MemoryDependenceAnalysis.cpp revision 5034dd318a9dfa0dc45a3ac01e58e60f2aa2498d
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/Analysis/ValueTracking.h" 29#include "llvm/ADT/Statistic.h" 30#include "llvm/ADT/STLExtras.h" 31#include "llvm/Support/PredIteratorCache.h" 32#include "llvm/Support/Debug.h" 33#include "llvm/Target/TargetData.h" 34using namespace llvm; 35 36STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 37STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 38STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 39 40STATISTIC(NumCacheNonLocalPtr, 41 "Number of fully cached non-local ptr responses"); 42STATISTIC(NumCacheDirtyNonLocalPtr, 43 "Number of cached, but dirty, non-local ptr responses"); 44STATISTIC(NumUncacheNonLocalPtr, 45 "Number of uncached non-local ptr responses"); 46STATISTIC(NumCacheCompleteNonLocalPtr, 47 "Number of block queries that were completely cached"); 48 49char MemoryDependenceAnalysis::ID = 0; 50 51// Register this pass... 52INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep", 53 "Memory Dependence Analysis", false, true) 54INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 55INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep", 56 "Memory Dependence Analysis", false, true) 57 58MemoryDependenceAnalysis::MemoryDependenceAnalysis() 59: FunctionPass(ID), PredCache(0) { 60 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry()); 61} 62MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 63} 64 65/// Clean up memory in between runs 66void MemoryDependenceAnalysis::releaseMemory() { 67 LocalDeps.clear(); 68 NonLocalDeps.clear(); 69 NonLocalPointerDeps.clear(); 70 ReverseLocalDeps.clear(); 71 ReverseNonLocalDeps.clear(); 72 ReverseNonLocalPtrDeps.clear(); 73 PredCache->clear(); 74} 75 76 77 78/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 79/// 80void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 81 AU.setPreservesAll(); 82 AU.addRequiredTransitive<AliasAnalysis>(); 83} 84 85bool MemoryDependenceAnalysis::runOnFunction(Function &) { 86 AA = &getAnalysis<AliasAnalysis>(); 87 TD = getAnalysisIfAvailable<TargetData>(); 88 if (PredCache == 0) 89 PredCache.reset(new PredIteratorCache()); 90 return false; 91} 92 93/// RemoveFromReverseMap - This is a helper function that removes Val from 94/// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry. 95template <typename KeyTy> 96static void RemoveFromReverseMap(DenseMap<Instruction*, 97 SmallPtrSet<KeyTy, 4> > &ReverseMap, 98 Instruction *Inst, KeyTy Val) { 99 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator 100 InstIt = ReverseMap.find(Inst); 101 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); 102 bool Found = InstIt->second.erase(Val); 103 assert(Found && "Invalid reverse map!"); Found=Found; 104 if (InstIt->second.empty()) 105 ReverseMap.erase(InstIt); 106} 107 108/// GetLocation - If the given instruction references a specific memory 109/// location, fill in Loc with the details, otherwise set Loc.Ptr to null. 110/// Return a ModRefInfo value describing the general behavior of the 111/// instruction. 112static 113AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst, 114 AliasAnalysis::Location &Loc, 115 AliasAnalysis *AA) { 116 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 117 if (LI->isVolatile()) { 118 Loc = AliasAnalysis::Location(); 119 return AliasAnalysis::ModRef; 120 } 121 Loc = AA->getLocation(LI); 122 return AliasAnalysis::Ref; 123 } 124 125 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 126 if (SI->isVolatile()) { 127 Loc = AliasAnalysis::Location(); 128 return AliasAnalysis::ModRef; 129 } 130 Loc = AA->getLocation(SI); 131 return AliasAnalysis::Mod; 132 } 133 134 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 135 Loc = AA->getLocation(V); 136 return AliasAnalysis::ModRef; 137 } 138 139 if (const CallInst *CI = isFreeCall(Inst)) { 140 // calls to free() deallocate the entire structure 141 Loc = AliasAnalysis::Location(CI->getArgOperand(0)); 142 return AliasAnalysis::Mod; 143 } 144 145 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) 146 switch (II->getIntrinsicID()) { 147 case Intrinsic::lifetime_start: 148 case Intrinsic::lifetime_end: 149 case Intrinsic::invariant_start: 150 Loc = AliasAnalysis::Location(II->getArgOperand(1), 151 cast<ConstantInt>(II->getArgOperand(0)) 152 ->getZExtValue(), 153 II->getMetadata(LLVMContext::MD_tbaa)); 154 // These intrinsics don't really modify the memory, but returning Mod 155 // will allow them to be handled conservatively. 156 return AliasAnalysis::Mod; 157 case Intrinsic::invariant_end: 158 Loc = AliasAnalysis::Location(II->getArgOperand(2), 159 cast<ConstantInt>(II->getArgOperand(1)) 160 ->getZExtValue(), 161 II->getMetadata(LLVMContext::MD_tbaa)); 162 // These intrinsics don't really modify the memory, but returning Mod 163 // will allow them to be handled conservatively. 164 return AliasAnalysis::Mod; 165 default: 166 break; 167 } 168 169 // Otherwise, just do the coarse-grained thing that always works. 