MemoryDependenceAnalysis.cpp revision f6f1f062cc8029aa75ca7d0e99fbc1e0b453d07e
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/// GetLocation - If the given instruction references a specific memory 106/// location, fill in Loc with the details, otherwise set Loc.Ptr to null. 107/// Return a ModRefInfo value describing the general behavior of the 108/// instruction. 109static 110AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst, 111 AliasAnalysis::Location &Loc, 112 AliasAnalysis *AA) { 113 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 114 if (LI->isVolatile()) { 115 Loc = AliasAnalysis::Location(); 116 return AliasAnalysis::ModRef; 117 } 118 Loc = AA->getLocation(LI); 119 return AliasAnalysis::Ref; 120 } 121 122 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 123 if (SI->isVolatile()) { 124 Loc = AliasAnalysis::Location(); 125 return AliasAnalysis::ModRef; 126 } 127 Loc = AA->getLocation(SI); 128 return AliasAnalysis::Mod; 129 } 130 131 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 132 Loc = AA->getLocation(V); 133 return AliasAnalysis::ModRef; 134 } 135 136 if (const CallInst *CI = isFreeCall(Inst)) { 137 // calls to free() deallocate the entire structure 138 Loc = AliasAnalysis::Location(CI->getArgOperand(0)); 139 return AliasAnalysis::Mod; 140 } 141 142 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) 143 switch (II->getIntrinsicID()) { 144 case Intrinsic::lifetime_start: 145 case Intrinsic::lifetime_end: 146 case Intrinsic::invariant_start: 147 Loc = AliasAnalysis::Location(II->getArgOperand(1), 148 cast<ConstantInt>(II->getArgOperand(0)) 149 ->getZExtValue(), 150 II->getMetadata(LLVMContext::MD_tbaa)); 151 // These intrinsics don't really modify the memory, but returning Mod 152 // will allow them to be handled conservatively. 153 return AliasAnalysis::Mod; 154 case Intrinsic::invariant_end: 155 Loc = AliasAnalysis::Location(II->getArgOperand(2), 156 cast<ConstantInt>(II->getArgOperand(1)) 157 ->getZExtValue(), 158 II->getMetadata(LLVMContext::MD_tbaa)); 159 // These intrinsics don't really modify the memory, but returning Mod 160 // will allow them to be handled conservatively. 161 return AliasAnalysis::Mod; 162 default: 163 break; 164 } 165 166 // Otherwise, just do the coarse-grained thing that always works. 167 if (Inst->mayWriteToMemory()) 168 return AliasAnalysis::ModRef; 169 if (Inst->mayReadFromMemory()) 170 return AliasAnalysis::Ref; 171 return AliasAnalysis::NoModRef; 172} 173 174/// getCallSiteDependencyFrom - Private helper for finding the local 175/// dependencies of a call site. 176MemDepResult MemoryDependenceAnalysis:: 177getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 178 BasicBlock::iterator ScanIt, BasicBlock *BB) { 179 // Walk backwards through the block, looking for dependencies 180 while (ScanIt != BB->begin()) { 181 Instruction *Inst = --ScanIt; 182 183 // If this inst is a memory op, get the pointer it accessed 184 AliasAnalysis::Location Loc; 185 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA); 186 if (Loc.Ptr) { 187 // A simple instruction. 188 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef) 189 return MemDepResult::getClobber(Inst); 190 continue; 191 } 192 193 if (CallSite InstCS = cast<Value>(Inst)) { 194 // Debug intrinsics don't cause dependences. 195 if (isa<DbgInfoIntrinsic>(Inst)) continue; 196 // If these two calls do not interfere, look past it. 197 switch (AA->getModRefInfo(CS, InstCS)) { 198 case AliasAnalysis::NoModRef: 199 // If the two calls are the same, return InstCS as a Def, so that 200 // CS can be found redundant and eliminated. 201 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) && 202 CS.getInstruction()->isIdenticalToWhenDefined(Inst)) 203 return MemDepResult::getDef(Inst); 204 205 // Otherwise if the two calls don't interact (e.g. InstCS is readnone) 206 // keep scanning. 207 break; 208 default: 209 return MemDepResult::getClobber(Inst); 210 } 211 } 212 } 213 214 // No dependence found. If this is the entry block of the function, it is a 215 // clobber, otherwise it is non-local. 216 if (BB != &BB->getParent()->getEntryBlock()) 217 return MemDepResult::getNonLocal(); 218 return MemDepResult::getClobber(ScanIt); 219} 220 221/// getPointerDependencyFrom - Return the instruction on which a memory 222/// location depends. If isLoad is true, this routine ignores may-aliases with 223/// read-only operations. If isLoad is false, this routine ignores may-aliases 224/// with reads from read-only locations. 225MemDepResult MemoryDependenceAnalysis:: 226getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad, 227 BasicBlock::iterator ScanIt, BasicBlock *BB) { 228 229 Value *InvariantTag = 0; 230 231 // Walk backwards through the basic block, looking for dependencies. 232 while (ScanIt != BB->begin()) { 233 Instruction *Inst = --ScanIt; 234 235 // If we're in an invariant region, no dependencies can be found before 236 // we pass an invariant-begin marker. 237 if (InvariantTag == Inst) { 238 InvariantTag = 0; 239 continue; 240 } 241 242 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 243 // Debug intrinsics don't (and can't) cause dependences. 244 if (isa<DbgInfoIntrinsic>(II)) continue; 245 246 // If we pass an invariant-end marker, then we've just entered an 247 // invariant region and can start ignoring dependencies. 248 if (II->getIntrinsicID() == Intrinsic::invariant_end) { 249 // FIXME: This only considers queries directly on the invariant-tagged 250 // pointer, not on query pointers that are indexed off of them. It'd 251 // be nice to handle that at some point. 252 AliasAnalysis::AliasResult R = 253 AA->alias(AliasAnalysis::Location(II->getArgOperand(2)), MemLoc); 254 if (R == AliasAnalysis::MustAlias) 255 InvariantTag = II->getArgOperand(0); 256 257 continue; 258 } 259 260 // If we reach a lifetime begin or end marker, then the query ends here 261 // because the value is undefined. 