1 //===- LazyValueInfo.cpp - Value constraint analysis ------------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file defines the interface for lazy computation of value constraint 11 // information. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "llvm/Analysis/LazyValueInfo.h" 16 #include "llvm/ADT/DenseSet.h" 17 #include "llvm/ADT/Optional.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/Analysis/AssumptionCache.h" 20 #include "llvm/Analysis/ConstantFolding.h" 21 #include "llvm/Analysis/InstructionSimplify.h" 22 #include "llvm/Analysis/TargetLibraryInfo.h" 23 #include "llvm/Analysis/ValueTracking.h" 24 #include "llvm/Analysis/ValueLattice.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/Instructions.h" 32 #include "llvm/IR/IntrinsicInst.h" 33 #include "llvm/IR/Intrinsics.h" 34 #include "llvm/IR/LLVMContext.h" 35 #include "llvm/IR/PatternMatch.h" 36 #include "llvm/IR/ValueHandle.h" 37 #include "llvm/Support/Debug.h" 38 #include "llvm/Support/FormattedStream.h" 39 #include "llvm/Support/raw_ostream.h" 40 #include <map> 41 using namespace llvm; 42 using namespace PatternMatch; 43 44 #define DEBUG_TYPE "lazy-value-info" 45 46 // This is the number of worklist items we will process to try to discover an 47 // answer for a given value. 48 static const unsigned MaxProcessedPerValue = 500; 49 50 char LazyValueInfoWrapperPass::ID = 0; 51 INITIALIZE_PASS_BEGIN(LazyValueInfoWrapperPass, "lazy-value-info", 52 "Lazy Value Information Analysis", false, true) 53 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 54 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 55 INITIALIZE_PASS_END(LazyValueInfoWrapperPass, "lazy-value-info", 56 "Lazy Value Information Analysis", false, true) 57 58 namespace llvm { 59 FunctionPass *createLazyValueInfoPass() { return new LazyValueInfoWrapperPass(); } 60 } 61 62 AnalysisKey LazyValueAnalysis::Key; 63 64 /// Returns true if this lattice value represents at most one possible value. 65 /// This is as precise as any lattice value can get while still representing 66 /// reachable code. 67 static bool hasSingleValue(const ValueLatticeElement &Val) { 68 if (Val.isConstantRange() && 69 Val.getConstantRange().isSingleElement()) 70 // Integer constants are single element ranges 71 return true; 72 if (Val.isConstant()) 73 // Non integer constants 74 return true; 75 return false; 76 } 77 78 /// Combine two sets of facts about the same value into a single set of 79 /// facts. Note that this method is not suitable for merging facts along 80 /// different paths in a CFG; that's what the mergeIn function is for. This 81 /// is for merging facts gathered about the same value at the same location 82 /// through two independent means. 83 /// Notes: 84 /// * This method does not promise to return the most precise possible lattice 85 /// value implied by A and B. It is allowed to return any lattice element 86 /// which is at least as strong as *either* A or B (unless our facts 87 /// conflict, see below). 88 /// * Due to unreachable code, the intersection of two lattice values could be 89 /// contradictory. If this happens, we return some valid lattice value so as 90 /// not confuse the rest of LVI. Ideally, we'd always return Undefined, but 91 /// we do not make this guarantee. TODO: This would be a useful enhancement. 92 static ValueLatticeElement intersect(const ValueLatticeElement &A, 93 const ValueLatticeElement &B) { 94 // Undefined is the strongest state. It means the value is known to be along 95 // an unreachable path. 96 if (A.isUndefined()) 97 return A; 98 if (B.isUndefined()) 99 return B; 100 101 // If we gave up for one, but got a useable fact from the other, use it. 102 if (A.isOverdefined()) 103 return B; 104 if (B.isOverdefined()) 105 return A; 106 107 // Can't get any more precise than constants. 108 if (hasSingleValue(A)) 109 return A; 110 if (hasSingleValue(B)) 111 return B; 112 113 // Could be either constant range or not constant here. 114 if (!A.isConstantRange() || !B.isConstantRange()) { 115 // TODO: Arbitrary choice, could be improved 116 return A; 117 } 118 119 // Intersect two constant ranges 120 ConstantRange Range = 121 A.getConstantRange().intersectWith(B.getConstantRange()); 122 // Note: An empty range is implicitly converted to overdefined internally. 123 // TODO: We could instead use Undefined here since we've proven a conflict 124 // and thus know this path must be unreachable. 125 return ValueLatticeElement::getRange(std::move(Range)); 126 } 127 128 //===----------------------------------------------------------------------===// 129 // LazyValueInfoCache Decl 130 //===----------------------------------------------------------------------===// 131 132 namespace { 133 /// A callback value handle updates the cache when values are erased. 134 class LazyValueInfoCache; 135 struct LVIValueHandle final : public CallbackVH { 136 // Needs to access getValPtr(), which is protected. 137 friend struct DenseMapInfo<LVIValueHandle>; 138 139 LazyValueInfoCache *Parent; 140 141 LVIValueHandle(Value *V, LazyValueInfoCache *P) 142 : CallbackVH(V), Parent(P) { } 143 144 void deleted() override; 145 void allUsesReplacedWith(Value *V) override { 146 deleted(); 147 } 148 }; 149 } // end anonymous namespace 150 151 namespace { 152 /// This is the cache kept by LazyValueInfo which 153 /// maintains information about queries across the clients' queries. 154 class LazyValueInfoCache { 155 /// This is all of the cached block information for exactly one Value*. 156 /// The entries are sorted by the BasicBlock* of the 157 /// entries, allowing us to do a lookup with a binary search. 158 /// Over-defined lattice values are recorded in OverDefinedCache to reduce 159 /// memory overhead. 160 struct ValueCacheEntryTy { 161 ValueCacheEntryTy(Value *V, LazyValueInfoCache *P) : Handle(V, P) {} 162 LVIValueHandle Handle; 163 SmallDenseMap<PoisoningVH<BasicBlock>, ValueLatticeElement, 4> BlockVals; 164 }; 165 166 /// This tracks, on a per-block basis, the set of values that are 167 /// over-defined at the end of that block. 168 typedef DenseMap<PoisoningVH<BasicBlock>, SmallPtrSet<Value *, 4>> 169 OverDefinedCacheTy; 170 /// Keep track of all blocks that we have ever seen, so we 171 /// don't spend time removing unused blocks from our caches. 172 DenseSet<PoisoningVH<BasicBlock> > SeenBlocks; 173 174 /// This is all of the cached information for all values, 175 /// mapped from Value* to key information. 176 DenseMap<Value *, std::unique_ptr<ValueCacheEntryTy>> ValueCache; 177 OverDefinedCacheTy OverDefinedCache; 178 179 180 public: 181 void insertResult(Value *Val, BasicBlock *BB, 182 const ValueLatticeElement &Result) { 183 SeenBlocks.insert(BB); 184 185 // Insert over-defined values into their own cache to reduce memory 186 // overhead. 187 if (Result.isOverdefined()) 188 OverDefinedCache[BB].insert(Val); 189 else { 190 auto It = ValueCache.find_as(Val); 191 if (It == ValueCache.end()) { 192 ValueCache[Val] = make_unique<ValueCacheEntryTy>(Val, this); 193 It = ValueCache.find_as(Val); 194 assert(It != ValueCache.end() && "Val was just added to the map!"); 195 } 196 It->second->BlockVals[BB] = Result; 197 } 198 } 199 200 bool isOverdefined(Value *V, BasicBlock *BB) const { 201 auto ODI = OverDefinedCache.find(BB); 202 203 if (ODI == OverDefinedCache.end()) 204 return false; 205 206 return ODI->second.count(V); 207 } 208 209 bool hasCachedValueInfo(Value *V, BasicBlock *BB) const { 210 if (isOverdefined(V, BB)) 211 return true; 212 213 auto I = ValueCache.find_as(V); 214 if (I == ValueCache.end()) 215 return false; 216 217 return I->second->BlockVals.count(BB); 218 } 219 220 ValueLatticeElement getCachedValueInfo(Value *V, BasicBlock *BB) const { 221 if (isOverdefined(V, BB)) 222 return ValueLatticeElement::getOverdefined(); 223 224 auto I = ValueCache.find_as(V); 225 if (I == ValueCache.end()) 226 return ValueLatticeElement(); 227 auto BBI = I->second->BlockVals.find(BB); 228 if (BBI == I->second->BlockVals.end()) 229 return ValueLatticeElement(); 230 return BBI->second; 231 } 232 233 /// clear - Empty the cache. 234 void clear() { 235 SeenBlocks.clear(); 236 ValueCache.clear(); 237 OverDefinedCache.clear(); 238 } 239 240 /// Inform the cache that a given value has been deleted. 241 void eraseValue(Value *V); 242 243 /// This is part of the update interface to inform the cache 244 /// that a block has been deleted. 245 void eraseBlock(BasicBlock *BB); 246 247 /// Updates the cache to remove any influence an overdefined value in 248 /// OldSucc might have (unless also overdefined in NewSucc). This just 249 /// flushes elements from the cache and does not add any. 250 void threadEdgeImpl(BasicBlock *OldSucc,BasicBlock *NewSucc); 251 252 friend struct LVIValueHandle; 253 }; 254 } 255 256 void LazyValueInfoCache::eraseValue(Value *V) { 257 for (auto I = OverDefinedCache.begin(), E = OverDefinedCache.end(); I != E;) { 258 // Copy and increment the iterator immediately so we can erase behind 259 // ourselves. 260 auto Iter = I++; 261 SmallPtrSetImpl<Value *> &ValueSet = Iter->second; 262 ValueSet.erase(V); 263 if (ValueSet.empty()) 264 OverDefinedCache.erase(Iter); 265 } 266 267 ValueCache.erase(V); 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(*this); 274 } 275 276 void LazyValueInfoCache::eraseBlock(BasicBlock *BB) { 277 // Shortcut if we have never seen this block. 