // %siv.next = constantrange<0, 17> not <-1, 17> 914 // In general, this can handle any clamp idiom which tests the edge 915 // condition via an equality or inequality. 916 if (auto *ICI = dyn_cast<ICmpInst>(Cond)) { 917 ICmpInst::Predicate Pred = ICI->getPredicate(); 918 Value *A = ICI->getOperand(0); 919 if (ConstantInt *CIBase = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 920 auto addConstants = [](ConstantInt *A, ConstantInt *B) { 921 assert(A->getType() == B->getType()); 922 return ConstantInt::get(A->getType(), A->getValue() + B->getValue()); 923 }; 924 // See if either input is A + C2, subject to the constraint from the 925 // condition that A != C when that input is used. We can assume that 926 // that input doesn't include C + C2. 927 ConstantInt *CIAdded; 928 switch (Pred) { 929 default: break; 930 case ICmpInst::ICMP_EQ: 931 if (match(SI->getFalseValue(), m_Add(m_Specific(A), 932 m_ConstantInt(CIAdded)))) { 933 auto ResNot = addConstants(CIBase, CIAdded); 934 FalseVal = intersect(FalseVal, 935 ValueLatticeElement::getNot(ResNot)); 936 } 937 break; 938 case ICmpInst::ICMP_NE: 939 if (match(SI->getTrueValue(), m_Add(m_Specific(A), 940 m_ConstantInt(CIAdded)))) { 941 auto ResNot = addConstants(CIBase, CIAdded); 942 TrueVal = intersect(TrueVal, 943 ValueLatticeElement::getNot(ResNot)); 944 } 945 break; 946 }; 947 } 948 } 949 950 ValueLatticeElement Result; // Start Undefined. 951 Result.mergeIn(TrueVal, DL); 952 Result.mergeIn(FalseVal, DL); 953 BBLV = Result; 954 return true; 955 } 956 957 Optional<ConstantRange> LazyValueInfoImpl::getRangeForOperand(unsigned Op, 958 Instruction *I, 959 BasicBlock *BB) { 960 if (!hasBlockValue(I->getOperand(Op), BB)) 961 if (pushBlockValue(std::make_pair(BB, I->getOperand(Op)))) 962 return None; 963 964 const unsigned OperandBitWidth = 965 DL.getTypeSizeInBits(I->getOperand(Op)->getType()); 966 ConstantRange Range = ConstantRange::getFull(OperandBitWidth); 967 if (hasBlockValue(I->getOperand(Op), BB)) { 968 ValueLatticeElement Val = getBlockValue(I->getOperand(Op), BB); 969 intersectAssumeOrGuardBlockValueConstantRange(I->getOperand(Op), Val, I); 970 if (Val.isConstantRange()) 971 Range = Val.getConstantRange(); 972 } 973 return Range; 974 } 975 976 bool LazyValueInfoImpl::solveBlockValueCast(ValueLatticeElement &BBLV, 977 CastInst *CI, 978 BasicBlock *BB) { 979 if (!CI->getOperand(0)->getType()->isSized()) { 980 // Without knowing how wide the input is, we can't analyze it in any useful 981 // way. 982 BBLV = ValueLatticeElement::getOverdefined(); 983 return true; 984 } 985 986 // Filter out casts we don't know how to reason about before attempting to 987 // recurse on our operand. This can cut a long search short if we know we're 988 // not going to be able to get any useful information anways. 989 switch (CI->getOpcode()) { 990 case Instruction::Trunc: 991 case Instruction::SExt: 992 case Instruction::ZExt: 993 case Instruction::BitCast: 994 break; 995 default: 996 // Unhandled instructions are overdefined. 997 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() 998 << "' - overdefined (unknown cast).\n"); 999 BBLV = ValueLatticeElement::getOverdefined(); 1000 return true; 1001 } 1002 1003 // Figure out the range of the LHS. If that fails, we still apply the 1004 // transfer rule on the full set since we may be able to locally infer 1005 // interesting facts. 1006 Optional<ConstantRange> LHSRes = getRangeForOperand(0, CI, BB); 1007 if (!LHSRes.hasValue()) 1008 // More work to do before applying this transfer rule. 1009 return false; 1010 ConstantRange LHSRange = LHSRes.getValue(); 1011 1012 const unsigned ResultBitWidth = CI->getType()->getIntegerBitWidth(); 1013 1014 // NOTE: We're currently limited by the set of operations that ConstantRange 1015 // can evaluate symbolically. Enhancing that set will allows us to analyze 1016 // more definitions. 1017 BBLV = ValueLatticeElement::getRange(LHSRange.castOp(CI->getOpcode(), 1018 ResultBitWidth)); 1019 return true; 1020 } 1021 1022 bool LazyValueInfoImpl::solveBlockValueBinaryOp(ValueLatticeElement &BBLV, 1023 BinaryOperator *BO, 1024 BasicBlock *BB) { 1025 1026 assert(BO->getOperand(0)->getType()->isSized() && 1027 "all operands to binary operators are sized"); 1028 1029 // Filter out operators we don't know how to reason about before attempting to 1030 // recurse on our operand(s). This can cut a long search short if we know 1031 // we're not going to be able to get any useful information anyways. 1032 switch (BO->getOpcode()) { 1033 case Instruction::Add: 1034 case Instruction::Sub: 1035 case Instruction::Mul: 1036 case Instruction::UDiv: 1037 case Instruction::Shl: 1038 case Instruction::LShr: 1039 case Instruction::AShr: 1040 case Instruction::And: 1041 case Instruction::Or: 1042 // continue into the code below 1043 break; 1044 default: 1045 // Unhandled instructions are overdefined. 1046 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() 1047 << "' - overdefined (unknown binary operator).