MemoryDependenceAnalysis.cpp revision 733c54da1e5432d9d64f88ea960121fa7a16076a
1//===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation --*- 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 implements an analysis that determines, for a given memory 11// operation, what preceding memory operations it depends on. It builds on 12// alias analysis information, and tries to provide a lazy, caching interface to 13// a common kind of alias information query. 14// 15//===----------------------------------------------------------------------===// 16 17#define DEBUG_TYPE "memdep" 18#include "llvm/Analysis/MemoryDependenceAnalysis.h" 19#include "llvm/Instructions.h" 20#include "llvm/IntrinsicInst.h" 21#include "llvm/Function.h" 22#include "llvm/LLVMContext.h" 23#include "llvm/Analysis/AliasAnalysis.h" 24#include "llvm/Analysis/Dominators.h" 25#include "llvm/Analysis/InstructionSimplify.h" 26#include "llvm/Analysis/MemoryBuiltins.h" 27#include "llvm/Analysis/PHITransAddr.h" 28#include "llvm/ADT/Statistic.h" 29#include "llvm/ADT/STLExtras.h" 30#include "llvm/Support/PredIteratorCache.h" 31#include "llvm/Support/Debug.h" 32using namespace llvm; 33 34STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 35STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 36STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 37 38STATISTIC(NumCacheNonLocalPtr, 39 "Number of fully cached non-local ptr responses"); 40STATISTIC(NumCacheDirtyNonLocalPtr, 41 "Number of cached, but dirty, non-local ptr responses"); 42STATISTIC(NumUncacheNonLocalPtr, 43 "Number of uncached non-local ptr responses"); 44STATISTIC(NumCacheCompleteNonLocalPtr, 45 "Number of block queries that were completely cached"); 46 47char MemoryDependenceAnalysis::ID = 0; 48 49// Register this pass... 50INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep", 51 "Memory Dependence Analysis", false, true) 52INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 53INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep", 54 "Memory Dependence Analysis", false, true) 55 56MemoryDependenceAnalysis::MemoryDependenceAnalysis() 57: FunctionPass(ID), PredCache(0) { 58 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry()); 59} 60MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 61} 62 63/// Clean up memory in between runs 64void MemoryDependenceAnalysis::releaseMemory() { 65 LocalDeps.clear(); 66 NonLocalDeps.clear(); 67 NonLocalPointerDeps.clear(); 68 ReverseLocalDeps.clear(); 69 ReverseNonLocalDeps.clear(); 70 ReverseNonLocalPtrDeps.clear(); 71 PredCache->clear(); 72} 73 74 75 76/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 77/// 78void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 79 AU.setPreservesAll(); 80 AU.addRequiredTransitive<AliasAnalysis>(); 81} 82 83bool MemoryDependenceAnalysis::runOnFunction(Function &) { 84 AA = &getAnalysis<AliasAnalysis>(); 85 if (PredCache == 0) 86 PredCache.reset(new PredIteratorCache()); 87 return false; 88} 89 90/// RemoveFromReverseMap - This is a helper function that removes Val from 91/// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry. 92template <typename KeyTy> 93static void RemoveFromReverseMap(DenseMap<Instruction*, 94 SmallPtrSet<KeyTy, 4> > &ReverseMap, 95 Instruction *Inst, KeyTy Val) { 96 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator 97 InstIt = ReverseMap.find(Inst); 98 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); 99 bool Found = InstIt->second.erase(Val); 100 assert(Found && "Invalid reverse map!"); Found=Found; 101 if (InstIt->second.empty()) 102 ReverseMap.erase(InstIt); 103} 104 105 106/// getCallSiteDependencyFrom - Private helper for finding the local 107/// dependencies of a call site. 108MemDepResult MemoryDependenceAnalysis:: 109getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 110 BasicBlock::iterator ScanIt, BasicBlock *BB) { 111 // Walk backwards through the block, looking for dependencies 112 while (ScanIt != BB->begin()) { 113 Instruction *Inst = --ScanIt; 114 115 // If this inst is a memory op, get the pointer it accessed 116 AliasAnalysis::Location Loc; 117 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) { 118 Loc = AliasAnalysis::Location(S->getPointerOperand(), 119 AA->getTypeStoreSize(S->getValueOperand() 120 ->getType()), 121 S->getMetadata(LLVMContext::MD_tbaa)); 122 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 123 Loc = AliasAnalysis::Location(V->getPointerOperand(), 124 AA->getTypeStoreSize(V->getType()), 125 V->getMetadata(LLVMContext::MD_tbaa)); 126 } else if (const CallInst *CI = isFreeCall(Inst)) { 127 // calls to free() erase the entire structure 128 Loc = AliasAnalysis::Location(CI->getArgOperand(0)); 129 } else if (CallSite InstCS = cast<Value>(Inst)) { 130 // Debug intrinsics don't cause dependences. 131 if (isa<DbgInfoIntrinsic>(Inst)) continue; 132 // If these two calls do not interfere, look past it. 133 switch (AA->getModRefInfo(CS, InstCS)) { 134 case AliasAnalysis::NoModRef: 135 // If the two calls are the same, return InstCS as a Def, so that 136 // CS can be found redundant and eliminated. 137 if (isReadOnlyCall && InstCS.onlyReadsMemory() && 138 CS.getInstruction()->isIdenticalToWhenDefined(Inst)) 139 return MemDepResult::getDef(Inst); 140 141 // Otherwise if the two calls don't interact (e.g. InstCS is readnone) 142 // keep scanning. 143 continue; 144 default: 145 return MemDepResult::getClobber(Inst); 146 } 147 } else { 148 // Non-memory instruction. 149 continue; 150 } 151 152 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef) 153 return MemDepResult::getClobber(Inst); 154 } 155 156 // No dependence found. If this is the entry block of the function, it is a 157 // clobber, otherwise it is non-local. 158 if (BB != &BB->getParent()->getEntryBlock()) 159 return MemDepResult::getNonLocal(); 160 return MemDepResult::getClobber(ScanIt); 161} 162 163/// getPointerDependencyFrom - Return the instruction on which a memory 164/// location depends. If isLoad is true, this routine ignores may-aliases with 165/// read-only operations. If isLoad is false, this routine ignores may-aliases 166/// with reads from read-only locations. 167MemDepResult MemoryDependenceAnalysis:: 168getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad, 169 BasicBlock::iterator ScanIt, BasicBlock *BB) { 170 171 Value *InvariantTag = 0; 172 173 // Walk backwards through the basic block, looking for dependencies. 