170 if (Inst->mayWriteToMemory()) 171 return AliasAnalysis::ModRef; 172 if (Inst->mayReadFromMemory()) 173 return AliasAnalysis::Ref; 174 return AliasAnalysis::NoModRef; 175} 176 177/// getCallSiteDependencyFrom - Private helper for finding the local 178/// dependencies of a call site. 179MemDepResult MemoryDependenceAnalysis:: 180getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 181 BasicBlock::iterator ScanIt, BasicBlock *BB) { 182 // Walk backwards through the block, looking for dependencies 183 while (ScanIt != BB->begin()) { 184 Instruction *Inst = --ScanIt; 185 186 // If this inst is a memory op, get the pointer it accessed 187 AliasAnalysis::Location Loc; 188 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA); 189 if (Loc.Ptr) { 190 // A simple instruction. 191 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef) 192 return MemDepResult::getClobber(Inst); 193 continue; 194 } 195 196 if (CallSite InstCS = cast<Value>(Inst)) { 197 // Debug intrinsics don't cause dependences. 198 if (isa<DbgInfoIntrinsic>(Inst)) continue; 199 // If these two calls do not interfere, look past it. 200 switch (AA->getModRefInfo(CS, InstCS)) { 201 case AliasAnalysis::NoModRef: 202 // If the two calls are the same, return InstCS as a Def, so that 203 // CS can be found redundant and eliminated. 204 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) && 205 CS.getInstruction()->isIdenticalToWhenDefined(Inst)) 206 return MemDepResult::getDef(Inst); 207 208 // Otherwise if the two calls don't interact (e.g. InstCS is readnone) 209 // keep scanning. 210 break; 211 default: 212 return MemDepResult::getClobber(Inst); 213 } 214 } 215 } 216 217 // No dependence found. If this is the entry block of the function, it is a 218 // clobber, otherwise it is non-local. 219 if (BB != &BB->getParent()->getEntryBlock()) 220 return MemDepResult::getNonLocal(); 221 return MemDepResult::getClobber(ScanIt); 222} 223 224/// getPointerDependencyFrom - Return the instruction on which a memory 225/// location depends. If isLoad is true, this routine ignores may-aliases with 226/// read-only operations. If isLoad is false, this routine ignores may-aliases 227/// with reads from read-only locations. 228MemDepResult MemoryDependenceAnalysis:: 229getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad, 230 BasicBlock::iterator ScanIt, BasicBlock *BB) { 231 232 Value *InvariantTag = 0; 233 234 // Walk backwards through the basic block, looking for dependencies. 235 while (ScanIt != BB->begin()) { 236 Instruction *Inst = --ScanIt; 237 238 // If we're in an invariant region, no dependencies can be found before 239 // we pass an invariant-begin marker. 240 if (InvariantTag == Inst) { 241 InvariantTag = 0; 242 continue; 243 } 244 245 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 246 // Debug intrinsics don't (and can't) cause dependences. 247 if (isa<DbgInfoIntrinsic>(II)) continue; 248 249 // If we pass an invariant-end marker, then we've just entered an 250 // invariant region and can start ignoring dependencies. 251 if (II->getIntrinsicID() == Intrinsic::invariant_end) { 252 // FIXME: This only considers queries directly on the invariant-tagged 253 // pointer, not on query pointers that are indexed off of them. It'd 254 // be nice to handle that at some point. 255 AliasAnalysis::AliasResult R = 256 AA->alias(AliasAnalysis::Location(II->getArgOperand(2)), MemLoc); 257 if (R == AliasAnalysis::MustAlias) 258 InvariantTag = II->getArgOperand(0); 259 260 continue; 261 } 262 263 // If we reach a lifetime begin or end marker, then the query ends here 264 // because the value is undefined. 265 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 266 // FIXME: This only considers queries directly on the invariant-tagged 267 // pointer, not on query pointers that are indexed off of them. It'd 268 // be nice to handle that at some point. 269 AliasAnalysis::AliasResult R = 270 AA->alias(AliasAnalysis::Location(II->getArgOperand(1)), MemLoc); 271 if (R == AliasAnalysis::MustAlias) 272 return MemDepResult::getDef(II); 273 continue; 274 } 275 } 276 277 // If we're querying on a load and we're in an invariant region, we're done 278 // at this point. Nothing a load depends on can live in an invariant region. 279 // 280 // FIXME: this will prevent us from returning load/load must-aliases, so GVN 281 // won't remove redundant loads. 282 if (isLoad && InvariantTag) continue; 283 284 // Values depend on loads if the pointers are must aliased. This means that 285 // a load depends on another must aliased load from the same value. 286 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 287 AliasAnalysis::Location LoadLoc = AA->getLocation(LI); 288 289 // If we found a pointer, check if it could be the same as our pointer. 290 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc); 291 if (R == AliasAnalysis::NoAlias) 292 continue; 293 294 // May-alias loads don't depend on each other without a dependence. 295 if (isLoad && R != AliasAnalysis::MustAlias) 296 continue; 297 298 // Stores don't alias loads from read-only memory. 299 if (!isLoad && AA->pointsToConstantMemory(LoadLoc)) 300 continue; 301 302 // Stores depend on may and must aliased loads, loads depend on must-alias 303 // loads. 304 return MemDepResult::getDef(Inst); 305 } 306 307 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 308 // There can't be stores to the value we care about inside an 309 // invariant region. 