\n"); 1048 BBLV = ValueLatticeElement::getOverdefined(); 1049 return true; 1050 }; 1051 1052 // Figure out the ranges of the operands. If that fails, use a 1053 // conservative range, but apply the transfer rule anyways. This 1054 // lets us pick up facts from expressions like "and i32 (call i32 1055 // @foo()), 32" 1056 Optional<ConstantRange> LHSRes = getRangeForOperand(0, BO, BB); 1057 Optional<ConstantRange> RHSRes = getRangeForOperand(1, BO, BB); 1058 1059 if (!LHSRes.hasValue() || !RHSRes.hasValue()) 1060 // More work to do before applying this transfer rule. 1061 return false; 1062 1063 ConstantRange LHSRange = LHSRes.getValue(); 1064 ConstantRange RHSRange = RHSRes.getValue(); 1065 1066 // NOTE: We're currently limited by the set of operations that ConstantRange 1067 // can evaluate symbolically. Enhancing that set will allows us to analyze 1068 // more definitions. 1069 Instruction::BinaryOps BinOp = BO->getOpcode(); 1070 BBLV = ValueLatticeElement::getRange(LHSRange.binaryOp(BinOp, RHSRange)); 1071 return true; 1072 } 1073 1074 static ValueLatticeElement getValueFromICmpCondition(Value *Val, ICmpInst *ICI, 1075 bool isTrueDest) { 1076 Value *LHS = ICI->getOperand(0); 1077 Value *RHS = ICI->getOperand(1); 1078 CmpInst::Predicate Predicate = ICI->getPredicate(); 1079 1080 if (isa<Constant>(RHS)) { 1081 if (ICI->isEquality() && LHS == Val) { 1082 // We know that V has the RHS constant if this is a true SETEQ or 1083 // false SETNE. 1084 if (isTrueDest == (Predicate == ICmpInst::ICMP_EQ)) 1085 return ValueLatticeElement::get(cast<Constant>(RHS)); 1086 else 1087 return ValueLatticeElement::getNot(cast<Constant>(RHS)); 1088 } 1089 } 1090 1091 if (!Val->getType()->isIntegerTy()) 1092 return ValueLatticeElement::getOverdefined(); 1093 1094 // Use ConstantRange::makeAllowedICmpRegion in order to determine the possible 1095 // range of Val guaranteed by the condition. Recognize comparisons in the from 1096 // of: 1097 // icmp <pred> Val, ... 1098 // icmp <pred> (add Val, Offset), ... 1099 // The latter is the range checking idiom that InstCombine produces. Subtract 1100 // the offset from the allowed range for RHS in this case. 1101 1102 // Val or (add Val, Offset) can be on either hand of the comparison 1103 if (LHS != Val && !match(LHS, m_Add(m_Specific(Val), m_ConstantInt()))) { 1104 std::swap(LHS, RHS); 1105 Predicate = CmpInst::getSwappedPredicate(Predicate); 1106 } 1107 1108 ConstantInt *Offset = nullptr; 1109 if (LHS != Val) 1110 match(LHS, m_Add(m_Specific(Val), m_ConstantInt(Offset))); 1111 1112 if (LHS == Val || Offset) { 1113 // Calculate the range of values that are allowed by the comparison 1114 ConstantRange RHSRange(RHS->getType()->getIntegerBitWidth(), 1115 /*isFullSet=*/true); 1116 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) 1117 RHSRange = ConstantRange(CI->getValue()); 1118 else if (Instruction *I = dyn_cast<Instruction>(RHS)) 1119 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range)) 1120 RHSRange = getConstantRangeFromMetadata(*Ranges); 1121 1122 // If we're interested in the false dest, invert the condition 1123 CmpInst::Predicate Pred = 1124 isTrueDest ? Predicate : CmpInst::getInversePredicate(Predicate); 1125 ConstantRange TrueValues = 1126 ConstantRange::makeAllowedICmpRegion(Pred, RHSRange); 1127 1128 if (Offset) // Apply the offset from above. 1129 TrueValues = TrueValues.subtract(Offset->getValue()); 1130 1131 return ValueLatticeElement::getRange(std::move(TrueValues)); 1132 } 1133 1134 return ValueLatticeElement::getOverdefined(); 1135 } 1136 1137 // Handle conditions of the form 1138 // extractvalue(op.with.overflow(%x, C), 1). 1139 static ValueLatticeElement getValueFromOverflowCondition( 1140 Value *Val, WithOverflowInst *WO, bool IsTrueDest) { 1141 // TODO: This only works with a constant RHS for now. We could also compute 1142 // the range of the RHS, but this doesn't fit into the current structure of 1143 // the edge value calculation. 1144 const APInt *C; 1145 if (WO->getLHS() != Val || !match(WO->getRHS(), m_APInt(C))) 1146 return ValueLatticeElement::getOverdefined(); 1147 1148 // Calculate the possible values of %x for which no overflow occurs. 1149 ConstantRange NWR = ConstantRange::makeExactNoWrapRegion( 1150 WO->getBinaryOp(), *C, WO->getNoWrapKind()); 1151 1152 // If overflow is false, %x is constrained to NWR. If overflow is true, %x is 1153 // constrained to it's inverse (all values that might cause overflow). 1154 if (IsTrueDest) 1155 NWR = NWR.inverse(); 1156 return ValueLatticeElement::getRange(NWR); 1157 } 1158 1159 static ValueLatticeElement 1160 getValueFromCondition(Value *Val, Value *Cond, bool isTrueDest, 1161 DenseMap<Value*, ValueLatticeElement> &Visited); 1162 1163 static ValueLatticeElement 1164 getValueFromConditionImpl(Value *Val, Value *Cond, bool isTrueDest, 1165 DenseMap<Value*, ValueLatticeElement> &Visited) { 1166 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cond)) 1167 return getValueFromICmpCondition(Val, ICI, isTrueDest); 1168 1169 if (auto *EVI = dyn_cast<ExtractValueInst>(Cond)) 1170 if (auto *WO = dyn_cast<WithOverflowInst>(EVI->getAggregateOperand())) 1171 if (EVI->getNumIndices() == 1 && *EVI->idx_begin() == 1) 1172 return getValueFromOverflowCondition(Val, WO, isTrueDest); 1173 1174 // Handle conditions in the form of (cond1 && cond2), we know that on the 1175 // true dest path both of the conditions hold. Similarly for conditions of 1176 // the form (cond1 || cond2), we know that on the false dest path neither 1177 // condition holds. 