174 while (ScanIt != BB->begin()) { 175 Instruction *Inst = --ScanIt; 176 177 // If we're in an invariant region, no dependencies can be found before 178 // we pass an invariant-begin marker. 179 if (InvariantTag == Inst) { 180 InvariantTag = 0; 181 continue; 182 } 183 184 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 185 // Debug intrinsics don't (and can't) cause dependences. 186 if (isa<DbgInfoIntrinsic>(II)) continue; 187 188 // If we pass an invariant-end marker, then we've just entered an 189 // invariant region and can start ignoring dependencies. 190 if (II->getIntrinsicID() == Intrinsic::invariant_end) { 191 // FIXME: This only considers queries directly on the invariant-tagged 192 // pointer, not on query pointers that are indexed off of them. It'd 193 // be nice to handle that at some point. 194 AliasAnalysis::AliasResult R = 195 AA->alias(AliasAnalysis::Location(II->getArgOperand(2)), MemLoc); 196 if (R == AliasAnalysis::MustAlias) 197 InvariantTag = II->getArgOperand(0); 198 199 continue; 200 } 201 202 // If we reach a lifetime begin or end marker, then the query ends here 203 // because the value is undefined. 204 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 205 // FIXME: This only considers queries directly on the invariant-tagged 206 // pointer, not on query pointers that are indexed off of them. It'd 207 // be nice to handle that at some point. 208 AliasAnalysis::AliasResult R = 209 AA->alias(AliasAnalysis::Location(II->getArgOperand(1)), MemLoc); 210 if (R == AliasAnalysis::MustAlias) 211 return MemDepResult::getDef(II); 212 continue; 213 } 214 } 215 216 // If we're querying on a load and we're in an invariant region, we're done 217 // at this point. Nothing a load depends on can live in an invariant region. 218 // 219 // FIXME: this will prevent us from returning load/load must-aliases, so GVN 220 // won't remove redundant loads. 221 if (isLoad && InvariantTag) continue; 222 223 // Values depend on loads if the pointers are must aliased. This means that 224 // a load depends on another must aliased load from the same value. 225 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 226 Value *Pointer = LI->getPointerOperand(); 227 uint64_t PointerSize = AA->getTypeStoreSize(LI->getType()); 228 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa); 229 AliasAnalysis::Location LoadLoc(Pointer, PointerSize, TBAATag); 230 231 // If we found a pointer, check if it could be the same as our pointer. 232 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc); 233 if (R == AliasAnalysis::NoAlias) 234 continue; 235 236 // May-alias loads don't depend on each other without a dependence. 237 if (isLoad && R == AliasAnalysis::MayAlias) 238 continue; 239 240 // Stores don't alias loads from read-only memory. 241 if (!isLoad && AA->pointsToConstantMemory(LoadLoc)) 242 continue; 243 244 // Stores depend on may and must aliased loads, loads depend on must-alias 245 // loads. 246 return MemDepResult::getDef(Inst); 247 } 248 249 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 250 // There can't be stores to the value we care about inside an 251 // invariant region. 252 if (InvariantTag) continue; 253 254 // If alias analysis can tell that this store is guaranteed to not modify 255 // the query pointer, ignore it. Use getModRefInfo to handle cases where 256 // the query pointer points to constant memory etc. 257 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef) 258 continue; 259 260 // Ok, this store might clobber the query pointer. Check to see if it is 261 // a must alias: in this case, we want to return this as a def. 262 Value *Pointer = SI->getPointerOperand(); 263 uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); 264 MDNode *TBAATag = SI->getMetadata(LLVMContext::MD_tbaa); 265 266 // If we found a pointer, check if it could be the same as our pointer. 267 AliasAnalysis::AliasResult R = 268 AA->alias(AliasAnalysis::Location(Pointer, PointerSize, TBAATag), 269 MemLoc); 270 271 if (R == AliasAnalysis::NoAlias) 272 continue; 273 if (R == AliasAnalysis::MayAlias) 274 return MemDepResult::getClobber(Inst); 275 return MemDepResult::getDef(Inst); 276 } 277 278 // If this is an allocation, and if we know that the accessed pointer is to 279 // the allocation, return Def. This means that there is no dependence and 280 // the access can be optimized based on that. For example, a load could 281 // turn into undef. 282 // Note: Only determine this to be a malloc if Inst is the malloc call, not 283 // a subsequent bitcast of the malloc call result. There can be stores to 284 // the malloced memory between the malloc call and its bitcast uses, and we 285 // need to continue scanning until the malloc call. 286 if (isa<AllocaInst>(Inst) || 287 (isa<CallInst>(Inst) && extractMallocCall(Inst))) { 288 const Value *AccessPtr = MemLoc.Ptr->getUnderlyingObject(); 289 290 if (AccessPtr == Inst || 291 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 292 return MemDepResult::getDef(Inst); 293 continue; 294 } 295 296 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 297 switch (AA->getModRefInfo(Inst, MemLoc)) { 298 case AliasAnalysis::NoModRef: 299 // If the call has no effect on the queried pointer, just ignore it. 300 continue; 301 case AliasAnalysis::Mod: 302 // If we're in an invariant region, we can ignore calls that ONLY 303 // modify the pointer. 304 if (InvariantTag) continue; 305 return MemDepResult::getClobber(Inst); 306 case AliasAnalysis::Ref: 307 // If the call is known to never store to the pointer, and if this is a 308 // load query, we can safely ignore it (scan past it). 309 if (isLoad) 310 continue; 311 default: 312 // Otherwise, there is a potential dependence. Return a clobber. 313 return MemDepResult::getClobber(Inst); 314 } 315 } 316 317 // No dependence found. If this is the entry block of the function, it is a 318 // clobber, otherwise it is non-local. 319 if (BB != &BB->getParent()->getEntryBlock()) 320 return MemDepResult::getNonLocal(); 321 return MemDepResult::getClobber(ScanIt); 322} 323 324/// getDependency - Return the instruction on which a memory operation 325/// depends. 326MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 327 Instruction *ScanPos = QueryInst; 328 329 // Check for a cached result 330 MemDepResult &LocalCache = LocalDeps[QueryInst]; 331 332 // If the cached entry is non-dirty, just return it. Note that this depends 333 // on MemDepResult's default constructing to 'dirty'. 