310 if (InvariantTag) continue; 311 312 // If alias analysis can tell that this store is guaranteed to not modify 313 // the query pointer, ignore it. Use getModRefInfo to handle cases where 314 // the query pointer points to constant memory etc. 315 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef) 316 continue; 317 318 // Ok, this store might clobber the query pointer. Check to see if it is 319 // a must alias: in this case, we want to return this as a def. 320 AliasAnalysis::Location StoreLoc = AA->getLocation(SI); 321 322 // If we found a pointer, check if it could be the same as our pointer. 323 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc); 324 325 if (R == AliasAnalysis::NoAlias) 326 continue; 327 if (R == AliasAnalysis::MustAlias) 328 return MemDepResult::getDef(Inst); 329 return MemDepResult::getClobber(Inst); 330 } 331 332 // If this is an allocation, and if we know that the accessed pointer is to 333 // the allocation, return Def. This means that there is no dependence and 334 // the access can be optimized based on that. For example, a load could 335 // turn into undef. 336 // Note: Only determine this to be a malloc if Inst is the malloc call, not 337 // a subsequent bitcast of the malloc call result. There can be stores to 338 // the malloced memory between the malloc call and its bitcast uses, and we 339 // need to continue scanning until the malloc call. 340 if (isa<AllocaInst>(Inst) || 341 (isa<CallInst>(Inst) && extractMallocCall(Inst))) { 342 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr); 343 344 if (AccessPtr == Inst || 345 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 346 return MemDepResult::getDef(Inst); 347 continue; 348 } 349 350 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 351 switch (AA->getModRefInfo(Inst, MemLoc)) { 352 case AliasAnalysis::NoModRef: 353 // If the call has no effect on the queried pointer, just ignore it. 354 continue; 355 case AliasAnalysis::Mod: 356 // If we're in an invariant region, we can ignore calls that ONLY 357 // modify the pointer. 358 if (InvariantTag) continue; 359 return MemDepResult::getClobber(Inst); 360 case AliasAnalysis::Ref: 361 // If the call is known to never store to the pointer, and if this is a 362 // load query, we can safely ignore it (scan past it). 363 if (isLoad) 364 continue; 365 default: 366 // Otherwise, there is a potential dependence. Return a clobber. 367 return MemDepResult::getClobber(Inst); 368 } 369 } 370 371 // No dependence found. If this is the entry block of the function, it is a 372 // clobber, otherwise it is non-local. 373 if (BB != &BB->getParent()->getEntryBlock()) 374 return MemDepResult::getNonLocal(); 375 return MemDepResult::getClobber(ScanIt); 376} 377 378/// getDependency - Return the instruction on which a memory operation 379/// depends. 380MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 381 Instruction *ScanPos = QueryInst; 382 383 // Check for a cached result 384 MemDepResult &LocalCache = LocalDeps[QueryInst]; 385 386 // If the cached entry is non-dirty, just return it. Note that this depends 387 // on MemDepResult's default constructing to 'dirty'. 388 if (!LocalCache.isDirty()) 389 return LocalCache; 390 391 // Otherwise, if we have a dirty entry, we know we can start the scan at that 392 // instruction, which may save us some work. 393 if (Instruction *Inst = LocalCache.getInst()) { 394 ScanPos = Inst; 395 396 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 397 } 398 399 BasicBlock *QueryParent = QueryInst->getParent(); 400 401 // Do the scan. 402 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 403 // No dependence found. If this is the entry block of the function, it is a 404 // clobber, otherwise it is non-local. 405 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 406 LocalCache = MemDepResult::getNonLocal(); 407 else 408 LocalCache = MemDepResult::getClobber(QueryInst); 409 } else { 410 AliasAnalysis::Location MemLoc; 411 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA); 412 if (MemLoc.Ptr) { 413 // If we can do a pointer scan, make it happen. 414 bool isLoad = !(MR & AliasAnalysis::Mod); 415 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst)) 416 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end; 417 418 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos, 419 QueryParent); 420 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 421 CallSite QueryCS(QueryInst); 422 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 423 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 424 QueryParent); 425 } else 426 // Non-memory instruction. 427 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 428 } 429 430 // Remember the result! 431 if (Instruction *I = LocalCache.getInst()) 432 ReverseLocalDeps[I].insert(QueryInst); 433 434 return LocalCache; 435} 436 437#ifndef NDEBUG 438/// AssertSorted - This method is used when -debug is specified to verify that 439/// cache arrays are properly kept sorted. 440static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 441 int Count = -1) { 442 if (Count == -1) Count = Cache.size(); 443 if (Count == 0) return; 444 445 for (unsigned i = 1; i != unsigned(Count); ++i) 446 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 447} 448#endif 449 450/// getNonLocalCallDependency - Perform a full dependency query for the 451/// specified call, returning the set of blocks that the value is 452/// potentially live across. The returned set of results will include a 453/// "NonLocal" result for all blocks where the value is live across. 