278 DenseSet<PoisoningVH<BasicBlock> >::iterator I = SeenBlocks.find(BB); 279 if (I == SeenBlocks.end()) 280 return; 281 SeenBlocks.erase(I); 282 283 auto ODI = OverDefinedCache.find(BB); 284 if (ODI != OverDefinedCache.end()) 285 OverDefinedCache.erase(ODI); 286 287 for (auto &I : ValueCache) 288 I.second->BlockVals.erase(BB); 289 } 290 291 void LazyValueInfoCache::threadEdgeImpl(BasicBlock *OldSucc, 292 BasicBlock *NewSucc) { 293 // When an edge in the graph has been threaded, values that we could not 294 // determine a value for before (i.e. were marked overdefined) may be 295 // possible to solve now. We do NOT try to proactively update these values. 296 // Instead, we clear their entries from the cache, and allow lazy updating to 297 // recompute them when needed. 298 299 // The updating process is fairly simple: we need to drop cached info 300 // for all values that were marked overdefined in OldSucc, and for those same 301 // values in any successor of OldSucc (except NewSucc) in which they were 302 // also marked overdefined. 303 std::vector<BasicBlock*> worklist; 304 worklist.push_back(OldSucc); 305 306 auto I = OverDefinedCache.find(OldSucc); 307 if (I == OverDefinedCache.end()) 308 return; // Nothing to process here. 309 SmallVector<Value *, 4> ValsToClear(I->second.begin(), I->second.end()); 310 311 // Use a worklist to perform a depth-first search of OldSucc's successors. 312 // NOTE: We do not need a visited list since any blocks we have already 313 // visited will have had their overdefined markers cleared already, and we 314 // thus won't loop to their successors. 315 while (!worklist.empty()) { 316 BasicBlock *ToUpdate = worklist.back(); 317 worklist.pop_back(); 318 319 // Skip blocks only accessible through NewSucc. 320 if (ToUpdate == NewSucc) continue; 321 322 // If a value was marked overdefined in OldSucc, and is here too... 323 auto OI = OverDefinedCache.find(ToUpdate); 324 if (OI == OverDefinedCache.end()) 325 continue; 326 SmallPtrSetImpl<Value *> &ValueSet = OI->second; 327 328 bool changed = false; 329 for (Value *V : ValsToClear) { 330 if (!ValueSet.erase(V)) 331 continue; 332 333 // If we removed anything, then we potentially need to update 334 // blocks successors too. 335 changed = true; 336 337 if (ValueSet.empty()) { 338 OverDefinedCache.erase(OI); 339 break; 340 } 341 } 342 343 if (!changed) continue; 344 345 worklist.insert(worklist.end(), succ_begin(ToUpdate), succ_end(ToUpdate)); 346 } 347 } 348 349 350 namespace { 351 /// An assembly annotator class to print LazyValueCache information in 352 /// comments. 353 class LazyValueInfoImpl; 354 class LazyValueInfoAnnotatedWriter : public AssemblyAnnotationWriter { 355 LazyValueInfoImpl *LVIImpl; 356 // While analyzing which blocks we can solve values for, we need the dominator 357 // information. Since this is an optional parameter in LVI, we require this 358 // DomTreeAnalysis pass in the printer pass, and pass the dominator 359 // tree to the LazyValueInfoAnnotatedWriter. 360 DominatorTree &DT; 361 362 public: 363 LazyValueInfoAnnotatedWriter(LazyValueInfoImpl *L, DominatorTree &DTree) 364 : LVIImpl(L), DT(DTree) {} 365 366 virtual void emitBasicBlockStartAnnot(const BasicBlock *BB, 367 formatted_raw_ostream &OS); 368 369 virtual void emitInstructionAnnot(const Instruction *I, 370 formatted_raw_ostream &OS); 371 }; 372 } 373 namespace { 374 // The actual implementation of the lazy analysis and update. Note that the 375 // inheritance from LazyValueInfoCache is intended to be temporary while 376 // splitting the code and then transitioning to a has-a relationship. 377 class LazyValueInfoImpl { 378 379 /// Cached results from previous queries 380 LazyValueInfoCache TheCache; 381 382 /// This stack holds the state of the value solver during a query. 383 /// It basically emulates the callstack of the naive 384 /// recursive value lookup process. 385 SmallVector<std::pair<BasicBlock*, Value*>, 8> BlockValueStack; 386 387 /// Keeps track of which block-value pairs are in BlockValueStack. 388 DenseSet<std::pair<BasicBlock*, Value*> > BlockValueSet; 389 390 /// Push BV onto BlockValueStack unless it's already in there. 391 /// Returns true on success. 392 bool pushBlockValue(const std::pair<BasicBlock *, Value *> &BV) { 393 if (!BlockValueSet.insert(BV).second) 394 return false; // It's already in the stack. 395 396 LLVM_DEBUG(dbgs() << "PUSH: " << *BV.second << " in " 397 << BV.first->getName() << "\n"); 398 BlockValueStack.push_back(BV); 399 return true; 400 } 401 402 AssumptionCache *AC; ///< A pointer to the cache of @llvm.assume calls. 403 const DataLayout &DL; ///< A mandatory DataLayout 404 DominatorTree *DT; ///< An optional DT pointer. 405 DominatorTree *DisabledDT; ///< Stores DT if it's disabled. 406 407 ValueLatticeElement getBlockValue(Value *Val, BasicBlock *BB); 408 bool getEdgeValue(Value *V, BasicBlock *F, BasicBlock *T, 409 ValueLatticeElement &Result, Instruction *CxtI = nullptr); 410 bool hasBlockValue(Value *Val, BasicBlock *BB); 411 412 // These methods process one work item and may add more. A false value 413 // returned means that the work item was not completely processed and must 414 // be revisited after going through the new items. 415 bool solveBlockValue(Value *Val, BasicBlock *BB); 416 bool solveBlockValueImpl(ValueLatticeElement &Res, Value *Val, 417 BasicBlock *BB); 418 bool solveBlockValueNonLocal(ValueLatticeElement &BBLV, Value *Val, 419 BasicBlock *BB); 420 bool solveBlockValuePHINode(ValueLatticeElement &BBLV, PHINode *PN, 421 BasicBlock *BB); 422 bool solveBlockValueSelect(ValueLatticeElement &BBLV, SelectInst *S, 423 BasicBlock *BB); 424 Optional<ConstantRange> getRangeForOperand(unsigned Op, Instruction *I, 425 BasicBlock *BB); 426 bool solveBlockValueBinaryOp(ValueLatticeElement &BBLV, BinaryOperator *BBI, 427 BasicBlock *BB); 428 bool solveBlockValueCast(ValueLatticeElement &BBLV, CastInst *CI, 429 BasicBlock *BB); 430 void intersectAssumeOrGuardBlockValueConstantRange(Value *Val, 431 ValueLatticeElement &BBLV, 432 Instruction *BBI); 433 434 void solve(); 435 436 public: 437 /// This is the query interface to determine the lattice 438 /// value for the specified Value* at the end of the specified block. 439 ValueLatticeElement getValueInBlock(Value *V, BasicBlock *BB, 440 Instruction *CxtI = nullptr); 441 442 /// This is the query interface to determine the lattice 443 /// value for the specified Value* at the specified instruction (generally 444 /// from an assume intrinsic). 445 ValueLatticeElement getValueAt(Value *V, Instruction *CxtI); 446 447 /// This is the query interface to determine the lattice 448 /// value for the specified Value* that is true on the specified edge. 449 ValueLatticeElement getValueOnEdge(Value *V, BasicBlock *FromBB, 450 BasicBlock *ToBB, 451 Instruction *CxtI = nullptr); 452 453 /// Complete flush all previously computed values 454 void clear() { 455 TheCache.clear(); 456 } 457 458 /// Printing the LazyValueInfo Analysis. 459 void printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) { 460 LazyValueInfoAnnotatedWriter Writer(this, DTree); 461 F.print(OS, &Writer); 462 } 463 464 /// This is part of the update interface to inform the cache 465 /// that a block has been deleted. 466 void eraseBlock(BasicBlock *BB) { 467 TheCache.eraseBlock(BB); 468 } 469 470 /// Disables use of the DominatorTree within LVI. 471 void disableDT() { 472 if (DT) { 473 assert(!DisabledDT && "Both DT and DisabledDT are not nullptr!"); 474 std::swap(DT, DisabledDT); 475 } 476 } 477 478 /// Enables use of the DominatorTree within LVI. Does nothing if the class 479 /// instance was initialized without a DT pointer. 480 void enableDT() { 481 if (DisabledDT) { 482 assert(!DT && "Both DT and DisabledDT are not nullptr!"); 483 std::swap(DT, DisabledDT); 484 } 485 } 486 487 /// This is the update interface to inform the cache that an edge from 488 /// PredBB to OldSucc has been threaded to be from PredBB to NewSucc. 489 void threadEdge(BasicBlock *PredBB,BasicBlock *OldSucc,BasicBlock *NewSucc); 490 491 LazyValueInfoImpl(AssumptionCache *AC, const DataLayout &DL, 492 DominatorTree *DT = nullptr) 493 : AC(AC), DL(DL), DT(DT), DisabledDT(nullptr) {} 494 }; 495 } // end anonymous namespace 496 497 498 void LazyValueInfoImpl::solve() { 499 SmallVector<std::pair<BasicBlock *, Value *>, 8> StartingStack( 500 BlockValueStack.begin(), BlockValueStack.end()); 501 502 unsigned processedCount = 0; 503 while (!BlockValueStack.empty()) { 504 processedCount++; 505 // Abort if we have to process too many values to get a result for this one. 506 // Because of the design of the overdefined cache currently being per-block 507 // to avoid naming-related issues (IE it wants to try to give different 508 // results for the same name in different blocks), overdefined results don't 509 // get cached globally, which in turn means we will often try to rediscover 510 // the same overdefined result again and again. Once something like 511 // PredicateInfo is used in LVI or CVP, we should be able to make the 512 // overdefined cache global, and remove this throttle. 