334 if (!LocalCache.isDirty()) 335 return LocalCache; 336 337 // Otherwise, if we have a dirty entry, we know we can start the scan at that 338 // instruction, which may save us some work. 339 if (Instruction *Inst = LocalCache.getInst()) { 340 ScanPos = Inst; 341 342 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 343 } 344 345 BasicBlock *QueryParent = QueryInst->getParent(); 346 347 AliasAnalysis::Location MemLoc; 348 349 // Do the scan. 350 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 351 // No dependence found. If this is the entry block of the function, it is a 352 // clobber, otherwise it is non-local. 353 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 354 LocalCache = MemDepResult::getNonLocal(); 355 else 356 LocalCache = MemDepResult::getClobber(QueryInst); 357 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) { 358 // If this is a volatile store, don't mess around with it. Just return the 359 // previous instruction as a clobber. 360 if (SI->isVolatile()) 361 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 362 else 363 MemLoc = AliasAnalysis::Location(SI->getPointerOperand(), 364 AA->getTypeStoreSize(SI->getOperand(0) 365 ->getType()), 366 SI->getMetadata(LLVMContext::MD_tbaa)); 367 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) { 368 // If this is a volatile load, don't mess around with it. Just return the 369 // previous instruction as a clobber. 370 if (LI->isVolatile()) 371 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 372 else 373 MemLoc = AliasAnalysis::Location(LI->getPointerOperand(), 374 AA->getTypeStoreSize(LI->getType()), 375 LI->getMetadata(LLVMContext::MD_tbaa)); 376 } else if (const CallInst *CI = isFreeCall(QueryInst)) { 377 // calls to free() erase the entire structure, not just a field. 378 MemLoc = AliasAnalysis::Location(CI->getArgOperand(0)); 379 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 380 int IntrinsicID = 0; // Intrinsic IDs start at 1. 381 IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst); 382 if (II) 383 IntrinsicID = II->getIntrinsicID(); 384 385 switch (IntrinsicID) { 386 case Intrinsic::lifetime_start: 387 case Intrinsic::lifetime_end: 388 case Intrinsic::invariant_start: 389 MemLoc = AliasAnalysis::Location(II->getArgOperand(1), 390 cast<ConstantInt>(II->getArgOperand(0)) 391 ->getZExtValue(), 392 II->getMetadata(LLVMContext::MD_tbaa)); 393 break; 394 case Intrinsic::invariant_end: 395 MemLoc = AliasAnalysis::Location(II->getArgOperand(2), 396 cast<ConstantInt>(II->getArgOperand(1)) 397 ->getZExtValue(), 398 II->getMetadata(LLVMContext::MD_tbaa)); 399 break; 400 default: 401 CallSite QueryCS(QueryInst); 402 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 403 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 404 QueryParent); 405 break; 406 } 407 } else { 408 // Non-memory instruction. 409 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 410 } 411 412 // If we need to do a pointer scan, make it happen. 413 if (MemLoc.Ptr) { 414 bool isLoad = !QueryInst->mayWriteToMemory(); 415 if (IntrinsicInst *II = dyn_cast<MemoryUseIntrinsic>(QueryInst)) { 416 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end; 417 } 418 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos, 419 QueryParent); 420 } 421 422 // Remember the result! 423 if (Instruction *I = LocalCache.getInst()) 424 ReverseLocalDeps[I].insert(QueryInst); 425 426 return LocalCache; 427} 428 429#ifndef NDEBUG 430/// AssertSorted - This method is used when -debug is specified to verify that 431/// cache arrays are properly kept sorted. 432static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 433 int Count = -1) { 434 if (Count == -1) Count = Cache.size(); 435 if (Count == 0) return; 436 437 for (unsigned i = 1; i != unsigned(Count); ++i) 438 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 439} 440#endif 441 442/// getNonLocalCallDependency - Perform a full dependency query for the 443/// specified call, returning the set of blocks that the value is 444/// potentially live across. The returned set of results will include a 445/// "NonLocal" result for all blocks where the value is live across. 446/// 447/// This method assumes the instruction returns a "NonLocal" dependency 448/// within its own block. 449/// 450/// This returns a reference to an internal data structure that may be 451/// invalidated on the next non-local query or when an instruction is 452/// removed. Clients must copy this data if they want it around longer than 453/// that. 454const MemoryDependenceAnalysis::NonLocalDepInfo & 455MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 456 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 457 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 458 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 459 NonLocalDepInfo &Cache = CacheP.first; 460 461 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 462 /// the cached case, this can happen due to instructions being deleted etc. In 463 /// the uncached case, this starts out as the set of predecessors we care 464 /// about. 465 SmallVector<BasicBlock*, 32> DirtyBlocks; 466 467 if (!Cache.empty()) { 468 // Okay, we have a cache entry. If we know it is not dirty, just return it 469 // with no computation. 470 if (!CacheP.second) { 471 ++NumCacheNonLocal; 472 return Cache; 473 } 474 475 // If we already have a partially computed set of results, scan them to 476 // determine what is dirty, seeding our initial DirtyBlocks worklist. 477 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 478 I != E; ++I) 479 if (I->getResult().isDirty()) 480 DirtyBlocks.push_back(I->getBB()); 481 482 // Sort the cache so that we can do fast binary search lookups below. 483 std::sort(Cache.begin(), Cache.end()); 484 485 ++NumCacheDirtyNonLocal; 486 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 487 // << Cache.size() << " cached: " << *QueryInst; 488 } else { 489 // Seed DirtyBlocks with each of the preds of QueryInst's block. 490 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 491 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 492 DirtyBlocks.push_back(*PI); 493 ++NumUncacheNonLocal; 494 } 495 496 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 497 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 498 499 SmallPtrSet<BasicBlock*, 64> Visited; 500 501 unsigned NumSortedEntries = Cache.size(); 502 DEBUG(AssertSorted(Cache)); 503 504 // Iterate while we still have blocks to update. 