1178 BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond); 1179 if (!BO || (isTrueDest && BO->getOpcode() != BinaryOperator::And) || 1180 (!isTrueDest && BO->getOpcode() != BinaryOperator::Or)) 1181 return ValueLatticeElement::getOverdefined(); 1182 1183 // Prevent infinite recursion if Cond references itself as in this example: 1184 // Cond: "%tmp4 = and i1 %tmp4, undef" 1185 // BL: "%tmp4 = and i1 %tmp4, undef" 1186 // BR: "i1 undef" 1187 Value *BL = BO->getOperand(0); 1188 Value *BR = BO->getOperand(1); 1189 if (BL == Cond || BR == Cond) 1190 return ValueLatticeElement::getOverdefined(); 1191 1192 return intersect(getValueFromCondition(Val, BL, isTrueDest, Visited), 1193 getValueFromCondition(Val, BR, isTrueDest, Visited)); 1194 } 1195 1196 static ValueLatticeElement 1197 getValueFromCondition(Value *Val, Value *Cond, bool isTrueDest, 1198 DenseMap<Value*, ValueLatticeElement> &Visited) { 1199 auto I = Visited.find(Cond); 1200 if (I != Visited.end()) 1201 return I->second; 1202 1203 auto Result = getValueFromConditionImpl(Val, Cond, isTrueDest, Visited); 1204 Visited[Cond] = Result; 1205 return Result; 1206 } 1207 1208 ValueLatticeElement getValueFromCondition(Value *Val, Value *Cond, 1209 bool isTrueDest) { 1210 assert(Cond && "precondition"); 1211 DenseMap<Value*, ValueLatticeElement> Visited; 1212 return getValueFromCondition(Val, Cond, isTrueDest, Visited); 1213 } 1214 1215 // Return true if Usr has Op as an operand, otherwise false. 1216 static bool usesOperand(User *Usr, Value *Op) { 1217 return find(Usr->operands(), Op) != Usr->op_end(); 1218 } 1219 1220 // Return true if the instruction type of Val is supported by 1221 // constantFoldUser(). Currently CastInst and BinaryOperator only. Call this 1222 // before calling constantFoldUser() to find out if it's even worth attempting 1223 // to call it. 1224 static bool isOperationFoldable(User *Usr) { 1225 return isa<CastInst>(Usr) || isa<BinaryOperator>(Usr); 1226 } 1227 1228 // Check if Usr can be simplified to an integer constant when the value of one 1229 // of its operands Op is an integer constant OpConstVal. If so, return it as an 1230 // lattice value range with a single element or otherwise return an overdefined 1231 // lattice value. 1232 static ValueLatticeElement constantFoldUser(User *Usr, Value *Op, 1233 const APInt &OpConstVal, 1234 const DataLayout &DL) { 1235 assert(isOperationFoldable(Usr) && "Precondition"); 1236 Constant* OpConst = Constant::getIntegerValue(Op->getType(), OpConstVal); 1237 // Check if Usr can be simplified to a constant. 1238 if (auto *CI = dyn_cast<CastInst>(Usr)) { 1239 assert(CI->getOperand(0) == Op && "Operand 0 isn't Op"); 1240 if (auto *C = dyn_cast_or_null<ConstantInt>( 1241 SimplifyCastInst(CI->getOpcode(), OpConst, 1242 CI->getDestTy(), DL))) { 1243 return ValueLatticeElement::getRange(ConstantRange(C->getValue())); 1244 } 1245 } else if (auto *BO = dyn_cast<BinaryOperator>(Usr)) { 1246 bool Op0Match = BO->getOperand(0) == Op; 1247 bool Op1Match = BO->getOperand(1) == Op; 1248 assert((Op0Match || Op1Match) && 1249 "Operand 0 nor Operand 1 isn't a match"); 1250 Value *LHS = Op0Match ? OpConst : BO->getOperand(0); 1251 Value *RHS = Op1Match ? OpConst : BO->getOperand(1); 1252 if (auto *C = dyn_cast_or_null<ConstantInt>( 1253 SimplifyBinOp(BO->getOpcode(), LHS, RHS, DL))) { 1254 return ValueLatticeElement::getRange(ConstantRange(C->getValue())); 1255 } 1256 } 1257 return ValueLatticeElement::getOverdefined(); 1258 } 1259 1260 /// Compute the value of Val on the edge BBFrom -> BBTo. Returns false if 1261 /// Val is not constrained on the edge. Result is unspecified if return value 1262 /// is false. 1263 static bool getEdgeValueLocal(Value *Val, BasicBlock *BBFrom, 1264 BasicBlock *BBTo, ValueLatticeElement &Result) { 1265 // TODO: Handle more complex conditionals. If (v == 0 || v2 < 1) is false, we 1266 // know that v != 0. 1267 if (BranchInst *BI = dyn_cast<BranchInst>(BBFrom->getTerminator())) { 1268 // If this is a conditional branch and only one successor goes to BBTo, then 1269 // we may be able to infer something from the condition. 1270 if (BI->isConditional() && 1271 BI->getSuccessor(0) != BI->getSuccessor(1)) { 1272 bool isTrueDest = BI->getSuccessor(0) == BBTo; 1273 assert(BI->getSuccessor(!isTrueDest) == BBTo && 1274 "BBTo isn't a successor of BBFrom"); 1275 Value *Condition = BI->getCondition(); 1276 1277 // If V is the condition of the branch itself, then we know exactly what 1278 // it is. 1279 if (Condition == Val) { 1280 Result = ValueLatticeElement::get(ConstantInt::get( 1281 Type::getInt1Ty(Val->getContext()), isTrueDest)); 1282 return true; 1283 } 1284 1285 // If the condition of the branch is an equality comparison, we may be 1286 // able to infer the value. 1287 Result = getValueFromCondition(Val, Condition, isTrueDest); 1288 if (!Result.isOverdefined()) 1289 return true; 1290 1291 if (User *Usr = dyn_cast<User>(Val)) { 1292 assert(Result.isOverdefined() && "Result isn't overdefined"); 1293 // Check with isOperationFoldable() first to avoid linearly iterating 1294 // over the operands unnecessarily which can be expensive for 1295 // instructions with many operands. 