513 if (processedCount > MaxProcessedPerValue) { 514 LLVM_DEBUG( 515 dbgs() << "Giving up on stack because we are getting too deep\n"); 516 // Fill in the original values 517 while (!StartingStack.empty()) { 518 std::pair<BasicBlock *, Value *> &e = StartingStack.back(); 519 TheCache.insertResult(e.second, e.first, 520 ValueLatticeElement::getOverdefined()); 521 StartingStack.pop_back(); 522 } 523 BlockValueSet.clear(); 524 BlockValueStack.clear(); 525 return; 526 } 527 std::pair<BasicBlock *, Value *> e = BlockValueStack.back(); 528 assert(BlockValueSet.count(e) && "Stack value should be in BlockValueSet!"); 529 530 if (solveBlockValue(e.second, e.first)) { 531 // The work item was completely processed. 532 assert(BlockValueStack.back() == e && "Nothing should have been pushed!"); 533 assert(TheCache.hasCachedValueInfo(e.second, e.first) && 534 "Result should be in cache!"); 535 536 LLVM_DEBUG( 537 dbgs() << "POP " << *e.second << " in " << e.first->getName() << " = " 538 << TheCache.getCachedValueInfo(e.second, e.first) << "\n"); 539 540 BlockValueStack.pop_back(); 541 BlockValueSet.erase(e); 542 } else { 543 // More work needs to be done before revisiting. 544 assert(BlockValueStack.back() != e && "Stack should have been pushed!"); 545 } 546 } 547 } 548 549 bool LazyValueInfoImpl::hasBlockValue(Value *Val, BasicBlock *BB) { 550 // If already a constant, there is nothing to compute. 551 if (isa<Constant>(Val)) 552 return true; 553 554 return TheCache.hasCachedValueInfo(Val, BB); 555 } 556 557 ValueLatticeElement LazyValueInfoImpl::getBlockValue(Value *Val, 558 BasicBlock *BB) { 559 // If already a constant, there is nothing to compute. 560 if (Constant *VC = dyn_cast<Constant>(Val)) 561 return ValueLatticeElement::get(VC); 562 563 return TheCache.getCachedValueInfo(Val, BB); 564 } 565 566 static ValueLatticeElement getFromRangeMetadata(Instruction *BBI) { 567 switch (BBI->getOpcode()) { 568 default: break; 569 case Instruction::Load: 570 case Instruction::Call: 571 case Instruction::Invoke: 572 if (MDNode *Ranges = BBI->getMetadata(LLVMContext::MD_range)) 573 if (isa<IntegerType>(BBI->getType())) { 574 return ValueLatticeElement::getRange( 575 getConstantRangeFromMetadata(*Ranges)); 576 } 577 break; 578 }; 579 // Nothing known - will be intersected with other facts 580 return ValueLatticeElement::getOverdefined(); 581 } 582 583 bool LazyValueInfoImpl::solveBlockValue(Value *Val, BasicBlock *BB) { 584 if (isa<Constant>(Val)) 585 return true; 586 587 if (TheCache.hasCachedValueInfo(Val, BB)) { 588 // If we have a cached value, use that. 589 LLVM_DEBUG(dbgs() << " reuse BB '" << BB->getName() << "' val=" 590 << TheCache.getCachedValueInfo(Val, BB) << '\n'); 591 592 // Since we're reusing a cached value, we don't need to update the 593 // OverDefinedCache. The cache will have been properly updated whenever the 594 // cached value was inserted. 595 return true; 596 } 597 598 // Hold off inserting this value into the Cache in case we have to return 599 // false and come back later. 600 ValueLatticeElement Res; 601 if (!solveBlockValueImpl(Res, Val, BB)) 602 // Work pushed, will revisit 603 return false; 604 605 TheCache.insertResult(Val, BB, Res); 606 return true; 607 } 608 609 bool LazyValueInfoImpl::solveBlockValueImpl(ValueLatticeElement &Res, 610 Value *Val, BasicBlock *BB) { 611 612 Instruction *BBI = dyn_cast<Instruction>(Val); 613 if (!BBI || BBI->getParent() != BB) 614 return solveBlockValueNonLocal(Res, Val, BB); 615 616 if (PHINode *PN = dyn_cast<PHINode>(BBI)) 617 return solveBlockValuePHINode(Res, PN, BB); 618 619 if (auto *SI = dyn_cast<SelectInst>(BBI)) 620 return solveBlockValueSelect(Res, SI, BB); 621 622 // If this value is a nonnull pointer, record it's range and bailout. Note 623 // that for all other pointer typed values, we terminate the search at the 624 // definition. We could easily extend this to look through geps, bitcasts, 625 // and the like to prove non-nullness, but it's not clear that's worth it 626 // compile time wise. The context-insensitive value walk done inside 627 // isKnownNonZero gets most of the profitable cases at much less expense. 628 // This does mean that we have a sensativity to where the defining 629 // instruction is placed, even if it could legally be hoisted much higher. 630 // That is unfortunate. 631 PointerType *PT = dyn_cast<PointerType>(BBI->getType()); 632 if (PT && isKnownNonZero(BBI, DL)) { 633 Res = ValueLatticeElement::getNot(ConstantPointerNull::get(PT)); 634 return true; 635 } 636 if (BBI->getType()->isIntegerTy()) { 637 if (auto *CI = dyn_cast<CastInst>(BBI)) 638 return solveBlockValueCast(Res, CI, BB); 639 640 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(BBI)) 641 return solveBlockValueBinaryOp(Res, BO, BB); 642 } 643 644 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() 645 << "' - unknown inst def found.\n"); 646 Res = getFromRangeMetadata(BBI); 647 return true; 648 } 649 650 static bool InstructionDereferencesPointer(Instruction *I, Value *Ptr) { 651 if (LoadInst *L = dyn_cast<LoadInst>(I)) { 652 return L->getPointerAddressSpace() == 0 && 653 GetUnderlyingObject(L->getPointerOperand(), 654 L->getModule()->getDataLayout()) == Ptr; 655 } 656 if (StoreInst *S = dyn_cast<StoreInst>(I)) { 657 return S->getPointerAddressSpace() == 0 && 658 GetUnderlyingObject(S->getPointerOperand(), 659 S->getModule()->getDataLayout()) == Ptr; 660 } 661 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I)) { 662 if (MI->isVolatile()) return false; 663 664 // FIXME: check whether it has a valuerange that excludes zero? 665 ConstantInt *Len = dyn_cast<ConstantInt>(MI->getLength()); 666 if (!Len || Len->isZero()) return false; 667 668 if (MI->getDestAddressSpace() == 0) 669 if (GetUnderlyingObject(MI->getRawDest(), 670 MI->getModule()->getDataLayout()) == Ptr) 671 return true; 672 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) 673 if (MTI->getSourceAddressSpace() == 0) 674 if (GetUnderlyingObject(MTI->getRawSource(), 675 MTI->getModule()->getDataLayout()) == Ptr) 676 return true; 677 } 678 return false; 679 } 680 681 /// Return true if the allocation associated with Val is ever dereferenced 682 /// within the given basic block. This establishes the fact Val is not null, 683 /// but does not imply that the memory at Val is dereferenceable. (Val may 684 /// point off the end of the dereferenceable part of the object.) 685 static bool isObjectDereferencedInBlock(Value *Val, BasicBlock *BB) { 686 assert(Val->getType()->isPointerTy()); 687 688 const DataLayout &DL = BB->getModule()->getDataLayout(); 689 Value *UnderlyingVal = GetUnderlyingObject(Val, DL); 690 // If 'GetUnderlyingObject' didn't converge, skip it. It won't converge 691 // inside InstructionDereferencesPointer either. 692 if (UnderlyingVal == GetUnderlyingObject(UnderlyingVal, DL, 1)) 693 for (Instruction &I : *BB) 694 if (InstructionDereferencesPointer(&I, UnderlyingVal)) 695 return true; 696 return false; 697 } 698 699 bool LazyValueInfoImpl::solveBlockValueNonLocal(ValueLatticeElement &BBLV, 700 Value *Val, BasicBlock *BB) { 701 ValueLatticeElement Result; // Start Undefined. 702 703 // If this is the entry block, we must be asking about an argument. The 704 // value is overdefined. 705 if (BB == &BB->getParent()->getEntryBlock()) { 706 assert(isa<Argument>(Val) && "Unknown live-in to the entry block"); 707 // Before giving up, see if we can prove the pointer non-null local to 708 // this particular block. 709 PointerType *PTy = dyn_cast<PointerType>(Val->getType()); 710 if (PTy && 711 (isKnownNonZero(Val, DL) || 712 (isObjectDereferencedInBlock(Val, BB) && 713 !NullPointerIsDefined(BB->getParent(), PTy->getAddressSpace())))) { 714 Result = ValueLatticeElement::getNot(ConstantPointerNull::get(PTy)); 715 } else { 716 Result = ValueLatticeElement::getOverdefined(); 717 } 718 BBLV = Result; 719 return true; 720 } 721 722 // Loop over all of our predecessors, merging what we know from them into 723 // result. If we encounter an unexplored predecessor, we eagerly explore it 724 // in a depth first manner. In practice, this has the effect of discovering 725 // paths we can't analyze eagerly without spending compile times analyzing 726 // other paths. This heuristic benefits from the fact that predecessors are 727 // frequently arranged such that dominating ones come first and we quickly 728 // find a path to function entry. TODO: We should consider explicitly 729 // canonicalizing to make this true rather than relying on this happy 730 // accident. 731 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 732 ValueLatticeElement EdgeResult; 733 if (!getEdgeValue(Val, *PI, BB, EdgeResult)) 734 // Explore that input, then return here 735 return false; 736 737 Result.mergeIn(EdgeResult, DL); 738 739 // If we hit overdefined, exit early. The BlockVals entry is already set 740 // to overdefined. 741 if (Result.isOverdefined()) { 742 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() 743 << "' - overdefined because of pred (non local).\n"); 744 // Before giving up, see if we can prove the pointer non-null local to 745 // this particular block. 