505 while (!DirtyBlocks.empty()) { 506 BasicBlock *DirtyBB = DirtyBlocks.back(); 507 DirtyBlocks.pop_back(); 508 509 // Already processed this block? 510 if (!Visited.insert(DirtyBB)) 511 continue; 512 513 // Do a binary search to see if we already have an entry for this block in 514 // the cache set. If so, find it. 515 DEBUG(AssertSorted(Cache, NumSortedEntries)); 516 NonLocalDepInfo::iterator Entry = 517 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 518 NonLocalDepEntry(DirtyBB)); 519 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB) 520 --Entry; 521 522 NonLocalDepEntry *ExistingResult = 0; 523 if (Entry != Cache.begin()+NumSortedEntries && 524 Entry->getBB() == DirtyBB) { 525 // If we already have an entry, and if it isn't already dirty, the block 526 // is done. 527 if (!Entry->getResult().isDirty()) 528 continue; 529 530 // Otherwise, remember this slot so we can update the value. 531 ExistingResult = &*Entry; 532 } 533 534 // If the dirty entry has a pointer, start scanning from it so we don't have 535 // to rescan the entire block. 536 BasicBlock::iterator ScanPos = DirtyBB->end(); 537 if (ExistingResult) { 538 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 539 ScanPos = Inst; 540 // We're removing QueryInst's use of Inst. 541 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 542 QueryCS.getInstruction()); 543 } 544 } 545 546 // Find out if this block has a local dependency for QueryInst. 547 MemDepResult Dep; 548 549 if (ScanPos != DirtyBB->begin()) { 550 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 551 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 552 // No dependence found. If this is the entry block of the function, it is 553 // a clobber, otherwise it is non-local. 554 Dep = MemDepResult::getNonLocal(); 555 } else { 556 Dep = MemDepResult::getClobber(ScanPos); 557 } 558 559 // If we had a dirty entry for the block, update it. Otherwise, just add 560 // a new entry. 561 if (ExistingResult) 562 ExistingResult->setResult(Dep); 563 else 564 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 565 566 // If the block has a dependency (i.e. it isn't completely transparent to 567 // the value), remember the association! 568 if (!Dep.isNonLocal()) { 569 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 570 // update this when we remove instructions. 571 if (Instruction *Inst = Dep.getInst()) 572 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 573 } else { 574 575 // If the block *is* completely transparent to the load, we need to check 576 // the predecessors of this block. Add them to our worklist. 577 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 578 DirtyBlocks.push_back(*PI); 579 } 580 } 581 582 return Cache; 583} 584 585/// getNonLocalPointerDependency - Perform a full dependency query for an 586/// access to the specified (non-volatile) memory location, returning the 587/// set of instructions that either define or clobber the value. 588/// 589/// This method assumes the pointer has a "NonLocal" dependency within its 590/// own block. 591/// 592void MemoryDependenceAnalysis:: 593getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad, 594 BasicBlock *FromBB, 595 SmallVectorImpl<NonLocalDepResult> &Result) { 596 assert(Loc.Ptr->getType()->isPointerTy() && 597 "Can't get pointer deps of a non-pointer!"); 598 Result.clear(); 599 600 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD); 601 602 // This is the set of blocks we've inspected, and the pointer we consider in 603 // each block. Because of critical edges, we currently bail out if querying 604 // a block with multiple different pointers. This can happen during PHI 605 // translation. 606 DenseMap<BasicBlock*, Value*> Visited; 607 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB, 608 Result, Visited, true)) 609 return; 610 Result.clear(); 611 Result.push_back(NonLocalDepResult(FromBB, 612 MemDepResult::getClobber(FromBB->begin()), 613 const_cast<Value *>(Loc.Ptr))); 614} 615 616/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 617/// Pointer/PointeeSize using either cached information in Cache or by doing a 618/// lookup (which may use dirty cache info if available). If we do a lookup, 619/// add the result to the cache. 620MemDepResult MemoryDependenceAnalysis:: 621GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc, 622 bool isLoad, BasicBlock *BB, 623 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 624 625 // Do a binary search to see if we already have an entry for this block in 626 // the cache set. If so, find it. 627 NonLocalDepInfo::iterator Entry = 628 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 629 NonLocalDepEntry(BB)); 630 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 631 --Entry; 632 633 NonLocalDepEntry *ExistingResult = 0; 634 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 635 ExistingResult = &*Entry; 636 637 // If we have a cached entry, and it is non-dirty, use it as the value for 638 // this dependency. 639 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 640 ++NumCacheNonLocalPtr; 641 return ExistingResult->getResult(); 642 } 643 644 // Otherwise, we have to scan for the value. If we have a dirty cache 645 // entry, start scanning from its position, otherwise we scan from the end 646 // of the block. 647 BasicBlock::iterator ScanPos = BB->end(); 648 if (ExistingResult && ExistingResult->getResult().getInst()) { 649 assert(ExistingResult->getResult().getInst()->getParent() == BB && 650 "Instruction invalidated?"); 651 ++NumCacheDirtyNonLocalPtr; 652 ScanPos = ExistingResult->getResult().getInst(); 653 654 // Eliminating the dirty entry from 'Cache', so update the reverse info. 655 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 656 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 657 } else { 658 ++NumUncacheNonLocalPtr; 659 } 660 661 // Scan the block for the dependency. 662 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB); 663 664 // If we had a dirty entry for the block, update it. Otherwise, just add 665 // a new entry. 666 if (ExistingResult) 667 ExistingResult->setResult(Dep); 668 else 669 Cache->push_back(NonLocalDepEntry(BB, Dep)); 670 671 // If the block has a dependency (i.e. it isn't completely transparent to 672 // the value), remember the reverse association because we just added it 673 // to Cache! 674 if (Dep.isNonLocal()) 675 return Dep; 676 677 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 678 // update MemDep when we remove instructions. 