454/// 455/// This method assumes the instruction returns a "NonLocal" dependency 456/// within its own block. 457/// 458/// This returns a reference to an internal data structure that may be 459/// invalidated on the next non-local query or when an instruction is 460/// removed. Clients must copy this data if they want it around longer than 461/// that. 462const MemoryDependenceAnalysis::NonLocalDepInfo & 463MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 464 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 465 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 466 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 467 NonLocalDepInfo &Cache = CacheP.first; 468 469 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 470 /// the cached case, this can happen due to instructions being deleted etc. In 471 /// the uncached case, this starts out as the set of predecessors we care 472 /// about. 473 SmallVector<BasicBlock*, 32> DirtyBlocks; 474 475 if (!Cache.empty()) { 476 // Okay, we have a cache entry. If we know it is not dirty, just return it 477 // with no computation. 478 if (!CacheP.second) { 479 ++NumCacheNonLocal; 480 return Cache; 481 } 482 483 // If we already have a partially computed set of results, scan them to 484 // determine what is dirty, seeding our initial DirtyBlocks worklist. 485 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 486 I != E; ++I) 487 if (I->getResult().isDirty()) 488 DirtyBlocks.push_back(I->getBB()); 489 490 // Sort the cache so that we can do fast binary search lookups below. 491 std::sort(Cache.begin(), Cache.end()); 492 493 ++NumCacheDirtyNonLocal; 494 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 495 // << Cache.size() << " cached: " << *QueryInst; 496 } else { 497 // Seed DirtyBlocks with each of the preds of QueryInst's block. 498 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 499 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 500 DirtyBlocks.push_back(*PI); 501 ++NumUncacheNonLocal; 502 } 503 504 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 505 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 506 507 SmallPtrSet<BasicBlock*, 64> Visited; 508 509 unsigned NumSortedEntries = Cache.size(); 510 DEBUG(AssertSorted(Cache)); 511 512 // Iterate while we still have blocks to update. 513 while (!DirtyBlocks.empty()) { 514 BasicBlock *DirtyBB = DirtyBlocks.back(); 515 DirtyBlocks.pop_back(); 516 517 // Already processed this block? 518 if (!Visited.insert(DirtyBB)) 519 continue; 520 521 // Do a binary search to see if we already have an entry for this block in 522 // the cache set. If so, find it. 523 DEBUG(AssertSorted(Cache, NumSortedEntries)); 524 NonLocalDepInfo::iterator Entry = 525 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 526 NonLocalDepEntry(DirtyBB)); 527 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB) 528 --Entry; 529 530 NonLocalDepEntry *ExistingResult = 0; 531 if (Entry != Cache.begin()+NumSortedEntries && 532 Entry->getBB() == DirtyBB) { 533 // If we already have an entry, and if it isn't already dirty, the block 534 // is done. 535 if (!Entry->getResult().isDirty()) 536 continue; 537 538 // Otherwise, remember this slot so we can update the value. 539 ExistingResult = &*Entry; 540 } 541 542 // If the dirty entry has a pointer, start scanning from it so we don't have 543 // to rescan the entire block. 544 BasicBlock::iterator ScanPos = DirtyBB->end(); 545 if (ExistingResult) { 546 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 547 ScanPos = Inst; 548 // We're removing QueryInst's use of Inst. 549 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 550 QueryCS.getInstruction()); 551 } 552 } 553 554 // Find out if this block has a local dependency for QueryInst. 555 MemDepResult Dep; 556 557 if (ScanPos != DirtyBB->begin()) { 558 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 559 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 560 // No dependence found. If this is the entry block of the function, it is 561 // a clobber, otherwise it is non-local. 562 Dep = MemDepResult::getNonLocal(); 563 } else { 564 Dep = MemDepResult::getClobber(ScanPos); 565 } 566 567 // If we had a dirty entry for the block, update it. Otherwise, just add 568 // a new entry. 569 if (ExistingResult) 570 ExistingResult->setResult(Dep); 571 else 572 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 573 574 // If the block has a dependency (i.e. it isn't completely transparent to 575 // the value), remember the association! 576 if (!Dep.isNonLocal()) { 577 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 578 // update this when we remove instructions. 579 if (Instruction *Inst = Dep.getInst()) 580 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 581 } else { 582 583 // If the block *is* completely transparent to the load, we need to check 584 // the predecessors of this block. Add them to our worklist. 585 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 586 DirtyBlocks.push_back(*PI); 587 } 588 } 589 590 return Cache; 591} 592 593/// getNonLocalPointerDependency - Perform a full dependency query for an 594/// access to the specified (non-volatile) memory location, returning the 595/// set of instructions that either define or clobber the value. 