746 PointerType *PTy = dyn_cast<PointerType>(Val->getType()); 747 if (PTy && isObjectDereferencedInBlock(Val, BB) && 748 !NullPointerIsDefined(BB->getParent(), PTy->getAddressSpace())) { 749 Result = ValueLatticeElement::getNot(ConstantPointerNull::get(PTy)); 750 } 751 752 BBLV = Result; 753 return true; 754 } 755 } 756 757 // Return the merged value, which is more precise than 'overdefined'. 758 assert(!Result.isOverdefined()); 759 BBLV = Result; 760 return true; 761 } 762 763 bool LazyValueInfoImpl::solveBlockValuePHINode(ValueLatticeElement &BBLV, 764 PHINode *PN, BasicBlock *BB) { 765 ValueLatticeElement Result; // Start Undefined. 766 767 // Loop over all of our predecessors, merging what we know from them into 768 // result. See the comment about the chosen traversal order in 769 // solveBlockValueNonLocal; the same reasoning applies here. 770 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 771 BasicBlock *PhiBB = PN->getIncomingBlock(i); 772 Value *PhiVal = PN->getIncomingValue(i); 773 ValueLatticeElement EdgeResult; 774 // Note that we can provide PN as the context value to getEdgeValue, even 775 // though the results will be cached, because PN is the value being used as 776 // the cache key in the caller. 777 if (!getEdgeValue(PhiVal, PhiBB, BB, EdgeResult, PN)) 778 // Explore that input, then return here 779 return false; 780 781 Result.mergeIn(EdgeResult, DL); 782 783 // If we hit overdefined, exit early. The BlockVals entry is already set 784 // to overdefined. 785 if (Result.isOverdefined()) { 786 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() 787 << "' - overdefined because of pred (local).\n"); 788 789 BBLV = Result; 790 return true; 791 } 792 } 793 794 // Return the merged value, which is more precise than 'overdefined'. 795 assert(!Result.isOverdefined() && "Possible PHI in entry block?"); 796 BBLV = Result; 797 return true; 798 } 799 800 static ValueLatticeElement getValueFromCondition(Value *Val, Value *Cond, 801 bool isTrueDest = true); 802 803 // If we can determine a constraint on the value given conditions assumed by 804 // the program, intersect those constraints with BBLV 805 void LazyValueInfoImpl::intersectAssumeOrGuardBlockValueConstantRange( 806 Value *Val, ValueLatticeElement &BBLV, Instruction *BBI) { 807 BBI = BBI ? BBI : dyn_cast<Instruction>(Val); 808 if (!BBI) 809 return; 810 811 for (auto &AssumeVH : AC->assumptionsFor(Val)) { 812 if (!AssumeVH) 813 continue; 814 auto *I = cast<CallInst>(AssumeVH); 815 if (!isValidAssumeForContext(I, BBI, DT)) 816 continue; 817 818 BBLV = intersect(BBLV, getValueFromCondition(Val, I->getArgOperand(0))); 819 } 820 821 // If guards are not used in the module, don't spend time looking for them 822 auto *GuardDecl = BBI->getModule()->getFunction( 823 Intrinsic::getName(Intrinsic::experimental_guard)); 824 if (!GuardDecl || GuardDecl->use_empty()) 825 return; 826 827 for (Instruction &I : make_range(BBI->getIterator().getReverse(), 828 BBI->getParent()->rend())) { 829 Value *Cond = nullptr; 830 if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>(m_Value(Cond)))) 831 BBLV = intersect(BBLV, getValueFromCondition(Val, Cond)); 832 } 833 } 834 835 bool LazyValueInfoImpl::solveBlockValueSelect(ValueLatticeElement &BBLV, 836 SelectInst *SI, BasicBlock *BB) { 837 838 // Recurse on our inputs if needed 839 if (!hasBlockValue(SI->getTrueValue(), BB)) { 840 if (pushBlockValue(std::make_pair(BB, SI->getTrueValue()))) 841 return false; 842 BBLV = ValueLatticeElement::getOverdefined(); 843 return true; 844 } 845 ValueLatticeElement TrueVal = getBlockValue(SI->getTrueValue(), BB); 846 // If we hit overdefined, don't ask more queries. We want to avoid poisoning 847 // extra slots in the table if we can. 848 if (TrueVal.isOverdefined()) { 849 BBLV = ValueLatticeElement::getOverdefined(); 850 return true; 851 } 852 853 if (!hasBlockValue(SI->getFalseValue(), BB)) { 854 if (pushBlockValue(std::make_pair(BB, SI->getFalseValue()))) 855 return false; 856 BBLV = ValueLatticeElement::getOverdefined(); 857 return true; 858 } 859 ValueLatticeElement FalseVal = getBlockValue(SI->getFalseValue(), BB); 860 // If we hit overdefined, don't ask more queries. We want to avoid poisoning 861 // extra slots in the table if we can. 862 if (FalseVal.isOverdefined()) { 863 BBLV = ValueLatticeElement::getOverdefined(); 864 return true; 865 } 866 867 if (TrueVal.isConstantRange() && FalseVal.isConstantRange()) { 868 const ConstantRange &TrueCR = TrueVal.getConstantRange(); 869 const ConstantRange &FalseCR = FalseVal.getConstantRange(); 870 Value *LHS = nullptr; 871 Value *RHS = nullptr; 872 SelectPatternResult SPR = matchSelectPattern(SI, LHS, RHS); 873 // Is this a min specifically of our two inputs? (Avoid the risk of 874 // ValueTracking getting smarter looking back past our immediate inputs.) 875 if (SelectPatternResult::isMinOrMax(SPR.Flavor) && 876 LHS == SI->getTrueValue() && RHS == SI->getFalseValue()) { 877 ConstantRange ResultCR = [&]() { 878 switch (SPR.Flavor) { 879 default: 880 llvm_unreachable("unexpected minmax type!"); 881 case SPF_SMIN: /// Signed minimum 882 return TrueCR.smin(FalseCR); 883 case SPF_UMIN: /// Unsigned minimum 884 return TrueCR.umin(FalseCR); 885 case SPF_SMAX: /// Signed maximum 886 return TrueCR.smax(FalseCR); 887 case SPF_UMAX: /// Unsigned maximum 888 return TrueCR.umax(FalseCR); 889 }; 890 }(); 891 BBLV = ValueLatticeElement::getRange(ResultCR); 892 return true; 893 } 894 895 // TODO: ABS, NABS from the SelectPatternResult 896 } 897 898 // Can we constrain the facts about the true and false values by using the 899 // condition itself? This shows up with idioms like e.g. select(a > 5, a, 5). 900 // TODO: We could potentially refine an overdefined true value above. 901 Value *Cond = SI->getCondition(); 902 TrueVal = intersect(TrueVal, 903 getValueFromCondition(SI->getTrueValue(), Cond, true)); 904 FalseVal = intersect(FalseVal, 905 getValueFromCondition(SI->getFalseValue(), Cond, false)); 906 907 // Handle clamp idioms such as: 908 // %24 = constantrange<0, 17> 909 // %39 = icmp eq i32 %24, 0 910 // %40 = add i32 %24, -1 911 // %siv.next = select i1 %39, i32 16, i32 %40 912 // %siv.next = constantrange<0, 17> not <-1, 17> 913 // In general, this can handle any clamp idiom which tests the edge 914 // condition via an equality or inequality. 915 if (auto *ICI = dyn_cast<ICmpInst>(Cond)) { 916 ICmpInst::Predicate Pred = ICI->getPredicate(); 917 Value *A = ICI->getOperand(0); 918 if (ConstantInt *CIBase = dyn_cast<ConstantInt>(ICI->getOperand(1))) { 919 auto addConstants = [](ConstantInt *A, ConstantInt *B) { 920 assert(A->getType() == B->getType()); 921 return ConstantInt::get(A->getType(), A->getValue() + B->getValue()); 922 }; 923 // See if either input is A + C2, subject to the constraint from the 924 // condition that A != C when that input is used. We can assume that 925 // that input doesn't include C + C2. 926 ConstantInt *CIAdded; 927 switch (Pred) { 928 default: break; 929 case ICmpInst::ICMP_EQ: 930 if (match(SI->getFalseValue(), m_Add(m_Specific(A), 931 m_ConstantInt(CIAdded)))) { 932 auto ResNot = addConstants(CIBase, CIAdded); 933 FalseVal = intersect(FalseVal, 934 ValueLatticeElement::getNot(ResNot)); 935 } 936 break; 937 case ICmpInst::ICMP_NE: 938 if (match(SI->getTrueValue(), m_Add(m_Specific(A), 939 m_ConstantInt(CIAdded)))) { 940 auto ResNot = addConstants(CIBase, CIAdded); 941 TrueVal = intersect(TrueVal, 942 ValueLatticeElement::getNot(ResNot)); 943 } 944 break; 945 }; 946 } 947 } 948 949 ValueLatticeElement Result; // Start Undefined. 950 Result.mergeIn(TrueVal, DL); 951 Result.mergeIn(FalseVal, DL); 952 BBLV = Result; 953 return true; 954 } 955 956 Optional<ConstantRange> LazyValueInfoImpl::getRangeForOperand(unsigned Op, 957 Instruction *I, 958 BasicBlock *BB) { 959 if (!hasBlockValue(I->getOperand(Op), BB)) 960 if (pushBlockValue(std::make_pair(BB, I->getOperand(Op)))) 961 return None; 962 963 const unsigned OperandBitWidth = 964 DL.getTypeSizeInBits(I->getOperand(Op)->getType()); 965 ConstantRange Range = ConstantRange(OperandBitWidth); 966 if (hasBlockValue(I->getOperand(Op), BB)) { 967 ValueLatticeElement Val = getBlockValue(I->getOperand(Op), BB); 968 intersectAssumeOrGuardBlockValueConstantRange(I->getOperand(Op), Val, I); 969 if (Val.isConstantRange()) 970 Range = Val.getConstantRange(); 971 } 972 return Range; 973 } 974 975 bool LazyValueInfoImpl::solveBlockValueCast(ValueLatticeElement &BBLV, 976 CastInst *CI, 977 BasicBlock *BB) { 978 if (!CI->getOperand(0)->getType()->isSized()) { 979 // Without knowing how wide the input is, we can't analyze it in any useful 980 // way. 981 BBLV = ValueLatticeElement::getOverdefined(); 982 return true; 983 } 984 985 // Filter out casts we don't know how to reason about before attempting to 986 // recurse on our operand. This can cut a long search short if we know we're 987 // not going to be able to get any useful information anways. 988 switch (CI->getOpcode()) { 989 case Instruction::Trunc: 990 case Instruction::SExt: 991 case Instruction::ZExt: 992 case Instruction::BitCast: 993 break; 994 default: 995 // Unhandled instructions are overdefined. 996 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() 997 << "' - overdefined (unknown cast).\n"); 998 BBLV = ValueLatticeElement::getOverdefined(); 999 return true; 1000 } 1001 1002 // Figure out the range of the LHS. If that fails, we still apply the 1003 // transfer rule on the full set since we may be able to locally infer 1004 // interesting facts. 