679 Instruction *Inst = Dep.getInst(); 680 assert(Inst && "Didn't depend on anything?"); 681 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 682 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 683 return Dep; 684} 685 686/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 687/// number of elements in the array that are already properly ordered. This is 688/// optimized for the case when only a few entries are added. 689static void 690SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 691 unsigned NumSortedEntries) { 692 switch (Cache.size() - NumSortedEntries) { 693 case 0: 694 // done, no new entries. 695 break; 696 case 2: { 697 // Two new entries, insert the last one into place. 698 NonLocalDepEntry Val = Cache.back(); 699 Cache.pop_back(); 700 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 701 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 702 Cache.insert(Entry, Val); 703 // FALL THROUGH. 704 } 705 case 1: 706 // One new entry, Just insert the new value at the appropriate position. 707 if (Cache.size() != 1) { 708 NonLocalDepEntry Val = Cache.back(); 709 Cache.pop_back(); 710 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 711 std::upper_bound(Cache.begin(), Cache.end(), Val); 712 Cache.insert(Entry, Val); 713 } 714 break; 715 default: 716 // Added many values, do a full scale sort. 717 std::sort(Cache.begin(), Cache.end()); 718 break; 719 } 720} 721 722/// getNonLocalPointerDepFromBB - Perform a dependency query based on 723/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 724/// results to the results vector and keep track of which blocks are visited in 725/// 'Visited'. 726/// 727/// This has special behavior for the first block queries (when SkipFirstBlock 728/// is true). In this special case, it ignores the contents of the specified 729/// block and starts returning dependence info for its predecessors. 730/// 731/// This function returns false on success, or true to indicate that it could 732/// not compute dependence information for some reason. This should be treated 733/// as a clobber dependence on the first instruction in the predecessor block. 734bool MemoryDependenceAnalysis:: 735getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, 736 const AliasAnalysis::Location &Loc, 737 bool isLoad, BasicBlock *StartBB, 738 SmallVectorImpl<NonLocalDepResult> &Result, 739 DenseMap<BasicBlock*, Value*> &Visited, 740 bool SkipFirstBlock) { 741 742 // Look up the cached info for Pointer. 743 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 744 745 // Set up a temporary NLPI value. If the map doesn't yet have an entry for 746 // CacheKey, this value will be inserted as the associated value. Otherwise, 747 // it'll be ignored, and we'll have to check to see if the cached size and 748 // tbaa tag are consistent with the current query. 749 NonLocalPointerInfo InitialNLPI; 750 InitialNLPI.Size = Loc.Size; 751 InitialNLPI.TBAATag = Loc.TBAATag; 752 753 // Get the NLPI for CacheKey, inserting one into the map if it doesn't 754 // already have one. 755 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair = 756 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI)); 757 NonLocalPointerInfo *CacheInfo = &Pair.first->second; 758 759 // If we already have a cache entry for this CacheKey, we may need to do some 760 // work to reconcile the cache entry and the current query. 761 if (!Pair.second) { 762 if (CacheInfo->Size < Loc.Size) { 763 // The query's Size is greater than the cached one. Throw out the 764 // cached data and procede with the query at the greater size. 765 CacheInfo->Pair = BBSkipFirstBlockPair(); 766 CacheInfo->Size = Loc.Size; 767 CacheInfo->NonLocalDeps.clear(); 768 } else if (CacheInfo->Size > Loc.Size) { 769 // This query's Size is less than the cached one. Conservatively restart 770 // the query using the greater size. 771 return getNonLocalPointerDepFromBB(Pointer, 772 Loc.getWithNewSize(CacheInfo->Size), 773 isLoad, StartBB, Result, Visited, 774 SkipFirstBlock); 775 } 776 777 // If the query's TBAATag is inconsistent with the cached one, 778 // conservatively throw out the cached data and restart the query with 779 // no tag if needed. 780 if (CacheInfo->TBAATag != Loc.TBAATag) { 781 if (CacheInfo->TBAATag) { 782 CacheInfo->Pair = BBSkipFirstBlockPair(); 783 CacheInfo->TBAATag = 0; 784 CacheInfo->NonLocalDeps.clear(); 785 } 786 if (Loc.TBAATag) 787 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(), 788 isLoad, StartBB, Result, Visited, 789 SkipFirstBlock); 790 } 791 } 792 793 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps; 794 795 // If we have valid cached information for exactly the block we are 796 // investigating, just return it with no recomputation. 797 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 798 // We have a fully cached result for this query then we can just return the 799 // cached results and populate the visited set. However, we have to verify 800 // that we don't already have conflicting results for these blocks. Check 801 // to ensure that if a block in the results set is in the visited set that 802 // it was for the same pointer query. 803 if (!Visited.empty()) { 804 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 805 I != E; ++I) { 806 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 807 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 808 continue; 809 810 // We have a pointer mismatch in a block. Just return clobber, saying 811 // that something was clobbered in this result. We could also do a 812 // non-fully cached query, but there is little point in doing this. 813 return true; 814 } 815 } 816 817 Value *Addr = Pointer.getAddr(); 818 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 819 I != E; ++I) { 820 Visited.insert(std::make_pair(I->getBB(), Addr)); 821 if (!I->getResult().isNonLocal()) 822 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 823 } 824 ++NumCacheCompleteNonLocalPtr; 825 return false; 826 } 827 828 // Otherwise, either this is a new block, a block with an invalid cache 829 // pointer or one that we're about to invalidate by putting more info into it 830 // than its valid cache info. If empty, the result will be valid cache info, 831 // otherwise it isn't. 832 if (Cache->empty()) 833 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 834 else { 835 CacheInfo->Pair = BBSkipFirstBlockPair(); 836 CacheInfo->Size = 0; 837 CacheInfo->TBAATag = 0; 838 } 839 840 SmallVector<BasicBlock*, 