596/// 597/// This method assumes the pointer has a "NonLocal" dependency within its 598/// own block. 599/// 600void MemoryDependenceAnalysis:: 601getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad, 602 BasicBlock *FromBB, 603 SmallVectorImpl<NonLocalDepResult> &Result) { 604 assert(Loc.Ptr->getType()->isPointerTy() && 605 "Can't get pointer deps of a non-pointer!"); 606 Result.clear(); 607 608 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD); 609 610 // This is the set of blocks we've inspected, and the pointer we consider in 611 // each block. Because of critical edges, we currently bail out if querying 612 // a block with multiple different pointers. This can happen during PHI 613 // translation. 614 DenseMap<BasicBlock*, Value*> Visited; 615 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB, 616 Result, Visited, true)) 617 return; 618 Result.clear(); 619 Result.push_back(NonLocalDepResult(FromBB, 620 MemDepResult::getClobber(FromBB->begin()), 621 const_cast<Value *>(Loc.Ptr))); 622} 623 624/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 625/// Pointer/PointeeSize using either cached information in Cache or by doing a 626/// lookup (which may use dirty cache info if available). If we do a lookup, 627/// add the result to the cache. 628MemDepResult MemoryDependenceAnalysis:: 629GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc, 630 bool isLoad, BasicBlock *BB, 631 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 632 633 // Do a binary search to see if we already have an entry for this block in 634 // the cache set. If so, find it. 635 NonLocalDepInfo::iterator Entry = 636 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 637 NonLocalDepEntry(BB)); 638 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 639 --Entry; 640 641 NonLocalDepEntry *ExistingResult = 0; 642 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 643 ExistingResult = &*Entry; 644 645 // If we have a cached entry, and it is non-dirty, use it as the value for 646 // this dependency. 647 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 648 ++NumCacheNonLocalPtr; 649 return ExistingResult->getResult(); 650 } 651 652 // Otherwise, we have to scan for the value. If we have a dirty cache 653 // entry, start scanning from its position, otherwise we scan from the end 654 // of the block. 655 BasicBlock::iterator ScanPos = BB->end(); 656 if (ExistingResult && ExistingResult->getResult().getInst()) { 657 assert(ExistingResult->getResult().getInst()->getParent() == BB && 658 "Instruction invalidated?"); 659 ++NumCacheDirtyNonLocalPtr; 660 ScanPos = ExistingResult->getResult().getInst(); 661 662 // Eliminating the dirty entry from 'Cache', so update the reverse info. 663 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 664 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 665 } else { 666 ++NumUncacheNonLocalPtr; 667 } 668 669 // Scan the block for the dependency. 670 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB); 671 672 // If we had a dirty entry for the block, update it. Otherwise, just add 673 // a new entry. 674 if (ExistingResult) 675 ExistingResult->setResult(Dep); 676 else 677 Cache->push_back(NonLocalDepEntry(BB, Dep)); 678 679 // If the block has a dependency (i.e. it isn't completely transparent to 680 // the value), remember the reverse association because we just added it 681 // to Cache! 682 if (Dep.isNonLocal()) 683 return Dep; 684 685 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 686 // update MemDep when we remove instructions. 687 Instruction *Inst = Dep.getInst(); 688 assert(Inst && "Didn't depend on anything?"); 689 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 690 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 691 return Dep; 692} 693 694/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 695/// number of elements in the array that are already properly ordered. This is 696/// optimized for the case when only a few entries are added. 697static void 698SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 699 unsigned NumSortedEntries) { 700 switch (Cache.size() - NumSortedEntries) { 701 case 0: 702 // done, no new entries. 703 break; 704 case 2: { 705 // Two new entries, insert the last one into place. 706 NonLocalDepEntry Val = Cache.back(); 707 Cache.pop_back(); 708 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 709 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 710 Cache.insert(Entry, Val); 711 // FALL THROUGH. 712 } 713 case 1: 714 // One new entry, Just insert the new value at the appropriate position. 715 if (Cache.size() != 1) { 716 NonLocalDepEntry Val = Cache.back(); 717 Cache.pop_back(); 718 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 719 std::upper_bound(Cache.begin(), Cache.end(), Val); 720 Cache.insert(Entry, Val); 721 } 722 break; 723 default: 724 // Added many values, do a full scale sort. 725 std::sort(Cache.begin(), Cache.end()); 726 break; 727 } 728} 729 730/// getNonLocalPointerDepFromBB - Perform a dependency query based on 731/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 732/// results to the results vector and keep track of which blocks are visited in 733/// 'Visited'. 