1296 if (isa<IntegerType>(Usr->getType()) && isOperationFoldable(Usr)) { 1297 const DataLayout &DL = BBTo->getModule()->getDataLayout(); 1298 if (usesOperand(Usr, Condition)) { 1299 // If Val has Condition as an operand and Val can be folded into a 1300 // constant with either Condition == true or Condition == false, 1301 // propagate the constant. 1302 // eg. 1303 // ; %Val is true on the edge to %then. 1304 // %Val = and i1 %Condition, true. 1305 // br %Condition, label %then, label %else 1306 APInt ConditionVal(1, isTrueDest ? 1 : 0); 1307 Result = constantFoldUser(Usr, Condition, ConditionVal, DL); 1308 } else { 1309 // If one of Val's operand has an inferred value, we may be able to 1310 // infer the value of Val. 1311 // eg. 1312 // ; %Val is 94 on the edge to %then. 1313 // %Val = add i8 %Op, 1 1314 // %Condition = icmp eq i8 %Op, 93 1315 // br i1 %Condition, label %then, label %else 1316 for (unsigned i = 0; i < Usr->getNumOperands(); ++i) { 1317 Value *Op = Usr->getOperand(i); 1318 ValueLatticeElement OpLatticeVal = 1319 getValueFromCondition(Op, Condition, isTrueDest); 1320 if (Optional<APInt> OpConst = OpLatticeVal.asConstantInteger()) { 1321 Result = constantFoldUser(Usr, Op, OpConst.getValue(), DL); 1322 break; 1323 } 1324 } 1325 } 1326 } 1327 } 1328 if (!Result.isOverdefined()) 1329 return true; 1330 } 1331 } 1332 1333 // If the edge was formed by a switch on the value, then we may know exactly 1334 // what it is. 1335 if (SwitchInst *SI = dyn_cast<SwitchInst>(BBFrom->getTerminator())) { 1336 Value *Condition = SI->getCondition(); 1337 if (!isa<IntegerType>(Val->getType())) 1338 return false; 1339 bool ValUsesConditionAndMayBeFoldable = false; 1340 if (Condition != Val) { 1341 // Check if Val has Condition as an operand. 1342 if (User *Usr = dyn_cast<User>(Val)) 1343 ValUsesConditionAndMayBeFoldable = isOperationFoldable(Usr) && 1344 usesOperand(Usr, Condition); 1345 if (!ValUsesConditionAndMayBeFoldable) 1346 return false; 1347 } 1348 assert((Condition == Val || ValUsesConditionAndMayBeFoldable) && 1349 "Condition != Val nor Val doesn't use Condition"); 1350 1351 bool DefaultCase = SI->getDefaultDest() == BBTo; 1352 unsigned BitWidth = Val->getType()->getIntegerBitWidth(); 1353 ConstantRange EdgesVals(BitWidth, DefaultCase/*isFullSet*/); 1354 1355 for (auto Case : SI->cases()) { 1356 APInt CaseValue = Case.getCaseValue()->getValue(); 1357 ConstantRange EdgeVal(CaseValue); 1358 if (ValUsesConditionAndMayBeFoldable) { 1359 User *Usr = cast<User>(Val); 1360 const DataLayout &DL = BBTo->getModule()->getDataLayout(); 1361 ValueLatticeElement EdgeLatticeVal = 1362 constantFoldUser(Usr, Condition, CaseValue, DL); 1363 if (EdgeLatticeVal.isOverdefined()) 1364 return false; 1365 EdgeVal = EdgeLatticeVal.getConstantRange(); 1366 } 1367 if (DefaultCase) { 1368 // It is possible that the default destination is the destination of 1369 // some cases. We cannot perform difference for those cases. 1370 // We know Condition != CaseValue in BBTo. In some cases we can use 1371 // this to infer Val == f(Condition) is != f(CaseValue). For now, we 1372 // only do this when f is identity (i.e. Val == Condition), but we 1373 // should be able to do this for any injective f. 1374 if (Case.getCaseSuccessor() != BBTo && Condition == Val) 1375 EdgesVals = EdgesVals.difference(EdgeVal); 1376 } else if (Case.getCaseSuccessor() == BBTo) 1377 EdgesVals = EdgesVals.unionWith(EdgeVal); 1378 } 1379 Result = ValueLatticeElement::getRange(std::move(EdgesVals)); 1380 return true; 1381 } 1382 return false; 1383 } 1384 1385 /// Compute the value of Val on the edge BBFrom -> BBTo or the value at 1386 /// the basic block if the edge does not constrain Val. 1387 bool LazyValueInfoImpl::getEdgeValue(Value *Val, BasicBlock *BBFrom, 1388 BasicBlock *BBTo, 1389 ValueLatticeElement &Result, 1390 Instruction *CxtI) { 1391 // If already a constant, there is nothing to compute. 1392 if (Constant *VC = dyn_cast<Constant>(Val)) { 1393 Result = ValueLatticeElement::get(VC); 1394 return true; 1395 } 1396 1397 ValueLatticeElement LocalResult; 1398 if (!getEdgeValueLocal(Val, BBFrom, BBTo, LocalResult)) 1399 // If we couldn't constrain the value on the edge, LocalResult doesn't 1400 // provide any information. 1401 LocalResult = ValueLatticeElement::getOverdefined(); 1402 1403 if (hasSingleValue(LocalResult)) { 1404 // Can't get any more precise here 1405 Result = LocalResult; 1406 return true; 1407 } 1408 1409 if (!hasBlockValue(Val, BBFrom)) { 1410 if (pushBlockValue(std::make_pair(BBFrom, Val))) 1411 return false; 1412 // No new information. 1413 Result = LocalResult; 1414 return true; 1415 } 1416 1417 // Try to intersect ranges of the BB and the constraint on the edge. 1418 ValueLatticeElement InBlock = getBlockValue(Val, BBFrom); 1419 intersectAssumeOrGuardBlockValueConstantRange(Val, InBlock, 1420 BBFrom->getTerminator()); 1421 // We can use the context instruction (generically the ultimate instruction 1422 // the calling pass is trying to simplify) here, even though the result of 1423 // this function is generally cached when called from the solve* functions 1424 // (and that cached result might be used with queries using a different 1425 // context instruction), because when this function is called from the solve* 1426 // functions, the context instruction is not provided. When called from 1427 // LazyValueInfoImpl::getValueOnEdge, the context instruction is provided, 1428 // but then the result is not cached. 