32> Worklist; 841 Worklist.push_back(StartBB); 842 843 // Keep track of the entries that we know are sorted. Previously cached 844 // entries will all be sorted. The entries we add we only sort on demand (we 845 // don't insert every element into its sorted position). We know that we 846 // won't get any reuse from currently inserted values, because we don't 847 // revisit blocks after we insert info for them. 848 unsigned NumSortedEntries = Cache->size(); 849 DEBUG(AssertSorted(*Cache)); 850 851 while (!Worklist.empty()) { 852 BasicBlock *BB = Worklist.pop_back_val(); 853 854 // Skip the first block if we have it. 855 if (!SkipFirstBlock) { 856 // Analyze the dependency of *Pointer in FromBB. See if we already have 857 // been here. 858 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 859 860 // Get the dependency info for Pointer in BB. If we have cached 861 // information, we will use it, otherwise we compute it. 862 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 863 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache, 864 NumSortedEntries); 865 866 // If we got a Def or Clobber, add this to the list of results. 867 if (!Dep.isNonLocal()) { 868 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 869 continue; 870 } 871 } 872 873 // If 'Pointer' is an instruction defined in this block, then we need to do 874 // phi translation to change it into a value live in the predecessor block. 875 // If not, we just add the predecessors to the worklist and scan them with 876 // the same Pointer. 877 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 878 SkipFirstBlock = false; 879 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 880 // Verify that we haven't looked at this block yet. 881 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 882 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 883 if (InsertRes.second) { 884 // First time we've looked at *PI. 885 Worklist.push_back(*PI); 886 continue; 887 } 888 889 // If we have seen this block before, but it was with a different 890 // pointer then we have a phi translation failure and we have to treat 891 // this as a clobber. 892 if (InsertRes.first->second != Pointer.getAddr()) 893 goto PredTranslationFailure; 894 } 895 continue; 896 } 897 898 // We do need to do phi translation, if we know ahead of time we can't phi 899 // translate this value, don't even try. 900 if (!Pointer.IsPotentiallyPHITranslatable()) 901 goto PredTranslationFailure; 902 903 // We may have added values to the cache list before this PHI translation. 904 // If so, we haven't done anything to ensure that the cache remains sorted. 905 // Sort it now (if needed) so that recursive invocations of 906 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 907 // value will only see properly sorted cache arrays. 908 if (Cache && NumSortedEntries != Cache->size()) { 909 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 910 NumSortedEntries = Cache->size(); 911 } 912 Cache = 0; 913 914 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 915 BasicBlock *Pred = *PI; 916 917 // Get the PHI translated pointer in this predecessor. This can fail if 918 // not translatable, in which case the getAddr() returns null. 919 PHITransAddr PredPointer(Pointer); 920 PredPointer.PHITranslateValue(BB, Pred, 0); 921 922 Value *PredPtrVal = PredPointer.getAddr(); 923 924 // Check to see if we have already visited this pred block with another 925 // pointer. If so, we can't do this lookup. This failure can occur 926 // with PHI translation when a critical edge exists and the PHI node in 927 // the successor translates to a pointer value different than the 928 // pointer the block was first analyzed with. 929 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 930 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 931 932 if (!InsertRes.second) { 933 // If the predecessor was visited with PredPtr, then we already did 934 // the analysis and can ignore it. 935 if (InsertRes.first->second == PredPtrVal) 936 continue; 937 938 // Otherwise, the block was previously analyzed with a different 939 // pointer. We can't represent the result of this case, so we just 940 // treat this as a phi translation failure. 941 goto PredTranslationFailure; 942 } 943 944 // If PHI translation was unable to find an available pointer in this 945 // predecessor, then we have to assume that the pointer is clobbered in 946 // that predecessor. We can still do PRE of the load, which would insert 947 // a computation of the pointer in this predecessor. 948 if (PredPtrVal == 0) { 949 // Add the entry to the Result list. 950 NonLocalDepResult Entry(Pred, 951 MemDepResult::getClobber(Pred->getTerminator()), 952 PredPtrVal); 953 Result.push_back(Entry); 954 955 // Since we had a phi translation failure, the cache for CacheKey won't 956 // include all of the entries that we need to immediately satisfy future 957 // queries. Mark this in NonLocalPointerDeps by setting the 958 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 959 // cached value to do more work but not miss the phi trans failure. 960 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey]; 961 NLPI.Pair = BBSkipFirstBlockPair(); 962 NLPI.Size = 0; 963 NLPI.TBAATag = 0; 964 continue; 965 } 966 967 // FIXME: it is entirely possible that PHI translating will end up with 968 // the same value. Consider PHI translating something like: 969 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 970 // to recurse here, pedantically speaking. 971 972 // If we have a problem phi translating, fall through to the code below 973 // to handle the failure condition. 974 if (getNonLocalPointerDepFromBB(PredPointer, 975 Loc.getWithNewPtr(PredPointer.getAddr()), 976 isLoad, Pred, 977 Result, Visited)) 978 goto PredTranslationFailure; 979 } 980 981 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 982 CacheInfo = &NonLocalPointerDeps[CacheKey]; 983 Cache = &CacheInfo->NonLocalDeps; 984 NumSortedEntries = Cache->size(); 985 986 // Since we did phi translation, the "Cache" set won't contain all of the 987 // results for the query. This is ok (we can still use it to accelerate 988 // specific block queries) but we can't do the fastpath "return all 989 // results from the set" Clear out the indicator for this. 990 CacheInfo->Pair = BBSkipFirstBlockPair(); 991 CacheInfo->Size = 0; 992 CacheInfo->TBAATag = 0; 993 SkipFirstBlock = false; 994 continue; 995 996 PredTranslationFailure: 997 998 if (Cache == 0) { 999 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1000 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1001 Cache = &CacheInfo->NonLocalDeps; 1002 NumSortedEntries = Cache->size(); 1003 } 1004 1005 // Since we failed phi translation, the "Cache" set won't contain all of the 1006 // results for the query. This is ok (we can still use it to accelerate 1007 // specific block queries) but we can't do the fastpath "return all 1008 // results from the set". Clear out the indicator for this. 1009 CacheInfo->Pair = BBSkipFirstBlockPair(); 1010 CacheInfo->Size = 0; 1011 CacheInfo->TBAATag = 0; 1012 1013 // If *nothing* works, mark the pointer as being clobbered by the first 1014 // instruction in this block. 