734/// 735/// This has special behavior for the first block queries (when SkipFirstBlock 736/// is true). In this special case, it ignores the contents of the specified 737/// block and starts returning dependence info for its predecessors. 738/// 739/// This function returns false on success, or true to indicate that it could 740/// not compute dependence information for some reason. This should be treated 741/// as a clobber dependence on the first instruction in the predecessor block. 742bool MemoryDependenceAnalysis:: 743getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, 744 const AliasAnalysis::Location &Loc, 745 bool isLoad, BasicBlock *StartBB, 746 SmallVectorImpl<NonLocalDepResult> &Result, 747 DenseMap<BasicBlock*, Value*> &Visited, 748 bool SkipFirstBlock) { 749 750 // Look up the cached info for Pointer. 751 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 752 753 // Set up a temporary NLPI value. If the map doesn't yet have an entry for 754 // CacheKey, this value will be inserted as the associated value. Otherwise, 755 // it'll be ignored, and we'll have to check to see if the cached size and 756 // tbaa tag are consistent with the current query. 757 NonLocalPointerInfo InitialNLPI; 758 InitialNLPI.Size = Loc.Size; 759 InitialNLPI.TBAATag = Loc.TBAATag; 760 761 // Get the NLPI for CacheKey, inserting one into the map if it doesn't 762 // already have one. 763 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair = 764 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI)); 765 NonLocalPointerInfo *CacheInfo = &Pair.first->second; 766 767 // If we already have a cache entry for this CacheKey, we may need to do some 768 // work to reconcile the cache entry and the current query. 769 if (!Pair.second) { 770 if (CacheInfo->Size < Loc.Size) { 771 // The query's Size is greater than the cached one. Throw out the 772 // cached data and procede with the query at the greater size. 773 CacheInfo->Pair = BBSkipFirstBlockPair(); 774 CacheInfo->Size = Loc.Size; 775 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 776 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 777 if (Instruction *Inst = DI->getResult().getInst()) 778 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 779 CacheInfo->NonLocalDeps.clear(); 780 } else if (CacheInfo->Size > Loc.Size) { 781 // This query's Size is less than the cached one. Conservatively restart 782 // the query using the greater size. 783 return getNonLocalPointerDepFromBB(Pointer, 784 Loc.getWithNewSize(CacheInfo->Size), 785 isLoad, StartBB, Result, Visited, 786 SkipFirstBlock); 787 } 788 789 // If the query's TBAATag is inconsistent with the cached one, 790 // conservatively throw out the cached data and restart the query with 791 // no tag if needed. 792 if (CacheInfo->TBAATag != Loc.TBAATag) { 793 if (CacheInfo->TBAATag) { 794 CacheInfo->Pair = BBSkipFirstBlockPair(); 795 CacheInfo->TBAATag = 0; 796 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 797 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 798 if (Instruction *Inst = DI->getResult().getInst()) 799 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 800 CacheInfo->NonLocalDeps.clear(); 801 } 802 if (Loc.TBAATag) 803 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(), 804 isLoad, StartBB, Result, Visited, 805 SkipFirstBlock); 806 } 807 } 808 809 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps; 810 811 // If we have valid cached information for exactly the block we are 812 // investigating, just return it with no recomputation. 813 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 814 // We have a fully cached result for this query then we can just return the 815 // cached results and populate the visited set. However, we have to verify 816 // that we don't already have conflicting results for these blocks. Check 817 // to ensure that if a block in the results set is in the visited set that 818 // it was for the same pointer query. 819 if (!Visited.empty()) { 820 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 821 I != E; ++I) { 822 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 823 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 824 continue; 825 826 // We have a pointer mismatch in a block. Just return clobber, saying 827 // that something was clobbered in this result. We could also do a 828 // non-fully cached query, but there is little point in doing this. 829 return true; 830 } 831 } 832 833 Value *Addr = Pointer.getAddr(); 834 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 835 I != E; ++I) { 836 Visited.insert(std::make_pair(I->getBB(), Addr)); 837 if (!I->getResult().isNonLocal()) 838 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 839 } 840 ++NumCacheCompleteNonLocalPtr; 841 return false; 842 } 843 844 // Otherwise, either this is a new block, a block with an invalid cache 845 // pointer or one that we're about to invalidate by putting more info into it 846 // than its valid cache info. If empty, the result will be valid cache info, 847 // otherwise it isn't. 848 if (Cache->empty()) 849 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 850 else 851 CacheInfo->Pair = BBSkipFirstBlockPair(); 852 853 SmallVector<BasicBlock*, 32> Worklist; 854 Worklist.push_back(StartBB); 855 856 // Keep track of the entries that we know are sorted. Previously cached 