1429 intersectAssumeOrGuardBlockValueConstantRange(Val, InBlock, CxtI); 1430 1431 Result = intersect(LocalResult, InBlock); 1432 return true; 1433 } 1434 1435 ValueLatticeElement LazyValueInfoImpl::getValueInBlock(Value *V, BasicBlock *BB, 1436 Instruction *CxtI) { 1437 LLVM_DEBUG(dbgs() << "LVI Getting block end value " << *V << " at '" 1438 << BB->getName() << "'\n"); 1439 1440 assert(BlockValueStack.empty() && BlockValueSet.empty()); 1441 if (!hasBlockValue(V, BB)) { 1442 pushBlockValue(std::make_pair(BB, V)); 1443 solve(); 1444 } 1445 ValueLatticeElement Result = getBlockValue(V, BB); 1446 intersectAssumeOrGuardBlockValueConstantRange(V, Result, CxtI); 1447 1448 LLVM_DEBUG(dbgs() << " Result = " << Result << "\n"); 1449 return Result; 1450 } 1451 1452 ValueLatticeElement LazyValueInfoImpl::getValueAt(Value *V, Instruction *CxtI) { 1453 LLVM_DEBUG(dbgs() << "LVI Getting value " << *V << " at '" << CxtI->getName() 1454 << "'\n"); 1455 1456 if (auto *C = dyn_cast<Constant>(V)) 1457 return ValueLatticeElement::get(C); 1458 1459 ValueLatticeElement Result = ValueLatticeElement::getOverdefined(); 1460 if (auto *I = dyn_cast<Instruction>(V)) 1461 Result = getFromRangeMetadata(I); 1462 intersectAssumeOrGuardBlockValueConstantRange(V, Result, CxtI); 1463 1464 LLVM_DEBUG(dbgs() << " Result = " << Result << "\n"); 1465 return Result; 1466 } 1467 1468 ValueLatticeElement LazyValueInfoImpl:: 1469 getValueOnEdge(Value *V, BasicBlock *FromBB, BasicBlock *ToBB, 1470 Instruction *CxtI) { 1471 LLVM_DEBUG(dbgs() << "LVI Getting edge value " << *V << " from '" 1472 << FromBB->getName() << "' to '" << ToBB->getName() 1473 << "'\n"); 1474 1475 ValueLatticeElement Result; 1476 if (!getEdgeValue(V, FromBB, ToBB, Result, CxtI)) { 1477 solve(); 1478 bool WasFastQuery = getEdgeValue(V, FromBB, ToBB, Result, CxtI); 1479 (void)WasFastQuery; 1480 assert(WasFastQuery && "More work to do after problem solved?"); 1481 } 1482 1483 LLVM_DEBUG(dbgs() << " Result = " << Result << "\n"); 1484 return Result; 1485 } 1486 1487 void LazyValueInfoImpl::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc, 1488 BasicBlock *NewSucc) { 1489 TheCache.threadEdgeImpl(OldSucc, NewSucc); 1490 } 1491 1492 //===----------------------------------------------------------------------===// 1493 // LazyValueInfo Impl 1494 //===----------------------------------------------------------------------===// 1495 1496 /// This lazily constructs the LazyValueInfoImpl. 1497 static LazyValueInfoImpl &getImpl(void *&PImpl, AssumptionCache *AC, 1498 const DataLayout *DL, 1499 DominatorTree *DT = nullptr) { 1500 if (!PImpl) { 1501 assert(DL && "getCache() called with a null DataLayout"); 1502 PImpl = new LazyValueInfoImpl(AC, *DL, DT); 1503 } 1504 return *static_cast<LazyValueInfoImpl*>(PImpl); 1505 } 1506 1507 bool LazyValueInfoWrapperPass::runOnFunction(Function &F) { 1508 Info.AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 1509 const DataLayout &DL = F.getParent()->getDataLayout(); 1510 1511 DominatorTreeWrapperPass *DTWP = 1512 getAnalysisIfAvailable<DominatorTreeWrapperPass>(); 1513 Info.DT = DTWP ? &DTWP->getDomTree() : nullptr; 1514 Info.TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 1515 1516 if (Info.PImpl) 1517 getImpl(Info.PImpl, Info.AC, &DL, Info.DT).clear(); 1518 1519 // Fully lazy. 1520 return false; 1521 } 1522 1523 void LazyValueInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 1524 AU.setPreservesAll(); 1525 AU.addRequired<AssumptionCacheTracker>(); 1526 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1527 } 1528 1529 LazyValueInfo &LazyValueInfoWrapperPass::getLVI() { return Info; } 1530 1531 LazyValueInfo::~LazyValueInfo() { releaseMemory(); } 1532 1533 void LazyValueInfo::releaseMemory() { 1534 // If the cache was allocated, free it. 1535 if (PImpl) { 1536 delete &getImpl(PImpl, AC, nullptr); 1537 PImpl = nullptr; 1538 } 1539 } 1540 1541 bool LazyValueInfo::invalidate(Function &F, const PreservedAnalyses &PA, 1542 FunctionAnalysisManager::Invalidator &Inv) { 1543 // We need to invalidate if we have either failed to preserve this analyses 1544 // result directly or if any of its dependencies have been invalidated. 1545 auto PAC = PA.getChecker<LazyValueAnalysis>(); 1546 if (!(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>()) || 1547 (DT && Inv.invalidate<DominatorTreeAnalysis>(F, PA))) 1548 return true; 1549 1550 return false; 1551 } 1552 1553 void LazyValueInfoWrapperPass::releaseMemory() { Info.releaseMemory(); } 1554 1555 LazyValueInfo LazyValueAnalysis::run(Function &F, 1556 FunctionAnalysisManager &FAM) { 1557 auto &AC = FAM.getResult<AssumptionAnalysis>(F); 1558 auto &TLI = FAM.getResult<TargetLibraryAnalysis>(F); 1559 auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(F); 1560 1561 return LazyValueInfo(&AC, &F.getParent()->getDataLayout(), &TLI, DT); 1562 } 1563 1564 /// Returns true if we can statically tell that this value will never be a 1565 /// "useful" constant. In practice, this means we've got something like an 1566 /// alloca or a malloc call for which a comparison against a constant can 1567 /// only be guarding dead code. Note that we are potentially giving up some 1568 /// precision in dead code (a constant result) in favour of avoiding a 1569 /// expensive search for a easily answered common query. 