262 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 263 // FIXME: This only considers queries directly on the invariant-tagged 264 // pointer, not on query pointers that are indexed off of them. It'd 265 // be nice to handle that at some point. 266 AliasAnalysis::AliasResult R = 267 AA->alias(AliasAnalysis::Location(II->getArgOperand(1)), MemLoc); 268 if (R == AliasAnalysis::MustAlias) 269 return MemDepResult::getDef(II); 270 continue; 271 } 272 } 273 274 // If we're querying on a load and we're in an invariant region, we're done 275 // at this point. Nothing a load depends on can live in an invariant region. 276 // 277 // FIXME: this will prevent us from returning load/load must-aliases, so GVN 278 // won't remove redundant loads. 279 if (isLoad && InvariantTag) continue; 280 281 // Values depend on loads if the pointers are must aliased. This means that 282 // a load depends on another must aliased load from the same value. 283 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 284 AliasAnalysis::Location LoadLoc = AA->getLocation(LI); 285 286 // If we found a pointer, check if it could be the same as our pointer. 287 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc); 288 if (R == AliasAnalysis::NoAlias) 289 continue; 290 291 // May-alias loads don't depend on each other without a dependence. 292 if (isLoad && R == AliasAnalysis::MayAlias) 293 continue; 294 295 // Stores don't alias loads from read-only memory. 296 if (!isLoad && AA->pointsToConstantMemory(LoadLoc)) 297 continue; 298 299 // Stores depend on may and must aliased loads, loads depend on must-alias 300 // loads. 301 return MemDepResult::getDef(Inst); 302 } 303 304 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 305 // There can't be stores to the value we care about inside an 306 // invariant region. 307 if (InvariantTag) continue; 308 309 // If alias analysis can tell that this store is guaranteed to not modify 310 // the query pointer, ignore it. Use getModRefInfo to handle cases where 311 // the query pointer points to constant memory etc. 312 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef) 313 continue; 314 315 // Ok, this store might clobber the query pointer. Check to see if it is 316 // a must alias: in this case, we want to return this as a def. 317 AliasAnalysis::Location StoreLoc = AA->getLocation(SI); 318 319 // If we found a pointer, check if it could be the same as our pointer. 320 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc); 321 322 if (R == AliasAnalysis::NoAlias) 323 continue; 324 if (R == AliasAnalysis::MayAlias) 325 return MemDepResult::getClobber(Inst); 326 return MemDepResult::getDef(Inst); 327 } 328 329 // If this is an allocation, and if we know that the accessed pointer is to 330 // the allocation, return Def. This means that there is no dependence and 331 // the access can be optimized based on that. For example, a load could 332 // turn into undef. 333 // Note: Only determine this to be a malloc if Inst is the malloc call, not 334 // a subsequent bitcast of the malloc call result. There can be stores to 335 // the malloced memory between the malloc call and its bitcast uses, and we 336 // need to continue scanning until the malloc call. 337 if (isa<AllocaInst>(Inst) || 338 (isa<CallInst>(Inst) && extractMallocCall(Inst))) { 339 const Value *AccessPtr = MemLoc.Ptr->getUnderlyingObject(); 340 341 if (AccessPtr == Inst || 342 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 343 return MemDepResult::getDef(Inst); 344 continue; 345 } 346 347 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 348 switch (AA->getModRefInfo(Inst, MemLoc)) { 349 case AliasAnalysis::NoModRef: 350 // If the call has no effect on the queried pointer, just ignore it. 351 continue; 352 case AliasAnalysis::Mod: 353 // If we're in an invariant region, we can ignore calls that ONLY 354 // modify the pointer. 355 if (InvariantTag) continue; 356 return MemDepResult::getClobber(Inst); 357 case AliasAnalysis::Ref: 358 // If the call is known to never store to the pointer, and if this is a 359 // load query, we can safely ignore it (scan past it). 360 if (isLoad) 361 continue; 362 default: 363 // Otherwise, there is a potential dependence. Return a clobber. 364 return MemDepResult::getClobber(Inst); 365 } 366 } 367 368 // No dependence found. If this is the entry block of the function, it is a 369 // clobber, otherwise it is non-local. 370 if (BB != &BB->getParent()->getEntryBlock()) 371 return MemDepResult::getNonLocal(); 372 return MemDepResult::getClobber(ScanIt); 373} 374 375/// getDependency - Return the instruction on which a memory operation 376/// depends. 377MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 378 Instruction *ScanPos = QueryInst; 379 380 // Check for a cached result 381 MemDepResult &LocalCache = LocalDeps[QueryInst]; 382 383 // If the cached entry is non-dirty, just return it. Note that this depends 384 // on MemDepResult's default constructing to 'dirty'. 385 if (!LocalCache.isDirty()) 386 return LocalCache; 387 388 // Otherwise, if we have a dirty entry, we know we can start the scan at that 389 // instruction, which may save us some work. 390 if (Instruction *Inst = LocalCache.getInst()) { 391 ScanPos = Inst; 392 393 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 394 } 395 396 BasicBlock *QueryParent = QueryInst->getParent(); 397 398 // Do the scan. 399 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 400 // No dependence found. If this is the entry block of the function, it is a 401 // clobber, otherwise it is non-local. 402 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 403 LocalCache = MemDepResult::getNonLocal(); 404 else 405 LocalCache = MemDepResult::getClobber(QueryInst); 406 } else { 407 AliasAnalysis::Location MemLoc; 408 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA); 409 if (MemLoc.Ptr) { 410 // If we can do a pointer scan, make it happen. 411 bool isLoad = !