1005 Optional<ConstantRange> LHSRes = getRangeForOperand(0, CI, BB); 1006 if (!LHSRes.hasValue()) 1007 // More work to do before applying this transfer rule. 1008 return false; 1009 ConstantRange LHSRange = LHSRes.getValue(); 1010 1011 const unsigned ResultBitWidth = CI->getType()->getIntegerBitWidth(); 1012 1013 // NOTE: We're currently limited by the set of operations that ConstantRange 1014 // can evaluate symbolically. Enhancing that set will allows us to analyze 1015 // more definitions. 1016 BBLV = ValueLatticeElement::getRange(LHSRange.castOp(CI->getOpcode(), 1017 ResultBitWidth)); 1018 return true; 1019 } 1020 1021 bool LazyValueInfoImpl::solveBlockValueBinaryOp(ValueLatticeElement &BBLV, 1022 BinaryOperator *BO, 1023 BasicBlock *BB) { 1024 1025 assert(BO->getOperand(0)->getType()->isSized() && 1026 "all operands to binary operators are sized"); 1027 1028 // Filter out operators we don't know how to reason about before attempting to 1029 // recurse on our operand(s). This can cut a long search short if we know 1030 // we're not going to be able to get any useful information anyways. 1031 switch (BO->getOpcode()) { 1032 case Instruction::Add: 1033 case Instruction::Sub: 1034 case Instruction::Mul: 1035 case Instruction::UDiv: 1036 case Instruction::Shl: 1037 case Instruction::LShr: 1038 case Instruction::AShr: 1039 case Instruction::And: 1040 case Instruction::Or: 1041 // continue into the code below 1042 break; 1043 default: 1044 // Unhandled instructions are overdefined. 1045 LLVM_DEBUG(dbgs() << " compute BB '" << BB->getName() 1046 << "' - overdefined (unknown binary operator).\n"); 1047 BBLV = ValueLatticeElement::getOverdefined(); 1048 return true; 1049 }; 1050 1051 // Figure out the ranges of the operands. If that fails, use a 1052 // conservative range, but apply the transfer rule anyways. This 1053 // lets us pick up facts from expressions like "and i32 (call i32 1054 // @foo()), 32" 1055 Optional<ConstantRange> LHSRes = getRangeForOperand(0, BO, BB); 1056 Optional<ConstantRange> RHSRes = getRangeForOperand(1, BO, BB); 1057 1058 if (!LHSRes.hasValue() || !RHSRes.hasValue()) 1059 // More work to do before applying this transfer rule. 1060 return false; 1061 1062 ConstantRange LHSRange = LHSRes.getValue(); 1063 ConstantRange RHSRange = RHSRes.getValue(); 1064 1065 // NOTE: We're currently limited by the set of operations that ConstantRange 1066 // can evaluate symbolically. Enhancing that set will allows us to analyze 1067 // more definitions. 1068 Instruction::BinaryOps BinOp = BO->getOpcode(); 1069 BBLV = ValueLatticeElement::getRange(LHSRange.binaryOp(BinOp, RHSRange)); 1070 return true; 1071 } 1072 1073 static ValueLatticeElement getValueFromICmpCondition(Value *Val, ICmpInst *ICI, 1074 bool isTrueDest) { 1075 Value *LHS = ICI->getOperand(0); 1076 Value *RHS = ICI->getOperand(1); 1077 CmpInst::Predicate Predicate = ICI->getPredicate(); 1078 1079 if (isa<Constant>(RHS)) { 1080 if (ICI->isEquality() && LHS == Val) { 1081 // We know that V has the RHS constant if this is a true SETEQ or 1082 // false SETNE. 1083 if (isTrueDest == (Predicate == ICmpInst::ICMP_EQ)) 1084 return ValueLatticeElement::get(cast<Constant>(RHS)); 1085 else 1086 return ValueLatticeElement::getNot(cast<Constant>(RHS)); 1087 } 1088 } 1089 1090 if (!Val->getType()->isIntegerTy()) 1091 return ValueLatticeElement::getOverdefined(); 1092 1093 // Use ConstantRange::makeAllowedICmpRegion in order to determine the possible 1094 // range of Val guaranteed by the condition. Recognize comparisons in the from 1095 // of: 1096 // icmp <pred> Val, ... 1097 // icmp <pred> (add Val, Offset), ... 1098 // The latter is the range checking idiom that InstCombine produces. Subtract 1099 // the offset from the allowed range for RHS in this case. 1100 1101 // Val or (add Val, Offset) can be on either hand of the comparison 1102 if (LHS != Val && !match(LHS, m_Add(m_Specific(Val), m_ConstantInt()))) { 1103 std::swap(LHS, RHS); 1104 Predicate = CmpInst::getSwappedPredicate(Predicate); 1105 } 1106 1107 ConstantInt *Offset = nullptr; 1108 if (LHS != Val) 1109 match(LHS, m_Add(m_Specific(Val), m_ConstantInt(Offset))); 1110 1111 if (LHS == Val || Offset) { 1112 // Calculate the range of values that are allowed by the comparison 1113 ConstantRange RHSRange(RHS->getType()->getIntegerBitWidth(), 1114 /*isFullSet=*/true); 1115 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) 1116 RHSRange = ConstantRange(CI->getValue()); 1117 else if (Instruction *I = dyn_cast<Instruction>(RHS)) 1118 if (auto *Ranges = I->getMetadata(LLVMContext::MD_range)) 1119 RHSRange = getConstantRangeFromMetadata(*Ranges); 1120 1121 // If we're interested in the false dest, invert the condition 1122 CmpInst::Predicate Pred = 1123 isTrueDest ? Predicate : CmpInst::getInversePredicate(Predicate); 1124 ConstantRange TrueValues = 1125 ConstantRange::makeAllowedICmpRegion(Pred, RHSRange); 1126 1127 if (Offset) // Apply the offset from above. 1128 TrueValues = TrueValues.subtract(Offset->getValue()); 1129 1130 return ValueLatticeElement::getRange(std::move(TrueValues)); 1131 } 1132 1133 return ValueLatticeElement::getOverdefined(); 1134 } 1135 1136 static ValueLatticeElement 1137 getValueFromCondition(Value *Val, Value *Cond, bool isTrueDest, 1138 DenseMap<Value*, ValueLatticeElement> &Visited); 1139 1140 static ValueLatticeElement 1141 getValueFromConditionImpl(Value *Val, Value *Cond, bool isTrueDest, 1142 DenseMap<Value*, ValueLatticeElement> &Visited) { 1143 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cond)) 1144 return getValueFromICmpCondition(Val, ICI, isTrueDest); 1145 1146 // Handle conditions in the form of (cond1 && cond2), we know that on the 1147 // true dest path both of the conditions hold. Similarly for conditions of 1148 // the form (cond1 || cond2), we know that on the false dest path neither 1149 // condition holds. 1150 BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond); 1151 if (!BO || (isTrueDest && BO->getOpcode() != BinaryOperator::And) || 1152 (!isTrueDest && BO->getOpcode() != BinaryOperator::Or)) 1153 return ValueLatticeElement::getOverdefined(); 1154 1155 // Prevent infinite recursion if Cond references itself as in this example: 1156 // Cond: "%tmp4 = and i1 %tmp4, undef" 1157 // BL: "%tmp4 = and i1 %tmp4, undef" 1158 // BR: "i1 undef" 1159 Value *BL = BO->getOperand(0); 1160 Value *BR = BO->getOperand(1); 1161 if (BL == Cond || BR == Cond) 1162 return ValueLatticeElement::getOverdefined(); 1163 1164 return intersect(getValueFromCondition(Val, BL, isTrueDest, Visited), 1165 getValueFromCondition(Val, BR, isTrueDest, Visited)); 1166 } 1167 1168 static ValueLatticeElement 1169 getValueFromCondition(Value *Val, Value *Cond, bool isTrueDest, 1170 DenseMap<Value*, ValueLatticeElement> &Visited) { 1171 auto I = Visited.find(Cond); 1172 if (I != Visited.end()) 1173 return I->second; 1174 1175 auto Result = getValueFromConditionImpl(Val, Cond, isTrueDest, Visited); 1176 Visited[Cond] = Result; 1177 return Result; 1178 } 1179 1180 ValueLatticeElement getValueFromCondition(Value *Val, Value *Cond, 1181 bool isTrueDest) { 1182 assert(Cond && "precondition"); 1183 DenseMap<Value*, ValueLatticeElement> Visited; 1184 return getValueFromCondition(Val, Cond, isTrueDest, Visited); 1185 } 1186 1187 // Return true if Usr has Op as an operand, otherwise false. 1188 static bool usesOperand(User *Usr, Value *Op) { 1189 return find(Usr->operands(), Op) != Usr->op_end(); 1190 } 1191 1192 // Return true if the instruction type of Val is supported by 1193 // constantFoldUser(). Currently CastInst and BinaryOperator only. Call this 1194 // before calling constantFoldUser() to find out if it's even worth attempting 1195 // to call it. 1196 static bool isOperationFoldable(User *Usr) { 1197 return isa<CastInst>(Usr) || isa<BinaryOperator>(Usr); 1198 } 1199 1200 // Check if Usr can be simplified to an integer constant when the value of one 1201 // of its operands Op is an integer constant OpConstVal. If so, return it as an 1202 // lattice value range with a single element or otherwise return an overdefined 1203 // lattice value. 1204 static ValueLatticeElement constantFoldUser(User *Usr, Value *Op, 1205 const APInt &OpConstVal, 1206 const DataLayout &DL) { 1207 assert(isOperationFoldable(Usr) && "Precondition"); 1208 Constant* OpConst = Constant::getIntegerValue(Op->getType(), OpConstVal); 1209 // Check if Usr can be simplified to a constant. 1210 if (auto *CI = dyn_cast<CastInst>(Usr)) { 1211 assert(CI->getOperand(0) == Op && "Operand 0 isn't Op"); 1212 if (auto *C = dyn_cast_or_null<ConstantInt>( 1213 SimplifyCastInst(CI->getOpcode(), OpConst, 1214 CI->getDestTy(), DL))) { 1215 return ValueLatticeElement::getRange(ConstantRange(C->getValue())); 1216 } 1217 } else if (auto *BO = dyn_cast<BinaryOperator>(Usr)) { 1218 bool Op0Match = BO->getOperand(0) == Op; 1219 bool Op1Match = BO->getOperand(1) == Op; 1220 assert((Op0Match || Op1Match) && 1221 "Operand 0 nor Operand 1 isn't a match"); 1222 Value *LHS = Op0Match ? OpConst : BO->getOperand(0); 1223 Value *RHS = Op1Match ? OpConst : BO->getOperand(1); 1224 if (auto *C = dyn_cast_or_null<ConstantInt>( 1225 SimplifyBinOp(BO->getOpcode(), LHS, RHS, DL))) { 1226 return ValueLatticeElement::getRange(ConstantRange(C->getValue())); 1227 } 1228 } 1229 return ValueLatticeElement::getOverdefined(); 1230 } 1231 1232 /// Compute the value of Val on the edge BBFrom -> BBTo. Returns false if 1233 /// Val is not constrained on the edge. Result is unspecified if return value 1234 /// is false. 