1570 static bool isKnownNonConstant(Value *V) { 1571 V = V->stripPointerCasts(); 1572 // The return val of alloc cannot be a Constant. 1573 if (isa<AllocaInst>(V)) 1574 return true; 1575 return false; 1576 } 1577 1578 Constant *LazyValueInfo::getConstant(Value *V, BasicBlock *BB, 1579 Instruction *CxtI) { 1580 // Bail out early if V is known not to be a Constant. 1581 if (isKnownNonConstant(V)) 1582 return nullptr; 1583 1584 const DataLayout &DL = BB->getModule()->getDataLayout(); 1585 ValueLatticeElement Result = 1586 getImpl(PImpl, AC, &DL, DT).getValueInBlock(V, BB, CxtI); 1587 1588 if (Result.isConstant()) 1589 return Result.getConstant(); 1590 if (Result.isConstantRange()) { 1591 const ConstantRange &CR = Result.getConstantRange(); 1592 if (const APInt *SingleVal = CR.getSingleElement()) 1593 return ConstantInt::get(V->getContext(), *SingleVal); 1594 } 1595 return nullptr; 1596 } 1597 1598 ConstantRange LazyValueInfo::getConstantRange(Value *V, BasicBlock *BB, 1599 Instruction *CxtI) { 1600 assert(V->getType()->isIntegerTy()); 1601 unsigned Width = V->getType()->getIntegerBitWidth(); 1602 const DataLayout &DL = BB->getModule()->getDataLayout(); 1603 ValueLatticeElement Result = 1604 getImpl(PImpl, AC, &DL, DT).getValueInBlock(V, BB, CxtI); 1605 if (Result.isUndefined()) 1606 return ConstantRange::getEmpty(Width); 1607 if (Result.isConstantRange()) 1608 return Result.getConstantRange(); 1609 // We represent ConstantInt constants as constant ranges but other kinds 1610 // of integer constants, i.e. ConstantExpr will be tagged as constants 1611 assert(!(Result.isConstant() && isa<ConstantInt>(Result.getConstant())) && 1612 "ConstantInt value must be represented as constantrange"); 1613 return ConstantRange::getFull(Width); 1614 } 1615 1616 /// Determine whether the specified value is known to be a 1617 /// constant on the specified edge. Return null if not. 1618 Constant *LazyValueInfo::getConstantOnEdge(Value *V, BasicBlock *FromBB, 1619 BasicBlock *ToBB, 1620 Instruction *CxtI) { 1621 const DataLayout &DL = FromBB->getModule()->getDataLayout(); 1622 ValueLatticeElement Result = 1623 getImpl(PImpl, AC, &DL, DT).getValueOnEdge(V, FromBB, ToBB, CxtI); 1624 1625 if (Result.isConstant()) 1626 return Result.getConstant(); 1627 if (Result.isConstantRange()) { 1628 const ConstantRange &CR = Result.getConstantRange(); 1629 if (const APInt *SingleVal = CR.getSingleElement()) 1630 return ConstantInt::get(V->getContext(), *SingleVal); 1631 } 1632 return nullptr; 1633 } 1634 1635 ConstantRange LazyValueInfo::getConstantRangeOnEdge(Value *V, 1636 BasicBlock *FromBB, 1637 BasicBlock *ToBB, 1638 Instruction *CxtI) { 1639 unsigned Width = V->getType()->getIntegerBitWidth(); 1640 const DataLayout &DL = FromBB->getModule()->getDataLayout(); 1641 ValueLatticeElement Result = 1642 getImpl(PImpl, AC, &DL, DT).getValueOnEdge(V, FromBB, ToBB, CxtI); 1643 1644 if (Result.isUndefined()) 1645 return ConstantRange::getEmpty(Width); 1646 if (Result.isConstantRange()) 1647 return Result.getConstantRange(); 1648 // We represent ConstantInt constants as constant ranges but other kinds 1649 // of integer constants, i.e. ConstantExpr will be tagged as constants 1650 assert(!(Result.isConstant() && isa<ConstantInt>(Result.getConstant())) && 1651 "ConstantInt value must be represented as constantrange"); 1652 return ConstantRange::getFull(Width); 1653 } 1654 1655 static LazyValueInfo::Tristate 1656 getPredicateResult(unsigned Pred, Constant *C, const ValueLatticeElement &Val, 1657 const DataLayout &DL, TargetLibraryInfo *TLI) { 1658 // If we know the value is a constant, evaluate the conditional. 1659 Constant *Res = nullptr; 1660 if (Val.isConstant()) { 1661 Res = ConstantFoldCompareInstOperands(Pred, Val.getConstant(), C, DL, TLI); 1662 if (ConstantInt *ResCI = dyn_cast<ConstantInt>(Res)) 1663 return ResCI->isZero() ? LazyValueInfo::False : LazyValueInfo::True; 1664 return LazyValueInfo::Unknown; 1665 } 1666 1667 if (Val.isConstantRange()) { 1668 ConstantInt *CI = dyn_cast<ConstantInt>(C); 1669 if (!CI) return LazyValueInfo::Unknown; 1670 1671 const ConstantRange &CR = Val.getConstantRange(); 1672 if (Pred == ICmpInst::ICMP_EQ) { 1673 if (!CR.contains(CI->getValue())) 1674 return LazyValueInfo::False; 1675 1676 if (CR.isSingleElement()) 1677 return LazyValueInfo::True; 1678 } else if (Pred == ICmpInst::ICMP_NE) { 1679 if (!CR.contains(CI->getValue())) 1680 return LazyValueInfo::True; 1681 1682 if (CR.isSingleElement()) 1683 return LazyValueInfo::False; 1684 } else { 1685 // Handle more complex predicates. 