(MR & AliasAnalysis::Mod); 412 if (IntrinsicInst *II = dyn_cast<MemoryUseIntrinsic>(QueryInst)) 413 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end; 414 415 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos, 416 QueryParent); 417 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 418 CallSite QueryCS(QueryInst); 419 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 420 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 421 QueryParent); 422 } else 423 // Non-memory instruction. 424 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 425 } 426 427 // Remember the result! 428 if (Instruction *I = LocalCache.getInst()) 429 ReverseLocalDeps[I].insert(QueryInst); 430 431 return LocalCache; 432} 433 434#ifndef NDEBUG 435/// AssertSorted - This method is used when -debug is specified to verify that 436/// cache arrays are properly kept sorted. 437static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 438 int Count = -1) { 439 if (Count == -1) Count = Cache.size(); 440 if (Count == 0) return; 441 442 for (unsigned i = 1; i != unsigned(Count); ++i) 443 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 444} 445#endif 446 447/// getNonLocalCallDependency - Perform a full dependency query for the 448/// specified call, returning the set of blocks that the value is 449/// potentially live across. The returned set of results will include a 450/// "NonLocal" result for all blocks where the value is live across. 451/// 452/// This method assumes the instruction returns a "NonLocal" dependency 453/// within its own block. 454/// 455/// This returns a reference to an internal data structure that may be 456/// invalidated on the next non-local query or when an instruction is 457/// removed. Clients must copy this data if they want it around longer than 458/// that. 459const MemoryDependenceAnalysis::NonLocalDepInfo & 460MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 461 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 462 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 463 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 464 NonLocalDepInfo &Cache = CacheP.first; 465 466 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 467 /// the cached case, this can happen due to instructions being deleted etc. In 468 /// the uncached case, this starts out as the set of predecessors we care 469 /// about. 470 SmallVector<BasicBlock*, 32> DirtyBlocks; 471 472 if (!Cache.empty()) { 473 // Okay, we have a cache entry. If we know it is not dirty, just return it 474 // with no computation. 475 if (!CacheP.second) { 476 ++NumCacheNonLocal; 477 return Cache; 478 } 479 480 // If we already have a partially computed set of results, scan them to 481 // determine what is dirty, seeding our initial DirtyBlocks worklist. 482 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 483 I != E; ++I) 484 if (I->getResult().isDirty()) 485 DirtyBlocks.push_back(I->getBB()); 486 487 // Sort the cache so that we can do fast binary search lookups below. 488 std::sort(Cache.begin(), Cache.end()); 489 490 ++NumCacheDirtyNonLocal; 491 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 492 // << Cache.size() << " cached: " << *QueryInst; 493 } else { 494 // Seed DirtyBlocks with each of the preds of QueryInst's block. 495 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 496 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 497 DirtyBlocks.push_back(*PI); 498 ++NumUncacheNonLocal; 499 } 500 501 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 502 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 503 504 SmallPtrSet<BasicBlock*, 64> Visited; 505 506 unsigned NumSortedEntries = Cache.size(); 507 DEBUG(AssertSorted(Cache)); 508 509 // Iterate while we still have blocks to update. 510 while (!DirtyBlocks.empty()) { 511 BasicBlock *DirtyBB = DirtyBlocks.back(); 512 DirtyBlocks.pop_back(); 513 514 // Already processed this block? 515 if (!Visited.insert(DirtyBB)) 516 continue; 517 518 // Do a binary search to see if we already have an entry for this block in 519 // the cache set. If so, find it. 520 DEBUG(AssertSorted(Cache, NumSortedEntries)); 521 NonLocalDepInfo::iterator Entry = 522 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 523 NonLocalDepEntry(DirtyBB)); 524 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB) 525 --Entry; 526 527 NonLocalDepEntry *ExistingResult = 0; 528 if (Entry != Cache.begin()+NumSortedEntries && 529 Entry->getBB() == DirtyBB) { 530 // If we already have an entry, and if it isn't already dirty, the block 531 // is done. 532 if (!Entry->getResult().isDirty()) 533 continue; 534 535 // Otherwise, remember this slot so we can update the value. 536 ExistingResult = &*Entry; 537 } 538 539 // If the dirty entry has a pointer, start scanning from it so we don't have 540 // to rescan the entire block. 541 BasicBlock::iterator ScanPos = DirtyBB->end(); 542 if (ExistingResult) { 543 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 544 ScanPos = Inst; 545 // We're removing QueryInst's use of Inst. 546 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 547 QueryCS.getInstruction()); 548 } 549 } 550 551 // Find out if this block has a local dependency for QueryInst. 552 MemDepResult Dep; 553 554 if (ScanPos != DirtyBB->begin()) { 555 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 556 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 557 // No dependence found. If this is the entry block of the function, it is 558 // a clobber, otherwise it is non-local. 559 Dep = MemDepResult::getNonLocal(); 560 } else { 561 Dep = MemDepResult::getClobber(ScanPos); 562 } 563 564 // If we had a dirty entry for the block, update it. Otherwise, just add 565 // a new entry. 566 if (ExistingResult) 567 ExistingResult->setResult(Dep); 568 else 569 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 570 571 // If the block has a dependency (i.e. it isn't completely transparent to 572 // the value), remember the association! 573 if (!Dep.isNonLocal()) { 574 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 575 // update this when we remove instructions. 