1235 static bool getEdgeValueLocal(Value *Val, BasicBlock *BBFrom, 1236 BasicBlock *BBTo, ValueLatticeElement &Result) { 1237 // TODO: Handle more complex conditionals. If (v == 0 || v2 < 1) is false, we 1238 // know that v != 0. 1239 if (BranchInst *BI = dyn_cast<BranchInst>(BBFrom->getTerminator())) { 1240 // If this is a conditional branch and only one successor goes to BBTo, then 1241 // we may be able to infer something from the condition. 1242 if (BI->isConditional() && 1243 BI->getSuccessor(0) != BI->getSuccessor(1)) { 1244 bool isTrueDest = BI->getSuccessor(0) == BBTo; 1245 assert(BI->getSuccessor(!isTrueDest) == BBTo && 1246 "BBTo isn't a successor of BBFrom"); 1247 Value *Condition = BI->getCondition(); 1248 1249 // If V is the condition of the branch itself, then we know exactly what 1250 // it is. 1251 if (Condition == Val) { 1252 Result = ValueLatticeElement::get(ConstantInt::get( 1253 Type::getInt1Ty(Val->getContext()), isTrueDest)); 1254 return true; 1255 } 1256 1257 // If the condition of the branch is an equality comparison, we may be 1258 // able to infer the value. 1259 Result = getValueFromCondition(Val, Condition, isTrueDest); 1260 if (!Result.isOverdefined()) 1261 return true; 1262 1263 if (User *Usr = dyn_cast<User>(Val)) { 1264 assert(Result.isOverdefined() && "Result isn't overdefined"); 1265 // Check with isOperationFoldable() first to avoid linearly iterating 1266 // over the operands unnecessarily which can be expensive for 1267 // instructions with many operands. 1268 if (isa<IntegerType>(Usr->getType()) && isOperationFoldable(Usr)) { 1269 const DataLayout &DL = BBTo->getModule()->getDataLayout(); 1270 if (usesOperand(Usr, Condition)) { 1271 // If Val has Condition as an operand and Val can be folded into a 1272 // constant with either Condition == true or Condition == false, 1273 // propagate the constant. 1274 // eg. 1275 // ; %Val is true on the edge to %then. 1276 // %Val = and i1 %Condition, true. 1277 // br %Condition, label %then, label %else 1278 APInt ConditionVal(1, isTrueDest ? 1 : 0); 1279 Result = constantFoldUser(Usr, Condition, ConditionVal, DL); 1280 } else { 1281 // If one of Val's operand has an inferred value, we may be able to 1282 // infer the value of Val. 1283 // eg. 1284 // ; %Val is 94 on the edge to %then. 1285 // %Val = add i8 %Op, 1 1286 // %Condition = icmp eq i8 %Op, 93 1287 // br i1 %Condition, label %then, label %else 1288 for (unsigned i = 0; i < Usr->getNumOperands(); ++i) { 1289 Value *Op = Usr->getOperand(i); 1290 ValueLatticeElement OpLatticeVal = 1291 getValueFromCondition(Op, Condition, isTrueDest); 1292 if (Optional<APInt> OpConst = OpLatticeVal.asConstantInteger()) { 1293 Result = constantFoldUser(Usr, Op, OpConst.getValue(), DL); 1294 break; 1295 } 1296 } 1297 } 1298 } 1299 } 1300 if (!Result.isOverdefined()) 1301 return true; 1302 } 1303 } 1304 1305 // If the edge was formed by a switch on the value, then we may know exactly 1306 // what it is. 1307 if (SwitchInst *SI = dyn_cast<SwitchInst>(BBFrom->getTerminator())) { 1308 Value *Condition = SI->getCondition(); 1309 if (!isa<IntegerType>(Val->getType())) 1310 return false; 1311 bool ValUsesConditionAndMayBeFoldable = false; 1312 if (Condition != Val) { 1313 // Check if Val has Condition as an operand. 1314 if (User *Usr = dyn_cast<User>(Val)) 1315 ValUsesConditionAndMayBeFoldable = isOperationFoldable(Usr) && 1316 usesOperand(Usr, Condition); 1317 if (!ValUsesConditionAndMayBeFoldable) 1318 return false; 1319 } 1320 assert((Condition == Val || ValUsesConditionAndMayBeFoldable) && 1321 "Condition != Val nor Val doesn't use Condition"); 1322 1323 bool DefaultCase = SI->getDefaultDest() == BBTo; 1324 unsigned BitWidth = Val->getType()->getIntegerBitWidth(); 1325 ConstantRange EdgesVals(BitWidth, DefaultCase/*isFullSet*/); 1326 1327 for (auto Case : SI->cases()) { 1328 APInt CaseValue = Case.getCaseValue()->getValue(); 1329 ConstantRange EdgeVal(CaseValue); 1330 if (ValUsesConditionAndMayBeFoldable) { 1331 User *Usr = cast<User>(Val); 1332 const DataLayout &DL = BBTo->getModule()->getDataLayout(); 1333 ValueLatticeElement EdgeLatticeVal = 1334 constantFoldUser(Usr, Condition, CaseValue, DL); 1335 if (EdgeLatticeVal.isOverdefined()) 1336 return false; 1337 EdgeVal = EdgeLatticeVal.getConstantRange(); 1338 } 1339 if (DefaultCase) { 1340 // It is possible that the default destination is the destination of 1341 // some cases. We cannot perform difference for those cases. 1342 // We know Condition != CaseValue in BBTo. In some cases we can use 1343 // this to infer Val == f(Condition) is != f(CaseValue). For now, we 1344 // only do this when f is identity (i.e. Val == Condition), but we 1345 // should be able to do this for any injective f. 1346 if (Case.getCaseSuccessor() != BBTo && Condition == Val) 1347 EdgesVals = EdgesVals.difference(EdgeVal); 1348 } else if (Case.getCaseSuccessor() == BBTo) 1349 EdgesVals = EdgesVals.unionWith(EdgeVal); 1350 } 1351 Result = ValueLatticeElement::getRange(std::move(EdgesVals)); 1352 return true; 1353 } 1354 return false; 1355 } 1356 1357 /// Compute the value of Val on the edge BBFrom -> BBTo or the value at 1358 /// the basic block if the edge does not constrain Val. 1359 bool LazyValueInfoImpl::getEdgeValue(Value *Val, BasicBlock *BBFrom, 1360 BasicBlock *BBTo, 1361 ValueLatticeElement &Result, 1362 Instruction *CxtI) { 1363 // If already a constant, there is nothing to compute. 1364 if (Constant *VC = dyn_cast<Constant>(Val)) { 1365 Result = ValueLatticeElement::get(VC); 1366 return true; 1367 } 1368 1369 ValueLatticeElement LocalResult; 1370 if (!getEdgeValueLocal(Val, BBFrom, BBTo, LocalResult)) 1371 // If we couldn't constrain the value on the edge, LocalResult doesn't 1372 // provide any information. 1373 LocalResult = ValueLatticeElement::getOverdefined(); 1374 1375 if (hasSingleValue(LocalResult)) { 1376 // Can't get any more precise here 1377 Result = LocalResult; 1378 return true; 1379 } 1380 1381 if (!hasBlockValue(Val, BBFrom)) { 1382 if (pushBlockValue(std::make_pair(BBFrom, Val))) 1383 return false; 1384 // No new information. 1385 Result = LocalResult; 1386 return true; 1387 } 1388 1389 // Try to intersect ranges of the BB and the constraint on the edge. 1390 ValueLatticeElement InBlock = getBlockValue(Val, BBFrom); 1391 intersectAssumeOrGuardBlockValueConstantRange(Val, InBlock, 1392 BBFrom->getTerminator()); 1393 // We can use the context instruction (generically the ultimate instruction 1394 // the calling pass is trying to simplify) here, even though the result of 1395 // this function is generally cached when called from the solve* functions 1396 // (and that cached result might be used with queries using a different 1397 // context instruction), because when this function is called from the solve* 1398 // functions, the context instruction is not provided. When called from 1399 // LazyValueInfoImpl::getValueOnEdge, the context instruction is provided, 1400 // but then the result is not cached. 1401 intersectAssumeOrGuardBlockValueConstantRange(Val, InBlock, CxtI); 1402 1403 Result = intersect(LocalResult, InBlock); 1404 return true; 1405 } 1406 1407 ValueLatticeElement LazyValueInfoImpl::getValueInBlock(Value *V, BasicBlock *BB, 1408 Instruction *CxtI) { 1409 LLVM_DEBUG(dbgs() << "LVI Getting block end value " << *V << " at '" 1410 << BB->getName() << "'\n"); 1411 1412 assert(BlockValueStack.empty() && BlockValueSet.empty()); 1413 if (!hasBlockValue(V, BB)) { 1414 pushBlockValue(std::make_pair(BB, V)); 1415 solve(); 1416 } 1417 ValueLatticeElement Result = getBlockValue(V, BB); 1418 intersectAssumeOrGuardBlockValueConstantRange(V, Result, CxtI); 1419 1420 LLVM_DEBUG(dbgs() << " Result = " << Result << "\n"); 1421 return Result; 1422 } 1423 1424 ValueLatticeElement LazyValueInfoImpl::getValueAt(Value *V, Instruction *CxtI) { 1425 LLVM_DEBUG(dbgs() << "LVI Getting value " << *V << " at '" << CxtI->getName() 1426 << "'\n"); 1427 1428 if (auto *C = dyn_cast<Constant>(V)) 1429 return ValueLatticeElement::get(C); 1430 1431 ValueLatticeElement Result = ValueLatticeElement::getOverdefined(); 1432 if (auto *I = dyn_cast<Instruction>(V)) 1433 Result = getFromRangeMetadata(I); 1434 intersectAssumeOrGuardBlockValueConstantRange(V, Result, CxtI); 1435 1436 LLVM_DEBUG(dbgs() << " Result = " << Result << "\n"); 1437 return Result; 1438 } 1439 1440 ValueLatticeElement LazyValueInfoImpl:: 1441 getValueOnEdge(Value *V, BasicBlock *FromBB, BasicBlock *ToBB, 1442 Instruction *CxtI) { 1443 LLVM_DEBUG(dbgs() << "LVI Getting edge value " << *V << " from '" 1444 << FromBB->getName() << "' to '" << ToBB->getName() 1445 << "'\n"); 1446 1447 ValueLatticeElement Result; 1448 if (!getEdgeValue(V, FromBB, ToBB, Result, CxtI)) { 1449 solve(); 1450 bool WasFastQuery = getEdgeValue(V, FromBB, ToBB, Result, CxtI); 1451 (void)WasFastQuery; 1452 assert(WasFastQuery && "More work to do after problem solved?"); 1453 } 1454 1455 LLVM_DEBUG(dbgs() << " Result = " << Result << "\n"); 1456 return Result; 1457 } 1458 1459 void LazyValueInfoImpl::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc, 1460 BasicBlock *NewSucc) { 1461 TheCache.threadEdgeImpl(OldSucc, NewSucc); 1462 } 1463 1464 //===----------------------------------------------------------------------===// 1465 // LazyValueInfo Impl 1466 //===----------------------------------------------------------------------===// 1467 1468 /// This lazily constructs the LazyValueInfoImpl. 