1686 ConstantRange TrueValues = ConstantRange::makeExactICmpRegion( 1687 (ICmpInst::Predicate)Pred, CI->getValue()); 1688 if (TrueValues.contains(CR)) 1689 return LazyValueInfo::True; 1690 if (TrueValues.inverse().contains(CR)) 1691 return LazyValueInfo::False; 1692 } 1693 return LazyValueInfo::Unknown; 1694 } 1695 1696 if (Val.isNotConstant()) { 1697 // If this is an equality comparison, we can try to fold it knowing that 1698 // "V != C1". 1699 if (Pred == ICmpInst::ICMP_EQ) { 1700 // !C1 == C -> false iff C1 == C. 1701 Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE, 1702 Val.getNotConstant(), C, DL, 1703 TLI); 1704 if (Res->isNullValue()) 1705 return LazyValueInfo::False; 1706 } else if (Pred == ICmpInst::ICMP_NE) { 1707 // !C1 != C -> true iff C1 == C. 1708 Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE, 1709 Val.getNotConstant(), C, DL, 1710 TLI); 1711 if (Res->isNullValue()) 1712 return LazyValueInfo::True; 1713 } 1714 return LazyValueInfo::Unknown; 1715 } 1716 1717 return LazyValueInfo::Unknown; 1718 } 1719 1720 /// Determine whether the specified value comparison with a constant is known to 1721 /// be true or false on the specified CFG edge. Pred is a CmpInst predicate. 1722 LazyValueInfo::Tristate 1723 LazyValueInfo::getPredicateOnEdge(unsigned Pred, Value *V, Constant *C, 1724 BasicBlock *FromBB, BasicBlock *ToBB, 1725 Instruction *CxtI) { 1726 const DataLayout &DL = FromBB->getModule()->getDataLayout(); 1727 ValueLatticeElement Result = 1728 getImpl(PImpl, AC, &DL, DT).getValueOnEdge(V, FromBB, ToBB, CxtI); 1729 1730 return getPredicateResult(Pred, C, Result, DL, TLI); 1731 } 1732 1733 LazyValueInfo::Tristate 1734 LazyValueInfo::getPredicateAt(unsigned Pred, Value *V, Constant *C, 1735 Instruction *CxtI) { 1736 // Is or is not NonNull are common predicates being queried. If 1737 // isKnownNonZero can tell us the result of the predicate, we can 1738 // return it quickly. But this is only a fastpath, and falling 1739 // through would still be correct. 1740 const DataLayout &DL = CxtI->getModule()->getDataLayout(); 1741 if (V->getType()->isPointerTy() && C->isNullValue() && 1742 isKnownNonZero(V->stripPointerCasts(), DL)) { 1743 if (Pred == ICmpInst::ICMP_EQ) 1744 return LazyValueInfo::False; 1745 else if (Pred == ICmpInst::ICMP_NE) 1746 return LazyValueInfo::True; 1747 } 1748 ValueLatticeElement Result = getImpl(PImpl, AC, &DL, DT).getValueAt(V, CxtI); 1749 Tristate Ret = getPredicateResult(Pred, C, Result, DL, TLI); 1750 if (Ret != Unknown) 1751 return Ret; 1752 1753 // Note: The following bit of code is somewhat distinct from the rest of LVI; 1754 // LVI as a whole tries to compute a lattice value which is conservatively 1755 // correct at a given location. In this case, we have a predicate which we 1756 // weren't able to prove about the merged result, and we're pushing that 1757 // predicate back along each incoming edge to see if we can prove it 1758 // separately for each input. As a motivating example, consider: 1759 // bb1: 1760 // %v1 = ... ; constantrange<1, 5> 1761 // br label %merge 1762 // bb2: 1763 // %v2 = ... ; constantrange<10, 20> 1764 // br label %merge 1765 // merge: 1766 // %phi = phi [%v1, %v2] ; constantrange<1,20> 1767 // %pred = icmp eq i32 %phi, 8 1768 // We can't tell from the lattice value for '%phi' that '%pred' is false 1769 // along each path, but by checking the predicate over each input separately, 1770 // we can. 1771 // We limit the search to one step backwards from the current BB and value. 1772 // We could consider extending this to search further backwards through the 1773 // CFG and/or value graph, but there are non-obvious compile time vs quality 1774 // tradeoffs. 1775 if (CxtI) { 1776 BasicBlock *BB = CxtI->getParent(); 1777 1778 // Function entry or an unreachable block. Bail to avoid confusing 1779 // analysis below. 1780 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 1781 if (PI == PE) 1782 return Unknown; 1783 1784 // If V is a PHI node in the same block as the context, we need to ask 1785 // questions about the predicate as applied to the incoming value along 1786 // each edge. This is useful for eliminating cases where the predicate is 1787 // known along all incoming edges. 1788 if (auto *PHI = dyn_cast<PHINode>(V)) 1789 if (PHI->getParent() == BB) { 1790 Tristate Baseline = Unknown; 1791 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i < e; i++) { 1792 Value *Incoming = PHI->getIncomingValue(i); 1793 BasicBlock *PredBB = PHI->getIncomingBlock(i); 1794 // Note that PredBB may be BB itself. 1795 Tristate Result = getPredicateOnEdge(Pred, Incoming, C, PredBB, BB, 1796 CxtI); 1797 1798 // Keep going as long as we've seen a consistent known result for 1799 // all inputs. 1800 Baseline = (i == 0) ? Result /* First iteration */ 1801 : (Baseline == Result ? Baseline : Unknown); /* All others */ 1802 if (Baseline == Unknown) 1803 break; 1804 } 1805 if (Baseline != Unknown) 1806 return Baseline; 1807 } 1808 1809 // For a comparison where the V is outside this block, it's possible 1810 // that we've branched on it before. Look to see if the value is known 1811 // on all incoming edges. 1812 if (!isa<Instruction>(V) || 1813 cast<Instruction>(V)->getParent() != BB) { 1814 // For predecessor edge, determine if the comparison is true or false 1815 // on that edge. If they're all true or all false, we can conclude 1816 // the value of the comparison in this block. 