576 if (Instruction *Inst = Dep.getInst()) 577 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 578 } else { 579 580 // If the block *is* completely transparent to the load, we need to check 581 // the predecessors of this block. Add them to our worklist. 582 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 583 DirtyBlocks.push_back(*PI); 584 } 585 } 586 587 return Cache; 588} 589 590/// getNonLocalPointerDependency - Perform a full dependency query for an 591/// access to the specified (non-volatile) memory location, returning the 592/// set of instructions that either define or clobber the value. 593/// 594/// This method assumes the pointer has a "NonLocal" dependency within its 595/// own block. 596/// 597void MemoryDependenceAnalysis:: 598getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad, 599 BasicBlock *FromBB, 600 SmallVectorImpl<NonLocalDepResult> &Result) { 601 assert(Loc.Ptr->getType()->isPointerTy() && 602 "Can't get pointer deps of a non-pointer!"); 603 Result.clear(); 604 605 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD); 606 607 // This is the set of blocks we've inspected, and the pointer we consider in 608 // each block. Because of critical edges, we currently bail out if querying 609 // a block with multiple different pointers. This can happen during PHI 610 // translation. 611 DenseMap<BasicBlock*, Value*> Visited; 612 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB, 613 Result, Visited, true)) 614 return; 615 Result.clear(); 616 Result.push_back(NonLocalDepResult(FromBB, 617 MemDepResult::getClobber(FromBB->begin()), 618 const_cast<Value *>(Loc.Ptr))); 619} 620 621/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 622/// Pointer/PointeeSize using either cached information in Cache or by doing a 623/// lookup (which may use dirty cache info if available). If we do a lookup, 624/// add the result to the cache. 625MemDepResult MemoryDependenceAnalysis:: 626GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc, 627 bool isLoad, BasicBlock *BB, 628 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 629 630 // Do a binary search to see if we already have an entry for this block in 631 // the cache set. If so, find it. 632 NonLocalDepInfo::iterator Entry = 633 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 634 NonLocalDepEntry(BB)); 635 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 636 --Entry; 637 638 NonLocalDepEntry *ExistingResult = 0; 639 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 640 ExistingResult = &*Entry; 641 642 // If we have a cached entry, and it is non-dirty, use it as the value for 643 // this dependency. 644 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 645 ++NumCacheNonLocalPtr; 646 return ExistingResult->getResult(); 647 } 648 649 // Otherwise, we have to scan for the value. If we have a dirty cache 650 // entry, start scanning from its position, otherwise we scan from the end 651 // of the block. 652 BasicBlock::iterator ScanPos = BB->end(); 653 if (ExistingResult && ExistingResult->getResult().getInst()) { 654 assert(ExistingResult->getResult().getInst()->getParent() == BB && 655 "Instruction invalidated?"); 656 ++NumCacheDirtyNonLocalPtr; 657 ScanPos = ExistingResult->getResult().getInst(); 658 659 // Eliminating the dirty entry from 'Cache', so update the reverse info. 660 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 661 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 662 } else { 663 ++NumUncacheNonLocalPtr; 664 } 665 666 // Scan the block for the dependency. 667 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB); 668 669 // If we had a dirty entry for the block, update it. Otherwise, just add 670 // a new entry. 671 if (ExistingResult) 672 ExistingResult->setResult(Dep); 673 else 674 Cache->push_back(NonLocalDepEntry(BB, Dep)); 675 676 // If the block has a dependency (i.e. it isn't completely transparent to 677 // the value), remember the reverse association because we just added it 678 // to Cache! 679 if (Dep.isNonLocal()) 680 return Dep; 681 682 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 683 // update MemDep when we remove instructions. 684 Instruction *Inst = Dep.getInst(); 685 assert(Inst && "Didn't depend on anything?"); 686 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 687 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 688 return Dep; 689} 690 691/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 692/// number of elements in the array that are already properly ordered. This is 693/// optimized for the case when only a few entries are added. 694static void 695SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 696 unsigned NumSortedEntries) { 697 switch (Cache.size() - NumSortedEntries) { 698 case 0: 699 // done, no new entries. 700 break; 701 case 2: { 702 // Two new entries, insert the last one into place. 703 NonLocalDepEntry Val = Cache.back(); 704 Cache.pop_back(); 705 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 706 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 707 Cache.insert(Entry, Val); 708 // FALL THROUGH. 709 } 710 case 1: 711 // One new entry, Just insert the new value at the appropriate position. 712 if (Cache.size() != 1) { 713 NonLocalDepEntry Val = Cache.back(); 714 Cache.pop_back(); 715 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 716 std::upper_bound(Cache.begin(), Cache.end(), Val); 717 Cache.insert(Entry, Val); 718 } 719 break; 720 default: 721 // Added many values, do a full scale sort. 722 std::sort(Cache.begin(), Cache.end()); 723 break; 724 } 725} 726 727/// getNonLocalPointerDepFromBB - Perform a dependency query based on 728/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 729/// results to the results vector and keep track of which blocks are visited in 730/// 'Visited'. 731/// 732/// This has special behavior for the first block queries (when SkipFirstBlock 733/// is true). In this special case, it ignores the contents of the specified 734/// block and starts returning dependence info for its predecessors. 