1469 static LazyValueInfoImpl &getImpl(void *&PImpl, AssumptionCache *AC, 1470 const DataLayout *DL, 1471 DominatorTree *DT = nullptr) { 1472 if (!PImpl) { 1473 assert(DL && "getCache() called with a null DataLayout"); 1474 PImpl = new LazyValueInfoImpl(AC, *DL, DT); 1475 } 1476 return *static_cast<LazyValueInfoImpl*>(PImpl); 1477 } 1478 1479 bool LazyValueInfoWrapperPass::runOnFunction(Function &F) { 1480 Info.AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 1481 const DataLayout &DL = F.getParent()->getDataLayout(); 1482 1483 DominatorTreeWrapperPass *DTWP = 1484 getAnalysisIfAvailable<DominatorTreeWrapperPass>(); 1485 Info.DT = DTWP ? &DTWP->getDomTree() : nullptr; 1486 Info.TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(); 1487 1488 if (Info.PImpl) 1489 getImpl(Info.PImpl, Info.AC, &DL, Info.DT).clear(); 1490 1491 // Fully lazy. 1492 return false; 1493 } 1494 1495 void LazyValueInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 1496 AU.setPreservesAll(); 1497 AU.addRequired<AssumptionCacheTracker>(); 1498 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1499 } 1500 1501 LazyValueInfo &LazyValueInfoWrapperPass::getLVI() { return Info; } 1502 1503 LazyValueInfo::~LazyValueInfo() { releaseMemory(); } 1504 1505 void LazyValueInfo::releaseMemory() { 1506 // If the cache was allocated, free it. 1507 if (PImpl) { 1508 delete &getImpl(PImpl, AC, nullptr); 1509 PImpl = nullptr; 1510 } 1511 } 1512 1513 bool LazyValueInfo::invalidate(Function &F, const PreservedAnalyses &PA, 1514 FunctionAnalysisManager::Invalidator &Inv) { 1515 // We need to invalidate if we have either failed to preserve this analyses 1516 // result directly or if any of its dependencies have been invalidated. 1517 auto PAC = PA.getChecker<LazyValueAnalysis>(); 1518 if (!(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>()) || 1519 (DT && Inv.invalidate<DominatorTreeAnalysis>(F, PA))) 1520 return true; 1521 1522 return false; 1523 } 1524 1525 void LazyValueInfoWrapperPass::releaseMemory() { Info.releaseMemory(); } 1526 1527 LazyValueInfo LazyValueAnalysis::run(Function &F, 1528 FunctionAnalysisManager &FAM) { 1529 auto &AC = FAM.getResult<AssumptionAnalysis>(F); 1530 auto &TLI = FAM.getResult<TargetLibraryAnalysis>(F); 1531 auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(F); 1532 1533 return LazyValueInfo(&AC, &F.getParent()->getDataLayout(), &TLI, DT); 1534 } 1535 1536 /// Returns true if we can statically tell that this value will never be a 1537 /// "useful" constant. In practice, this means we've got something like an 1538 /// alloca or a malloc call for which a comparison against a constant can 1539 /// only be guarding dead code. Note that we are potentially giving up some 1540 /// precision in dead code (a constant result) in favour of avoiding a 1541 /// expensive search for a easily answered common query. 1542 static bool isKnownNonConstant(Value *V) { 1543 V = V->stripPointerCasts(); 1544 // The return val of alloc cannot be a Constant. 1545 if (isa<AllocaInst>(V)) 1546 return true; 1547 return false; 1548 } 1549 1550 Constant *LazyValueInfo::getConstant(Value *V, BasicBlock *BB, 1551 Instruction *CxtI) { 1552 // Bail out early if V is known not to be a Constant. 1553 if (isKnownNonConstant(V)) 1554 return nullptr; 1555 1556 const DataLayout &DL = BB->getModule()->getDataLayout(); 1557 ValueLatticeElement Result = 1558 getImpl(PImpl, AC, &DL, DT).getValueInBlock(V, BB, CxtI); 1559 1560 if (Result.isConstant()) 1561 return Result.getConstant(); 1562 if (Result.isConstantRange()) { 1563 const ConstantRange &CR = Result.getConstantRange(); 1564 if (const APInt *SingleVal = CR.getSingleElement()) 1565 return ConstantInt::get(V->getContext(), *SingleVal); 1566 } 1567 return nullptr; 1568 } 1569 1570 ConstantRange LazyValueInfo::getConstantRange(Value *V, BasicBlock *BB, 1571 Instruction *CxtI) { 1572 assert(V->getType()->isIntegerTy()); 1573 unsigned Width = V->getType()->getIntegerBitWidth(); 1574 const DataLayout &DL = BB->getModule()->getDataLayout(); 1575 ValueLatticeElement Result = 1576 getImpl(PImpl, AC, &DL, DT).getValueInBlock(V, BB, CxtI); 1577 if (Result.isUndefined()) 1578 return ConstantRange(Width, /*isFullSet=*/false); 1579 if (Result.isConstantRange()) 1580 return Result.getConstantRange(); 1581 // We represent ConstantInt constants as constant ranges but other kinds 1582 // of integer constants, i.e. ConstantExpr will be tagged as constants 1583 assert(!(Result.isConstant() && isa<ConstantInt>(Result.getConstant())) && 1584 "ConstantInt value must be represented as constantrange"); 1585 return ConstantRange(Width, /*isFullSet=*/true); 1586 } 1587 1588 /// Determine whether the specified value is known to be a 1589 /// constant on the specified edge. Return null if not. 1590 Constant *LazyValueInfo::getConstantOnEdge(Value *V, BasicBlock *FromBB, 1591 BasicBlock *ToBB, 1592 Instruction *CxtI) { 1593 const DataLayout &DL = FromBB->getModule()->getDataLayout(); 1594 ValueLatticeElement Result = 1595 getImpl(PImpl, AC, &DL, DT).getValueOnEdge(V, FromBB, ToBB, CxtI); 1596 1597 if (Result.isConstant()) 1598 return Result.getConstant(); 1599 if (Result.isConstantRange()) { 1600 const ConstantRange &CR = Result.getConstantRange(); 1601 if (const APInt *SingleVal = CR.getSingleElement()) 1602 return ConstantInt::get(V->getContext(), *SingleVal); 1603 } 1604 return nullptr; 1605 } 1606 1607 ConstantRange LazyValueInfo::getConstantRangeOnEdge(Value *V, 1608 BasicBlock *FromBB, 1609 BasicBlock *ToBB, 1610 Instruction *CxtI) { 1611 unsigned Width = V->getType()->getIntegerBitWidth(); 1612 const DataLayout &DL = FromBB->getModule()->getDataLayout(); 1613 ValueLatticeElement Result = 1614 getImpl(PImpl, AC, &DL, DT).getValueOnEdge(V, FromBB, ToBB, CxtI); 1615 1616 if (Result.isUndefined()) 1617 return ConstantRange(Width, /*isFullSet=*/false); 1618 if (Result.isConstantRange()) 1619 return Result.getConstantRange(); 1620 // We represent ConstantInt constants as constant ranges but other kinds 1621 // of integer constants, i.e. ConstantExpr will be tagged as constants 1622 assert(!(Result.isConstant() && isa<ConstantInt>(Result.getConstant())) && 1623 "ConstantInt value must be represented as constantrange"); 1624 return ConstantRange(Width, /*isFullSet=*/true); 1625 } 1626 1627 static LazyValueInfo::Tristate 1628 getPredicateResult(unsigned Pred, Constant *C, const ValueLatticeElement &Val, 1629 const DataLayout &DL, TargetLibraryInfo *TLI) { 1630 // If we know the value is a constant, evaluate the conditional. 1631 Constant *Res = nullptr; 1632 if (Val.isConstant()) { 1633 Res = ConstantFoldCompareInstOperands(Pred, Val.getConstant(), C, DL, TLI); 1634 if (ConstantInt *ResCI = dyn_cast<ConstantInt>(Res)) 1635 return ResCI->isZero() ? LazyValueInfo::False : LazyValueInfo::True; 1636 return LazyValueInfo::Unknown; 1637 } 1638 1639 if (Val.isConstantRange()) { 1640 ConstantInt *CI = dyn_cast<ConstantInt>(C); 1641 if (!CI) return LazyValueInfo::Unknown; 1642 1643 const ConstantRange &CR = Val.getConstantRange(); 1644 if (Pred == ICmpInst::ICMP_EQ) { 1645 if (!CR.contains(CI->getValue())) 1646 return LazyValueInfo::False; 1647 1648 if (CR.isSingleElement()) 1649 return LazyValueInfo::True; 1650 } else if (Pred == ICmpInst::ICMP_NE) { 1651 if (!CR.contains(CI->getValue())) 1652 return LazyValueInfo::True; 1653 1654 if (CR.isSingleElement()) 1655 return LazyValueInfo::False; 1656 } else { 1657 // Handle more complex predicates. 1658 ConstantRange TrueValues = ConstantRange::makeExactICmpRegion( 1659 (ICmpInst::Predicate)Pred, CI->getValue()); 1660 if (TrueValues.contains(CR)) 1661 return LazyValueInfo::True; 1662 if (TrueValues.inverse().contains(CR)) 1663 return LazyValueInfo::False; 1664 } 1665 return LazyValueInfo::Unknown; 1666 } 1667 1668 if (Val.isNotConstant()) { 1669 // If this is an equality comparison, we can try to fold it knowing that 1670 // "V != C1". 1671 if (Pred == ICmpInst::ICMP_EQ) { 1672 // !C1 == C -> false iff C1 == C. 1673 Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE, 1674 Val.getNotConstant(), C, DL, 1675 TLI); 1676 if (Res->isNullValue()) 1677 return LazyValueInfo::False; 1678 } else if (Pred == ICmpInst::ICMP_NE) { 1679 // !C1 != C -> true iff C1 == C. 1680 Res = ConstantFoldCompareInstOperands(ICmpInst::ICMP_NE, 1681 Val.getNotConstant(), C, DL, 1682 TLI); 1683 if (Res->isNullValue()) 1684 return LazyValueInfo::True; 1685 } 1686 return LazyValueInfo::Unknown; 1687 } 1688 1689 return LazyValueInfo::Unknown; 1690 } 1691 1692 /// Determine whether the specified value comparison with a constant is known to 1693 /// be true or false on the specified CFG edge. Pred is a CmpInst predicate. 