1015 // 1016 // If this is the magic first block, return this as a clobber of the whole 1017 // incoming value. Since we can't phi translate to one of the predecessors, 1018 // we have to bail out. 1019 if (SkipFirstBlock) 1020 return true; 1021 1022 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 1023 assert(I != Cache->rend() && "Didn't find current block??"); 1024 if (I->getBB() != BB) 1025 continue; 1026 1027 assert(I->getResult().isNonLocal() && 1028 "Should only be here with transparent block"); 1029 I->setResult(MemDepResult::getClobber(BB->begin())); 1030 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey); 1031 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 1032 Pointer.getAddr())); 1033 break; 1034 } 1035 } 1036 1037 // Okay, we're done now. If we added new values to the cache, re-sort it. 1038 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1039 DEBUG(AssertSorted(*Cache)); 1040 return false; 1041} 1042 1043/// RemoveCachedNonLocalPointerDependencies - If P exists in 1044/// CachedNonLocalPointerInfo, remove it. 1045void MemoryDependenceAnalysis:: 1046RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 1047 CachedNonLocalPointerInfo::iterator It = 1048 NonLocalPointerDeps.find(P); 1049 if (It == NonLocalPointerDeps.end()) return; 1050 1051 // Remove all of the entries in the BB->val map. This involves removing 1052 // instructions from the reverse map. 1053 NonLocalDepInfo &PInfo = It->second.NonLocalDeps; 1054 1055 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 1056 Instruction *Target = PInfo[i].getResult().getInst(); 1057 if (Target == 0) continue; // Ignore non-local dep results. 1058 assert(Target->getParent() == PInfo[i].getBB()); 1059 1060 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1061 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1062 } 1063 1064 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1065 NonLocalPointerDeps.erase(It); 1066} 1067 1068 1069/// invalidateCachedPointerInfo - This method is used to invalidate cached 1070/// information about the specified pointer, because it may be too 1071/// conservative in memdep. This is an optional call that can be used when 1072/// the client detects an equivalence between the pointer and some other 1073/// value and replaces the other value with ptr. This can make Ptr available 1074/// in more places that cached info does not necessarily keep. 1075void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1076 // If Ptr isn't really a pointer, just ignore it. 1077 if (!Ptr->getType()->isPointerTy()) return; 1078 // Flush store info for the pointer. 1079 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1080 // Flush load info for the pointer. 1081 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1082} 1083 1084/// invalidateCachedPredecessors - Clear the PredIteratorCache info. 1085/// This needs to be done when the CFG changes, e.g., due to splitting 1086/// critical edges. 1087void MemoryDependenceAnalysis::invalidateCachedPredecessors() { 1088 PredCache->clear(); 1089} 1090 1091/// removeInstruction - Remove an instruction from the dependence analysis, 1092/// updating the dependence of instructions that previously depended on it. 1093/// This method attempts to keep the cache coherent using the reverse map. 1094void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1095 // Walk through the Non-local dependencies, removing this one as the value 1096 // for any cached queries. 1097 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1098 if (NLDI != NonLocalDeps.end()) { 1099 NonLocalDepInfo &BlockMap = NLDI->second.first; 1100 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1101 DI != DE; ++DI) 1102 if (Instruction *Inst = DI->getResult().getInst()) 1103 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1104 NonLocalDeps.erase(NLDI); 1105 } 1106 1107 // If we have a cached local dependence query for this instruction, remove it. 1108 // 1109 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1110 if (LocalDepEntry != LocalDeps.end()) { 1111 // Remove us from DepInst's reverse set now that the local dep info is gone. 1112 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1113 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1114 1115 // Remove this local dependency info. 1116 LocalDeps.erase(LocalDepEntry); 1117 } 1118 1119 // If we have any cached pointer dependencies on this instruction, remove 1120 // them. If the instruction has non-pointer type, then it can't be a pointer 1121 // base. 1122 1123 // Remove it from both the load info and the store info. The instruction 1124 // can't be in either of these maps if it is non-pointer. 1125 if (RemInst->getType()->isPointerTy()) { 1126 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1127 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1128 } 1129 1130 // Loop over all of the things that depend on the instruction we're removing. 1131 // 1132 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1133 1134 // If we find RemInst as a clobber or Def in any of the maps for other values, 1135 // we need to replace its entry with a dirty version of the instruction after 1136 // it. If RemInst is a terminator, we use a null dirty value. 1137 // 1138 // Using a dirty version of the instruction after RemInst saves having to scan 1139 // the entire block to get to this point. 