735/// 736/// This function returns false on success, or true to indicate that it could 737/// not compute dependence information for some reason. This should be treated 738/// as a clobber dependence on the first instruction in the predecessor block. 739bool MemoryDependenceAnalysis:: 740getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, 741 const AliasAnalysis::Location &Loc, 742 bool isLoad, BasicBlock *StartBB, 743 SmallVectorImpl<NonLocalDepResult> &Result, 744 DenseMap<BasicBlock*, Value*> &Visited, 745 bool SkipFirstBlock) { 746 747 // Look up the cached info for Pointer. 748 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 749 750 // Set up a temporary NLPI value. If the map doesn't yet have an entry for 751 // CacheKey, this value will be inserted as the associated value. Otherwise, 752 // it'll be ignored, and we'll have to check to see if the cached size and 753 // tbaa tag are consistent with the current query. 754 NonLocalPointerInfo InitialNLPI; 755 InitialNLPI.Size = Loc.Size; 756 InitialNLPI.TBAATag = Loc.TBAATag; 757 758 // Get the NLPI for CacheKey, inserting one into the map if it doesn't 759 // already have one. 760 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair = 761 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI)); 762 NonLocalPointerInfo *CacheInfo = &Pair.first->second; 763 764 // If we already have a cache entry for this CacheKey, we may need to do some 765 // work to reconcile the cache entry and the current query. 766 if (!Pair.second) { 767 if (CacheInfo->Size < Loc.Size) { 768 // The query's Size is greater than the cached one. Throw out the 769 // cached data and procede with the query at the greater size. 770 CacheInfo->Pair = BBSkipFirstBlockPair(); 771 CacheInfo->Size = Loc.Size; 772 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 773 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 774 if (Instruction *Inst = DI->getResult().getInst()) 775 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 776 CacheInfo->NonLocalDeps.clear(); 777 } else if (CacheInfo->Size > Loc.Size) { 778 // This query's Size is less than the cached one. Conservatively restart 779 // the query using the greater size. 780 return getNonLocalPointerDepFromBB(Pointer, 781 Loc.getWithNewSize(CacheInfo->Size), 782 isLoad, StartBB, Result, Visited, 783 SkipFirstBlock); 784 } 785 786 // If the query's TBAATag is inconsistent with the cached one, 787 // conservatively throw out the cached data and restart the query with 788 // no tag if needed. 789 if (CacheInfo->TBAATag != Loc.TBAATag) { 790 if (CacheInfo->TBAATag) { 791 CacheInfo->Pair = BBSkipFirstBlockPair(); 792 CacheInfo->TBAATag = 0; 793 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 794 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 795 if (Instruction *Inst = DI->getResult().getInst()) 796 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 797 CacheInfo->NonLocalDeps.clear(); 798 } 799 if (Loc.TBAATag) 800 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(), 801 isLoad, StartBB, Result, Visited, 802 SkipFirstBlock); 803 } 804 } 805 806 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps; 807 808 // If we have valid cached information for exactly the block we are 809 // investigating, just return it with no recomputation. 810 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 811 // We have a fully cached result for this query then we can just return the 812 // cached results and populate the visited set. However, we have to verify 813 // that we don't already have conflicting results for these blocks. Check 814 // to ensure that if a block in the results set is in the visited set that 815 // it was for the same pointer query. 816 if (!Visited.empty()) { 817 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 818 I != E; ++I) { 819 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 820 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 821 continue; 822 823 // We have a pointer mismatch in a block. Just return clobber, saying 824 // that something was clobbered in this result. We could also do a 825 // non-fully cached query, but there is little point in doing this. 826 return true; 827 } 828 } 829 830 Value *Addr = Pointer.getAddr(); 831 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 832 I != E; ++I) { 833 Visited.insert(std::make_pair(I->getBB(), Addr)); 834 if (!I->getResult().isNonLocal()) 835 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 836 } 837 ++NumCacheCompleteNonLocalPtr; 838 return false; 839 } 840 841 // Otherwise, either this is a new block, a block with an invalid cache 842 // pointer or one that we're about to invalidate by putting more info into it 843 // than its valid cache info. If empty, the result will be valid cache info, 844 // otherwise it isn't. 845 if (Cache->empty()) 846 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 847 else 848 CacheInfo->Pair = BBSkipFirstBlockPair(); 849 850 SmallVector<BasicBlock*, 32> Worklist; 851 Worklist.push_back(StartBB); 852 853 // Keep track of the entries that we know are sorted. Previously cached 854 // entries will all be sorted. The entries we add we only sort on demand (we 855 // don't insert every element into its sorted position). We know that we 856 // won't get any reuse from currently inserted values, because we don't 857 // revisit blocks after we insert info for them. 858 unsigned NumSortedEntries = Cache->size(); 859 DEBUG(AssertSorted(*Cache)); 860 861 while (!Worklist.empty()) { 862 BasicBlock *BB = Worklist.pop_back_val(); 863 864 // Skip the first block if we have it. 865 if (!SkipFirstBlock) { 866 // Analyze the dependency of *Pointer in FromBB. See if we already have 867 // been here. 868 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 869 870 // Get the dependency info for Pointer in BB. If we have cached 871 // information, we will use it, otherwise we compute it. 872 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 873 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache, 874 NumSortedEntries); 875 876 // If we got a Def or Clobber, add this to the list of results. 