857 // entries will all be sorted. The entries we add we only sort on demand (we 858 // don't insert every element into its sorted position). We know that we 859 // won't get any reuse from currently inserted values, because we don't 860 // revisit blocks after we insert info for them. 861 unsigned NumSortedEntries = Cache->size(); 862 DEBUG(AssertSorted(*Cache)); 863 864 while (!Worklist.empty()) { 865 BasicBlock *BB = Worklist.pop_back_val(); 866 867 // Skip the first block if we have it. 868 if (!SkipFirstBlock) { 869 // Analyze the dependency of *Pointer in FromBB. See if we already have 870 // been here. 871 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 872 873 // Get the dependency info for Pointer in BB. If we have cached 874 // information, we will use it, otherwise we compute it. 875 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 876 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache, 877 NumSortedEntries); 878 879 // If we got a Def or Clobber, add this to the list of results. 880 if (!Dep.isNonLocal()) { 881 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 882 continue; 883 } 884 } 885 886 // If 'Pointer' is an instruction defined in this block, then we need to do 887 // phi translation to change it into a value live in the predecessor block. 888 // If not, we just add the predecessors to the worklist and scan them with 889 // the same Pointer. 890 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 891 SkipFirstBlock = false; 892 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 893 // Verify that we haven't looked at this block yet. 894 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 895 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 896 if (InsertRes.second) { 897 // First time we've looked at *PI. 898 Worklist.push_back(*PI); 899 continue; 900 } 901 902 // If we have seen this block before, but it was with a different 903 // pointer then we have a phi translation failure and we have to treat 904 // this as a clobber. 905 if (InsertRes.first->second != Pointer.getAddr()) 906 goto PredTranslationFailure; 907 } 908 continue; 909 } 910 911 // We do need to do phi translation, if we know ahead of time we can't phi 912 // translate this value, don't even try. 913 if (!Pointer.IsPotentiallyPHITranslatable()) 914 goto PredTranslationFailure; 915 916 // We may have added values to the cache list before this PHI translation. 917 // If so, we haven't done anything to ensure that the cache remains sorted. 918 // Sort it now (if needed) so that recursive invocations of 919 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 920 // value will only see properly sorted cache arrays. 921 if (Cache && NumSortedEntries != Cache->size()) { 922 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 923 NumSortedEntries = Cache->size(); 924 } 925 Cache = 0; 926 927 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 928 BasicBlock *Pred = *PI; 929 930 // Get the PHI translated pointer in this predecessor. This can fail if 931 // not translatable, in which case the getAddr() returns null. 932 PHITransAddr PredPointer(Pointer); 933 PredPointer.PHITranslateValue(BB, Pred, 0); 934 935 Value *PredPtrVal = PredPointer.getAddr(); 936 937 // Check to see if we have already visited this pred block with another 938 // pointer. If so, we can't do this lookup. This failure can occur 939 // with PHI translation when a critical edge exists and the PHI node in 940 // the successor translates to a pointer value different than the 941 // pointer the block was first analyzed with. 942 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 943 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 944 945 if (!InsertRes.second) { 946 // If the predecessor was visited with PredPtr, then we already did 947 // the analysis and can ignore it. 948 if (InsertRes.first->second == PredPtrVal) 949 continue; 950 951 // Otherwise, the block was previously analyzed with a different 952 // pointer. We can't represent the result of this case, so we just 953 // treat this as a phi translation failure. 954 goto PredTranslationFailure; 955 } 956 957 // If PHI translation was unable to find an available pointer in this 958 // predecessor, then we have to assume that the pointer is clobbered in 959 // that predecessor. We can still do PRE of the load, which would insert 960 // a computation of the pointer in this predecessor. 961 if (PredPtrVal == 0) { 962 // Add the entry to the Result list. 963 NonLocalDepResult Entry(Pred, 964 MemDepResult::getClobber(Pred->getTerminator()), 965 PredPtrVal); 966 Result.push_back(Entry); 967 968 // Since we had a phi translation failure, the cache for CacheKey won't 969 // include all of the entries that we need to immediately satisfy future 970 // queries. Mark this in NonLocalPointerDeps by setting the 971 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 972 // cached value to do more work but not miss the phi trans failure. 973 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey]; 974 NLPI.Pair = BBSkipFirstBlockPair(); 975 continue; 976 } 977 978 // FIXME: it is entirely possible that PHI translating will end up with 979 // the same value. Consider PHI translating something like: 980 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 981 // to recurse here, pedantically speaking. 