1694 LazyValueInfo::Tristate 1695 LazyValueInfo::getPredicateOnEdge(unsigned Pred, Value *V, Constant *C, 1696 BasicBlock *FromBB, BasicBlock *ToBB, 1697 Instruction *CxtI) { 1698 const DataLayout &DL = FromBB->getModule()->getDataLayout(); 1699 ValueLatticeElement Result = 1700 getImpl(PImpl, AC, &DL, DT).getValueOnEdge(V, FromBB, ToBB, CxtI); 1701 1702 return getPredicateResult(Pred, C, Result, DL, TLI); 1703 } 1704 1705 LazyValueInfo::Tristate 1706 LazyValueInfo::getPredicateAt(unsigned Pred, Value *V, Constant *C, 1707 Instruction *CxtI) { 1708 // Is or is not NonNull are common predicates being queried. If 1709 // isKnownNonZero can tell us the result of the predicate, we can 1710 // return it quickly. But this is only a fastpath, and falling 1711 // through would still be correct. 1712 const DataLayout &DL = CxtI->getModule()->getDataLayout(); 1713 if (V->getType()->isPointerTy() && C->isNullValue() && 1714 isKnownNonZero(V->stripPointerCasts(), DL)) { 1715 if (Pred == ICmpInst::ICMP_EQ) 1716 return LazyValueInfo::False; 1717 else if (Pred == ICmpInst::ICMP_NE) 1718 return LazyValueInfo::True; 1719 } 1720 ValueLatticeElement Result = getImpl(PImpl, AC, &DL, DT).getValueAt(V, CxtI); 1721 Tristate Ret = getPredicateResult(Pred, C, Result, DL, TLI); 1722 if (Ret != Unknown) 1723 return Ret; 1724 1725 // Note: The following bit of code is somewhat distinct from the rest of LVI; 1726 // LVI as a whole tries to compute a lattice value which is conservatively 1727 // correct at a given location. In this case, we have a predicate which we 1728 // weren't able to prove about the merged result, and we're pushing that 1729 // predicate back along each incoming edge to see if we can prove it 1730 // separately for each input. As a motivating example, consider: 1731 // bb1: 1732 // %v1 = ... ; constantrange<1, 5> 1733 // br label %merge 1734 // bb2: 1735 // %v2 = ... ; constantrange<10, 20> 1736 // br label %merge 1737 // merge: 1738 // %phi = phi [%v1, %v2] ; constantrange<1,20> 1739 // %pred = icmp eq i32 %phi, 8 1740 // We can't tell from the lattice value for '%phi' that '%pred' is false 1741 // along each path, but by checking the predicate over each input separately, 1742 // we can. 1743 // We limit the search to one step backwards from the current BB and value. 1744 // We could consider extending this to search further backwards through the 1745 // CFG and/or value graph, but there are non-obvious compile time vs quality 1746 // tradeoffs. 1747 if (CxtI) { 1748 BasicBlock *BB = CxtI->getParent(); 1749 1750 // Function entry or an unreachable block. Bail to avoid confusing 1751 // analysis below. 1752 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 1753 if (PI == PE) 1754 return Unknown; 1755 1756 // If V is a PHI node in the same block as the context, we need to ask 1757 // questions about the predicate as applied to the incoming value along 1758 // each edge. This is useful for eliminating cases where the predicate is 1759 // known along all incoming edges. 1760 if (auto *PHI = dyn_cast<PHINode>(V)) 1761 if (PHI->getParent() == BB) { 1762 Tristate Baseline = Unknown; 1763 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i < e; i++) { 1764 Value *Incoming = PHI->getIncomingValue(i); 1765 BasicBlock *PredBB = PHI->getIncomingBlock(i); 1766 // Note that PredBB may be BB itself. 1767 Tristate Result = getPredicateOnEdge(Pred, Incoming, C, PredBB, BB, 1768 CxtI); 1769 1770 // Keep going as long as we've seen a consistent known result for 1771 // all inputs. 1772 Baseline = (i == 0) ? Result /* First iteration */ 1773 : (Baseline == Result ? Baseline : Unknown); /* All others */ 1774 if (Baseline == Unknown) 1775 break; 1776 } 1777 if (Baseline != Unknown) 1778 return Baseline; 1779 } 1780 1781 // For a comparison where the V is outside this block, it's possible 1782 // that we've branched on it before. Look to see if the value is known 1783 // on all incoming edges. 1784 if (!isa<Instruction>(V) || 1785 cast<Instruction>(V)->getParent() != BB) { 1786 // For predecessor edge, determine if the comparison is true or false 1787 // on that edge. If they're all true or all false, we can conclude 1788 // the value of the comparison in this block. 1789 Tristate Baseline = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI); 1790 if (Baseline != Unknown) { 1791 // Check that all remaining incoming values match the first one. 1792 while (++PI != PE) { 1793 Tristate Ret = getPredicateOnEdge(Pred, V, C, *PI, BB, CxtI); 1794 if (Ret != Baseline) break; 1795 } 1796 // If we terminated early, then one of the values didn't match. 1797 if (PI == PE) { 1798 return Baseline; 1799 } 1800 } 1801 } 1802 } 1803 return Unknown; 1804 } 1805 1806 void LazyValueInfo::threadEdge(BasicBlock *PredBB, BasicBlock *OldSucc, 1807 BasicBlock *NewSucc) { 1808 if (PImpl) { 1809 const DataLayout &DL = PredBB->getModule()->getDataLayout(); 1810 getImpl(PImpl, AC, &DL, DT).threadEdge(PredBB, OldSucc, NewSucc); 1811 } 1812 } 1813 1814 void LazyValueInfo::eraseBlock(BasicBlock *BB) { 1815 if (PImpl) { 1816 const DataLayout &DL = BB->getModule()->getDataLayout(); 1817 getImpl(PImpl, AC, &DL, DT).eraseBlock(BB); 1818 } 1819 } 1820 1821 1822 void LazyValueInfo::printLVI(Function &F, DominatorTree &DTree, raw_ostream &OS) { 1823 if (PImpl) { 1824 getImpl(PImpl, AC, DL, DT).printLVI(F, DTree, OS); 1825 } 1826 } 1827 1828 void LazyValueInfo::disableDT() { 1829 if (PImpl) 1830 getImpl(PImpl, AC, DL, DT).disableDT(); 1831 } 1832 1833 void LazyValueInfo::enableDT() { 1834 if (PImpl) 1835 getImpl(PImpl, AC, DL, DT).enableDT(); 1836 } 1837 1838 // Print the LVI for the function arguments at the start of each basic block. 1839 void LazyValueInfoAnnotatedWriter::emitBasicBlockStartAnnot( 1840 const BasicBlock *BB, formatted_raw_ostream &OS) { 1841 // Find if there are latticevalues defined for arguments of the function. 1842 auto *F = BB->getParent(); 1843 for (auto &Arg : F->args()) { 1844 ValueLatticeElement Result = LVIImpl->getValueInBlock( 1845 const_cast<Argument *>(&Arg), const_cast<BasicBlock *>(BB)); 1846 if (Result.isUndefined()) 1847 continue; 1848 OS << "; LatticeVal for: '" << Arg << "' is: " << Result << "\n"; 1849 } 1850 } 1851 1852 // This function prints the LVI analysis for the instruction I at the beginning 1853 // of various basic blocks. It relies on calculated values that are stored in 1854 // the LazyValueInfoCache, and in the absence of cached values, recalculate the 1855 // LazyValueInfo for `I`, and print that info. 1856 void LazyValueInfoAnnotatedWriter::emitInstructionAnnot( 1857 const Instruction *I, formatted_raw_ostream &OS) { 1858 1859 auto *ParentBB = I->getParent(); 1860 SmallPtrSet<const BasicBlock*, 16> BlocksContainingLVI; 1861 // We can generate (solve) LVI values only for blocks that are dominated by 1862 // the I's parent. However, to avoid generating LVI for all dominating blocks, 1863 // that contain redundant/uninteresting information, we print LVI for 1864 // blocks that may use this LVI information (such as immediate successor 1865 // blocks, and blocks that contain uses of `I`). 1866 auto printResult = [&](const BasicBlock *BB) { 1867 if (!BlocksContainingLVI.insert(BB).second) 1868 return; 1869 ValueLatticeElement Result = LVIImpl->getValueInBlock( 1870 const_cast<Instruction *>(I), const_cast<BasicBlock *>(BB)); 1871 OS << "; LatticeVal for: '" << *I << "' in BB: '"; 1872 BB->printAsOperand(OS, false); 1873 OS << "' is: " << Result << "\n"; 1874 }; 1875 1876 printResult(ParentBB); 1877 // Print the LVI analysis results for the immediate successor blocks, that 1878 // are dominated by `ParentBB`. 1879 for (auto *BBSucc : successors(ParentBB)) 1880 if (DT.dominates(ParentBB, BBSucc)) 1881 printResult(BBSucc); 1882 1883 // Print LVI in blocks where `I` is used. 1884 for (auto *U : I->users()) 1885 if (auto *UseI = dyn_cast<Instruction>(U)) 1886 if (!isa<PHINode>(UseI) || DT.dominates(ParentBB, UseI->getParent())) 1887 printResult(UseI->getParent()); 1888 1889 } 1890 1891 namespace { 1892 // Printer class for LazyValueInfo results. 1893 class LazyValueInfoPrinter : public FunctionPass { 1894 public: 1895 static char ID; // Pass identification, replacement for typeid 1896 LazyValueInfoPrinter() : FunctionPass(ID) { 1897 initializeLazyValueInfoPrinterPass(*PassRegistry::getPassRegistry()); 1898 } 1899 1900 void getAnalysisUsage(AnalysisUsage &AU) const override { 1901 AU.setPreservesAll(); 1902 AU.addRequired<LazyValueInfoWrapperPass>(); 1903 AU.addRequired<DominatorTreeWrapperPass>(); 1904 } 1905 1906 // Get the mandatory dominator tree analysis and pass this in to the 1907 // LVIPrinter. We cannot rely on the LVI's DT, since it's optional. 1908 bool runOnFunction(Function &F) override { 1909 dbgs() << "LVI for function '" << F.getName() << "':\n"; 1910 auto &LVI = getAnalysis<LazyValueInfoWrapperPass>().getLVI(); 1911 auto &DTree = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1912 LVI.printLVI(F, DTree, dbgs()); 1913 return false; 1914 } 1915 }; 1916 } 1917 1918 char LazyValueInfoPrinter::ID = 0; 1919 INITIALIZE_PASS_BEGIN(LazyValueInfoPrinter, "print-lazy-value-info", 1920 "Lazy Value Info Printer Pass", false, false) 1921 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass) 1922 INITIALIZE_PASS_END(LazyValueInfoPrinter, "print-lazy-value-info", 1923 "Lazy Value Info Printer Pass", false, false) 1924