877 if (!Dep.isNonLocal()) { 878 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 879 continue; 880 } 881 } 882 883 // If 'Pointer' is an instruction defined in this block, then we need to do 884 // phi translation to change it into a value live in the predecessor block. 885 // If not, we just add the predecessors to the worklist and scan them with 886 // the same Pointer. 887 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 888 SkipFirstBlock = false; 889 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 890 // Verify that we haven't looked at this block yet. 891 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 892 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 893 if (InsertRes.second) { 894 // First time we've looked at *PI. 895 Worklist.push_back(*PI); 896 continue; 897 } 898 899 // If we have seen this block before, but it was with a different 900 // pointer then we have a phi translation failure and we have to treat 901 // this as a clobber. 902 if (InsertRes.first->second != Pointer.getAddr()) 903 goto PredTranslationFailure; 904 } 905 continue; 906 } 907 908 // We do need to do phi translation, if we know ahead of time we can't phi 909 // translate this value, don't even try. 910 if (!Pointer.IsPotentiallyPHITranslatable()) 911 goto PredTranslationFailure; 912 913 // We may have added values to the cache list before this PHI translation. 914 // If so, we haven't done anything to ensure that the cache remains sorted. 915 // Sort it now (if needed) so that recursive invocations of 916 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 917 // value will only see properly sorted cache arrays. 918 if (Cache && NumSortedEntries != Cache->size()) { 919 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 920 NumSortedEntries = Cache->size(); 921 } 922 Cache = 0; 923 924 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 925 BasicBlock *Pred = *PI; 926 927 // Get the PHI translated pointer in this predecessor. This can fail if 928 // not translatable, in which case the getAddr() returns null. 929 PHITransAddr PredPointer(Pointer); 930 PredPointer.PHITranslateValue(BB, Pred, 0); 931 932 Value *PredPtrVal = PredPointer.getAddr(); 933 934 // Check to see if we have already visited this pred block with another 935 // pointer. If so, we can't do this lookup. This failure can occur 936 // with PHI translation when a critical edge exists and the PHI node in 937 // the successor translates to a pointer value different than the 938 // pointer the block was first analyzed with. 939 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 940 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 941 942 if (!InsertRes.second) { 943 // If the predecessor was visited with PredPtr, then we already did 944 // the analysis and can ignore it. 945 if (InsertRes.first->second == PredPtrVal) 946 continue; 947 948 // Otherwise, the block was previously analyzed with a different 949 // pointer. We can't represent the result of this case, so we just 950 // treat this as a phi translation failure. 951 goto PredTranslationFailure; 952 } 953 954 // If PHI translation was unable to find an available pointer in this 955 // predecessor, then we have to assume that the pointer is clobbered in 956 // that predecessor. We can still do PRE of the load, which would insert 957 // a computation of the pointer in this predecessor. 958 if (PredPtrVal == 0) { 959 // Add the entry to the Result list. 960 NonLocalDepResult Entry(Pred, 961 MemDepResult::getClobber(Pred->getTerminator()), 962 PredPtrVal); 963 Result.push_back(Entry); 964 965 // Since we had a phi translation failure, the cache for CacheKey won't 966 // include all of the entries that we need to immediately satisfy future 967 // queries. Mark this in NonLocalPointerDeps by setting the 968 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 969 // cached value to do more work but not miss the phi trans failure. 970 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey]; 971 NLPI.Pair = BBSkipFirstBlockPair(); 972 continue; 973 } 974 975 // FIXME: it is entirely possible that PHI translating will end up with 976 // the same value. Consider PHI translating something like: 977 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 978 // to recurse here, pedantically speaking. 979 980 // If we have a problem phi translating, fall through to the code below 981 // to handle the failure condition. 982 if (getNonLocalPointerDepFromBB(PredPointer, 983 Loc.getWithNewPtr(PredPointer.getAddr()), 984 isLoad, Pred, 985 Result, Visited)) 986 goto PredTranslationFailure; 987 } 988 989 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 990 CacheInfo = &NonLocalPointerDeps[CacheKey]; 991 Cache = &CacheInfo->NonLocalDeps; 992 NumSortedEntries = Cache->size(); 993 994 // Since we did phi translation, the "Cache" set won't contain all of the 995 // results for the query. This is ok (we can still use it to accelerate 996 // specific block queries) but we can't do the fastpath "return all 997 // results from the set" Clear out the indicator for this. 998 CacheInfo->Pair = BBSkipFirstBlockPair(); 999 SkipFirstBlock = false; 1000 continue; 1001 1002 PredTranslationFailure: 1003 1004 if (Cache == 0) { 1005 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1006 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1007 Cache = &CacheInfo->NonLocalDeps; 1008 NumSortedEntries = Cache->size(); 1009 } 1010 1011 // Since we failed phi translation, the "Cache" set won't contain all of the 1012 // results for the query. This is ok (we can still use it to accelerate 1013 // specific block queries) but we can't do the fastpath "return all 1014 // results from the set". Clear out the indicator for this. 1015 CacheInfo->Pair = BBSkipFirstBlockPair(); 1016 1017 // If *nothing* works, mark the pointer as being clobbered by the first 1018 // instruction in this block. 