1140 MemDepResult NewDirtyVal; 1141 if (!RemInst->isTerminator()) 1142 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1143 1144 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1145 if (ReverseDepIt != ReverseLocalDeps.end()) { 1146 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1147 // RemInst can't be the terminator if it has local stuff depending on it. 1148 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1149 "Nothing can locally depend on a terminator"); 1150 1151 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1152 E = ReverseDeps.end(); I != E; ++I) { 1153 Instruction *InstDependingOnRemInst = *I; 1154 assert(InstDependingOnRemInst != RemInst && 1155 "Already removed our local dep info"); 1156 1157 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1158 1159 // Make sure to remember that new things depend on NewDepInst. 1160 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1161 "a local dep on this if it is a terminator!"); 1162 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1163 InstDependingOnRemInst)); 1164 } 1165 1166 ReverseLocalDeps.erase(ReverseDepIt); 1167 1168 // Add new reverse deps after scanning the set, to avoid invalidating the 1169 // 'ReverseDeps' reference. 1170 while (!ReverseDepsToAdd.empty()) { 1171 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1172 .insert(ReverseDepsToAdd.back().second); 1173 ReverseDepsToAdd.pop_back(); 1174 } 1175 } 1176 1177 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1178 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1179 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1180 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1181 I != E; ++I) { 1182 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1183 1184 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1185 // The information is now dirty! 1186 INLD.second = true; 1187 1188 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1189 DE = INLD.first.end(); DI != DE; ++DI) { 1190 if (DI->getResult().getInst() != RemInst) continue; 1191 1192 // Convert to a dirty entry for the subsequent instruction. 1193 DI->setResult(NewDirtyVal); 1194 1195 if (Instruction *NextI = NewDirtyVal.getInst()) 1196 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1197 } 1198 } 1199 1200 ReverseNonLocalDeps.erase(ReverseDepIt); 1201 1202 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1203 while (!ReverseDepsToAdd.empty()) { 1204 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1205 .insert(ReverseDepsToAdd.back().second); 1206 ReverseDepsToAdd.pop_back(); 1207 } 1208 } 1209 1210 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1211 // value in the NonLocalPointerDeps info. 1212 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1213 ReverseNonLocalPtrDeps.find(RemInst); 1214 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1215 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1216 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1217 1218 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1219 E = Set.end(); I != E; ++I) { 1220 ValueIsLoadPair P = *I; 1221 assert(P.getPointer() != RemInst && 1222 "Already removed NonLocalPointerDeps info for RemInst"); 1223 1224 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps; 1225 1226 // The cache is not valid for any specific block anymore. 1227 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair(); 1228 NonLocalPointerDeps[P].Size = 0; 1229 NonLocalPointerDeps[P].TBAATag = 0; 1230 1231 // Update any entries for RemInst to use the instruction after it. 1232 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1233 DI != DE; ++DI) { 1234 if (DI->getResult().getInst() != RemInst) continue; 1235 1236 // Convert to a dirty entry for the subsequent instruction. 1237 DI->setResult(NewDirtyVal); 1238 1239 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1240 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1241 } 1242 1243 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1244 // subsequent value may invalidate the sortedness. 1245 std::sort(NLPDI.begin(), NLPDI.end()); 1246 } 1247 1248 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1249 1250 while (!ReversePtrDepsToAdd.empty()) { 1251 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1252 .insert(ReversePtrDepsToAdd.back().second); 1253 ReversePtrDepsToAdd.pop_back(); 1254 } 1255 } 1256 1257 1258 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1259 AA->deleteValue(RemInst); 1260 DEBUG(verifyRemoved(RemInst)); 1261} 1262/// verifyRemoved - Verify that the specified instruction does not occur 1263/// in our internal data structures. 1264void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1265 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1266 E = LocalDeps.end(); I != E; ++I) { 1267 assert(I->first != D && "Inst occurs in data structures"); 1268 assert(I->second.getInst() != D && 1269 "Inst occurs in data structures"); 1270 } 1271 1272 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1273 E = NonLocalPointerDeps.end(); I != E; ++I) { 1274 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1275 const NonLocalDepInfo &Val = I->second.NonLocalDeps; 1276 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1277 II != E; ++II) 1278 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1279 } 1280 1281 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1282 E = NonLocalDeps.end(); I != E; ++I) { 1283 assert(I->first != D && "Inst occurs in data structures"); 1284 const PerInstNLInfo &INLD = I->second; 1285 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1286 EE = INLD.first.end(); II != EE; ++II) 1287 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1288 } 1289 1290 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1291 E = ReverseLocalDeps.end(); I != E; ++I) { 1292 assert(I->first != D && "Inst occurs in data structures"); 1293 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1294 EE = I->second.end(); II != EE; ++II) 1295 assert(*II != D && "Inst occurs in data structures"); 1296 } 1297 1298 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1299 E = ReverseNonLocalDeps.end(); 1300 I != E; ++I) { 1301 assert(I->first != D && "Inst occurs in data structures"); 1302 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1303 EE = I->second.end(); II != EE; ++II) 1304 assert(*II != D && "Inst occurs in data structures"); 1305 } 1306 1307 for (ReverseNonLocalPtrDepTy::const_iterator 1308 I = ReverseNonLocalPtrDeps.begin(), 1309 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1310 assert(I->first != D && "Inst occurs in rev NLPD map"); 1311 1312 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1313 E = I->second.end(); II != E; ++II) 1314 assert(*II != ValueIsLoadPair(D, false) && 1315 *II != ValueIsLoadPair(D, true) && 1316 "Inst occurs in ReverseNonLocalPtrDeps map"); 1317 } 1318 1319} 1320