1817 Tristate Baseline = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI); 1818 if (Baseline != Unknown) { 1819 // Check that all remaining incoming values match the first one. 1820 while (++PI != PE) { 1821 Tristate Ret = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI); 1822 if (Ret != Baseline) break; 1823 } 1824 // If we terminated early, then one of the values didn't match. 1825 if (PI == PE) { 1826 return Baseline; 1827 } 1828 } 1829 } 1830 } 1831 return Unknown; 1832 } 1833 1834 void LazyValueInfo::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc, 1835 BasicBlock *NewSucc) { 1836 if (PImpl) { 1837 const DataLayout &DL = PredBB->getModule()->getDataLayout(); 1838 getImpl(PImpl, AC, &DL, DT).threadEdge(PredBB, OldSucc, NewSucc); 1839 } 1840 } 1841 1842 void LazyValueInfo::eraseBlock(BasicBlock *BB) { 1843 if (PImpl) { 1844 const DataLayout &DL = BB->getModule()->getDataLayout(); 1845 getImpl(PImpl, AC, &DL, DT).eraseBlock(BB); 1846 } 1847 } 1848 1849 1850 void LazyValueInfo::printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) { 1851 if (PImpl) { 1852 getImpl(PImpl, AC, DL, DT).printLVI(F, DTree, OS); 1853 } 1854 } 1855 1856 void LazyValueInfo::disableDT() { 1857 if (PImpl) 1858 getImpl(PImpl, AC, DL, DT).disableDT(); 1859 } 1860 1861 void LazyValueInfo::enableDT() { 1862 if (PImpl) 1863 getImpl(PImpl, AC, DL, DT).enableDT(); 1864 } 1865 1866 // Print the LVI for the function arguments at the start of each basic block. 1867 void LazyValueInfoAnnotatedWriter::emitBasicBlockStartAnnot( 1868 const BasicBlock *BB, formatted_raw_ostream &OS) { 1869 // Find if there are latticevalues defined for arguments of the function. 1870 auto *F = BB->getParent(); 1871 for (auto &Arg : F->args()) { 1872 ValueLatticeElement Result = LVIImpl->getValueInBlock( 1873 const_cast<Argument *>(&Arg), const_cast<BasicBlock *>(BB)); 1874 if (Result.isUndefined()) 1875 continue; 1876 OS << "; LatticeVal for: '" << Arg << "' is: " << Result << "\n"; 1877 } 1878 } 1879 1880 // This function prints the LVI analysis for the instruction I at the beginning 1881 // of various basic blocks. It relies on calculated values that are stored in 1882 // the LazyValueInfoCache, and in the absence of cached values, recalculate the 1883 // LazyValueInfo for `I`, and print that info. 1884 void LazyValueInfoAnnotatedWriter::emitInstructionAnnot( 1885 const Instruction *I, formatted_raw_ostream &OS) { 1886 1887 auto *ParentBB = I->getParent(); 1888 SmallPtrSet<const BasicBlock*, 16> BlocksContainingLVI; 1889 // We can generate (solve) LVI values only for blocks that are dominated by 1890 // the I's parent. However, to avoid generating LVI for all dominating blocks, 1891 // that contain redundant/uninteresting information, we print LVI for 1892 // blocks that may use this LVI information (such as immediate successor 1893 // blocks, and blocks that contain uses of `I`). 1894 auto printResult = [&](const BasicBlock *BB) { 1895 if (!BlocksContainingLVI.insert(BB).second) 1896 return; 1897 ValueLatticeElement Result = LVIImpl->getValueInBlock( 1898 const_cast<Instruction *>(I), const_cast<BasicBlock *>(BB)); 1899 OS << "; LatticeVal for: '" << *I << "' in BB: '"; 1900 BB->printAsOperand(OS, false); 1901 OS << "' is: " << Result << "\n"; 1902 }; 1903 1904 printResult(ParentBB); 1905 // Print the LVI analysis results for the immediate successor blocks, that 1906 // are dominated by `ParentBB`. 1907 for (auto *BBSucc : successors(ParentBB)) 1908 if (DT.dominates(ParentBB, BBSucc)) 1909 printResult(BBSucc); 1910 1911 // Print LVI in blocks where `I` is used. 1912 for (auto *U : I->users()) 1913 if (auto *UseI = dyn_cast<Instruction>(U)) 1914 if (!isa<PHINode>(UseI) || DT.dominates(ParentBB, UseI->getParent())) 1915 printResult(UseI->getParent()); 1916 1917 } 1918 1919 namespace { 1920 // Printer class for LazyValueInfo results. 1921 class LazyValueInfoPrinter : public FunctionPass { 1922 public: 1923 static char ID; // Pass identification, replacement for typeid 1924 LazyValueInfoPrinter() : FunctionPass(ID) { 1925 initializeLazyValueInfoPrinterPass(*PassRegistry::getPassRegistry()); 1926 } 1927 1928 void getAnalysisUsage(AnalysisUsage &AU) const override { 1929 AU.setPreservesAll(); 1930 AU.addRequired<LazyValueInfoWrapperPass>(); 1931 AU.addRequired<DominatorTreeWrapperPass>(); 1932 } 1933 1934 // Get the mandatory dominator tree analysis and pass this in to the 1935 // LVIPrinter. We cannot rely on the LVI's DT, since it's optional. 1936 bool runOnFunction(Function &F) override { 1937 dbgs() << "LVI for function '" << F.getName() << "':\n"; 1938 auto &LVI = getAnalysis<LazyValueInfoWrapperPass>().getLVI(); 1939 auto &DTree = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1940 LVI.printLVI(F, DTree, dbgs()); 1941 return false; 1942 } 1943 }; 1944 } 1945 1946 char LazyValueInfoPrinter::ID = 0; 1947 INITIALIZE_PASS_BEGIN(LazyValueInfoPrinter, "print-lazy-value-info", 1948 "Lazy Value Info Printer Pass", false, false) 1949 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass) 1950 INITIALIZE_PASS_END(LazyValueInfoPrinter, "print-lazy-value-info", 1951 "Lazy Value Info Printer Pass", false, false) 1952