1019 // 1020 // If this is the magic first block, return this as a clobber of the whole 1021 // incoming value. Since we can't phi translate to one of the predecessors, 1022 // we have to bail out. 1023 if (SkipFirstBlock) 1024 return true; 1025 1026 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 1027 assert(I != Cache->rend() && "Didn't find current block??"); 1028 if (I->getBB() != BB) 1029 continue; 1030 1031 assert(I->getResult().isNonLocal() && 1032 "Should only be here with transparent block"); 1033 I->setResult(MemDepResult::getClobber(BB->begin())); 1034 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey); 1035 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 1036 Pointer.getAddr())); 1037 break; 1038 } 1039 } 1040 1041 // Okay, we're done now. If we added new values to the cache, re-sort it. 1042 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1043 DEBUG(AssertSorted(*Cache)); 1044 return false; 1045} 1046 1047/// RemoveCachedNonLocalPointerDependencies - If P exists in 1048/// CachedNonLocalPointerInfo, remove it. 1049void MemoryDependenceAnalysis:: 1050RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 1051 CachedNonLocalPointerInfo::iterator It = 1052 NonLocalPointerDeps.find(P); 1053 if (It == NonLocalPointerDeps.end()) return; 1054 1055 // Remove all of the entries in the BB->val map. This involves removing 1056 // instructions from the reverse map. 1057 NonLocalDepInfo &PInfo = It->second.NonLocalDeps; 1058 1059 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 1060 Instruction *Target = PInfo[i].getResult().getInst(); 1061 if (Target == 0) continue; // Ignore non-local dep results. 1062 assert(Target->getParent() == PInfo[i].getBB()); 1063 1064 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1065 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1066 } 1067 1068 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1069 NonLocalPointerDeps.erase(It); 1070} 1071 1072 1073/// invalidateCachedPointerInfo - This method is used to invalidate cached 1074/// information about the specified pointer, because it may be too 1075/// conservative in memdep. This is an optional call that can be used when 1076/// the client detects an equivalence between the pointer and some other 1077/// value and replaces the other value with ptr. This can make Ptr available 1078/// in more places that cached info does not necessarily keep. 1079void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1080 // If Ptr isn't really a pointer, just ignore it. 1081 if (!Ptr->getType()->isPointerTy()) return; 1082 // Flush store info for the pointer. 1083 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1084 // Flush load info for the pointer. 1085 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1086} 1087 1088/// invalidateCachedPredecessors - Clear the PredIteratorCache info. 1089/// This needs to be done when the CFG changes, e.g., due to splitting 1090/// critical edges. 1091void MemoryDependenceAnalysis::invalidateCachedPredecessors() { 1092 PredCache->clear(); 1093} 1094 1095/// removeInstruction - Remove an instruction from the dependence analysis, 1096/// updating the dependence of instructions that previously depended on it. 1097/// This method attempts to keep the cache coherent using the reverse map. 1098void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1099 // Walk through the Non-local dependencies, removing this one as the value 1100 // for any cached queries. 1101 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1102 if (NLDI != NonLocalDeps.end()) { 1103 NonLocalDepInfo &BlockMap = NLDI->second.first; 1104 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1105 DI != DE; ++DI) 1106 if (Instruction *Inst = DI->getResult().getInst()) 1107 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1108 NonLocalDeps.erase(NLDI); 1109 } 1110 1111 // If we have a cached local dependence query for this instruction, remove it. 1112 // 1113 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1114 if (LocalDepEntry != LocalDeps.end()) { 1115 // Remove us from DepInst's reverse set now that the local dep info is gone. 1116 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1117 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1118 1119 // Remove this local dependency info. 1120 LocalDeps.erase(LocalDepEntry); 1121 } 1122 1123 // If we have any cached pointer dependencies on this instruction, remove 1124 // them. If the instruction has non-pointer type, then it can't be a pointer 1125 // base. 1126 1127 // Remove it from both the load info and the store info. The instruction 1128 // can't be in either of these maps if it is non-pointer. 1129 if (RemInst->getType()->isPointerTy()) { 1130 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1131 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1132 } 1133 1134 // Loop over all of the things that depend on the instruction we're removing. 1135 // 1136 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1137 1138 // If we find RemInst as a clobber or Def in any of the maps for other values, 1139 // we need to replace its entry with a dirty version of the instruction after 1140 // it. If RemInst is a terminator, we use a null dirty value. 1141 // 1142 // Using a dirty version of the instruction after RemInst saves having to scan 1143 // the entire block to get to this point. 1144 MemDepResult NewDirtyVal; 1145 if (!RemInst->isTerminator()) 1146 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1147 1148 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1149 if (ReverseDepIt != ReverseLocalDeps.end()) { 1150 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1151 // RemInst can't be the terminator if it has local stuff depending on it. 1152 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1153 "Nothing can locally depend on a terminator"); 1154 1155 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1156 E = ReverseDeps.end(); I != E; ++I) { 1157 Instruction *InstDependingOnRemInst = *I; 1158 assert(InstDependingOnRemInst != RemInst && 1159 "Already removed our local dep info"); 1160 1161 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1162 1163 // Make sure to remember that new things depend on NewDepInst. 1164 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1165 "a local dep on this if it is a terminator!"); 1166 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1167 InstDependingOnRemInst)); 1168 } 1169 1170 ReverseLocalDeps.erase(ReverseDepIt); 1171 1172 // Add new reverse deps after scanning the set, to avoid invalidating the 1173 // 'ReverseDeps' reference. 1174 while (!ReverseDepsToAdd.empty()) { 1175 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1176 .insert(ReverseDepsToAdd.back().second); 1177 ReverseDepsToAdd.pop_back(); 1178 } 1179 } 1180 1181 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1182 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1183 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1184 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1185 I != E; ++I) { 1186 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1187 1188 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1189 // The information is now dirty! 1190 INLD.second = true; 1191 1192 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1193 DE = INLD.first.end(); DI != DE; ++DI) { 1194 if (DI->getResult().getInst() != RemInst) continue; 1195 1196 // Convert to a dirty entry for the subsequent instruction. 1197 DI->setResult(NewDirtyVal); 1198 1199 if (Instruction *NextI = NewDirtyVal.getInst()) 1200 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1201 } 1202 } 1203 1204 ReverseNonLocalDeps.erase(ReverseDepIt); 1205 1206 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1207 while (!ReverseDepsToAdd.empty()) { 1208 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1209 .insert(ReverseDepsToAdd.back().second); 1210 ReverseDepsToAdd.pop_back(); 1211 } 1212 } 1213 1214 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1215 // value in the NonLocalPointerDeps info. 1216 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1217 ReverseNonLocalPtrDeps.find(RemInst); 1218 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1219 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1220 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1221 1222 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1223 E = Set.end(); I != E; ++I) { 1224 ValueIsLoadPair P = *I; 1225 assert(P.getPointer() != RemInst && 1226 "Already removed NonLocalPointerDeps info for RemInst"); 1227 1228 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps; 1229 1230 // The cache is not valid for any specific block anymore. 1231 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair(); 1232 1233 // Update any entries for RemInst to use the instruction after it. 1234 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1235 DI != DE; ++DI) { 1236 if (DI->getResult().getInst() != RemInst) continue; 1237 1238 // Convert to a dirty entry for the subsequent instruction. 1239 DI->setResult(NewDirtyVal); 1240 1241 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1242 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1243 } 1244 1245 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1246 // subsequent value may invalidate the sortedness. 1247 std::sort(NLPDI.begin(), NLPDI.end()); 1248 } 1249 1250 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1251 1252 while (!ReversePtrDepsToAdd.empty()) { 1253 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1254 .insert(ReversePtrDepsToAdd.back().second); 1255 ReversePtrDepsToAdd.pop_back(); 1256 } 1257 } 1258 1259 1260 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1261 AA->deleteValue(RemInst); 1262 DEBUG(verifyRemoved(RemInst)); 1263} 1264/// verifyRemoved - Verify that the specified instruction does not occur 1265/// in our internal data structures. 1266void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1267 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1268 E = LocalDeps.end(); I != E; ++I) { 1269 assert(I->first != D && "Inst occurs in data structures"); 1270 assert(I->second.getInst() != D && 1271 "Inst occurs in data structures"); 1272 } 1273 1274 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1275 E = NonLocalPointerDeps.end(); I != E; ++I) { 1276 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1277 const NonLocalDepInfo &Val = I->second.NonLocalDeps; 1278 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1279 II != E; ++II) 1280 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1281 } 1282 1283 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1284 E = NonLocalDeps.end(); I != E; ++I) { 1285 assert(I->first != D && "Inst occurs in data structures"); 1286 const PerInstNLInfo &INLD = I->second; 1287 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1288 EE = INLD.first.end(); II != EE; ++II) 1289 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1290 } 1291 1292 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1293 E = ReverseLocalDeps.end(); I != E; ++I) { 1294 assert(I->first != D && "Inst occurs in data structures"); 1295 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1296 EE = I->second.end(); II != EE; ++II) 1297 assert(*II != D && "Inst occurs in data structures"); 1298 } 1299 1300 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1301 E = ReverseNonLocalDeps.end(); 1302 I != E; ++I) { 1303 assert(I->first != D && "Inst occurs in data structures"); 1304 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1305 EE = I->second.end(); II != EE; ++II) 1306 assert(*II != D && "Inst occurs in data structures"); 1307 } 1308 1309 for (ReverseNonLocalPtrDepTy::const_iterator 1310 I = ReverseNonLocalPtrDeps.begin(), 1311 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1312 assert(I->first != D && "Inst occurs in rev NLPD map"); 1313 1314 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1315 E = I->second.end(); II != E; ++II) 1316 assert(*II != ValueIsLoadPair(D, false) && 1317 *II != ValueIsLoadPair(D, true) && 1318 "Inst occurs in ReverseNonLocalPtrDeps map"); 1319 } 1320 1321} 1322