982 983 // If we have a problem phi translating, fall through to the code below 984 // to handle the failure condition. 985 if (getNonLocalPointerDepFromBB(PredPointer, 986 Loc.getWithNewPtr(PredPointer.getAddr()), 987 isLoad, Pred, 988 Result, Visited)) 989 goto PredTranslationFailure; 990 } 991 992 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 993 CacheInfo = &NonLocalPointerDeps[CacheKey]; 994 Cache = &CacheInfo->NonLocalDeps; 995 NumSortedEntries = Cache->size(); 996 997 // Since we did phi translation, the "Cache" set won't contain all of the 998 // results for the query. This is ok (we can still use it to accelerate 999 // specific block queries) but we can't do the fastpath "return all 1000 // results from the set" Clear out the indicator for this. 1001 CacheInfo->Pair = BBSkipFirstBlockPair(); 1002 SkipFirstBlock = false; 1003 continue; 1004 1005 PredTranslationFailure: 1006 1007 if (Cache == 0) { 1008 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1009 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1010 Cache = &CacheInfo->NonLocalDeps; 1011 NumSortedEntries = Cache->size(); 1012 } 1013 1014 // Since we failed phi translation, the "Cache" set won't contain all of the 1015 // results for the query. This is ok (we can still use it to accelerate 1016 // specific block queries) but we can't do the fastpath "return all 1017 // results from the set". Clear out the indicator for this. 1018 CacheInfo->Pair = BBSkipFirstBlockPair(); 1019 1020 // If *nothing* works, mark the pointer as being clobbered by the first 1021 // instruction in this block. 1022 // 1023 // If this is the magic first block, return this as a clobber of the whole 1024 // incoming value. Since we can't phi translate to one of the predecessors, 1025 // we have to bail out. 1026 if (SkipFirstBlock) 1027 return true; 1028 1029 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 1030 assert(I != Cache->rend() && "Didn't find current block??"); 1031 if (I->getBB() != BB) 1032 continue; 1033 1034 assert(I->getResult().isNonLocal() && 1035 "Should only be here with transparent block"); 1036 I->setResult(MemDepResult::getClobber(BB->begin())); 1037 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey); 1038 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 1039 Pointer.getAddr())); 1040 break; 1041 } 1042 } 1043 1044 // Okay, we're done now. If we added new values to the cache, re-sort it. 1045 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1046 DEBUG(AssertSorted(*Cache)); 1047 return false; 1048} 1049 1050/// RemoveCachedNonLocalPointerDependencies - If P exists in 1051/// CachedNonLocalPointerInfo, remove it. 1052void MemoryDependenceAnalysis:: 1053RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 1054 CachedNonLocalPointerInfo::iterator It = 1055 NonLocalPointerDeps.find(P); 1056 if (It == NonLocalPointerDeps.end()) return; 1057 1058 // Remove all of the entries in the BB->val map. This involves removing 1059 // instructions from the reverse map. 1060 NonLocalDepInfo &PInfo = It->second.NonLocalDeps; 1061 1062 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 1063 Instruction *Target = PInfo[i].getResult().getInst(); 1064 if (Target == 0) continue; // Ignore non-local dep results. 1065 assert(Target->getParent() == PInfo[i].getBB()); 1066 1067 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1068 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1069 } 1070 1071 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1072 NonLocalPointerDeps.erase(It); 1073} 1074 1075 1076/// invalidateCachedPointerInfo - This method is used to invalidate cached 1077/// information about the specified pointer, because it may be too 1078/// conservative in memdep. This is an optional call that can be used when 1079/// the client detects an equivalence between the pointer and some other 1080/// value and replaces the other value with ptr. This can make Ptr available 1081/// in more places that cached info does not necessarily keep. 1082void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1083 // If Ptr isn't really a pointer, just ignore it. 1084 if (!Ptr->getType()->isPointerTy()) return; 1085 // Flush store info for the pointer. 1086 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1087 // Flush load info for the pointer. 1088 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1089} 1090 1091/// invalidateCachedPredecessors - Clear the PredIteratorCache info. 1092/// This needs to be done when the CFG changes, e.g., due to splitting 1093/// critical edges. 1094void MemoryDependenceAnalysis::invalidateCachedPredecessors() { 1095 PredCache->clear(); 1096} 1097 1098/// removeInstruction - Remove an instruction from the dependence analysis, 1099/// updating the dependence of instructions that previously depended on it. 1100/// This method attempts to keep the cache coherent using the reverse map. 1101void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1102 // Walk through the Non-local dependencies, removing this one as the value 1103 // for any cached queries. 1104 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1105 if (NLDI != NonLocalDeps.end()) { 1106 NonLocalDepInfo &BlockMap = NLDI->second.first; 1107 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1108 DI != DE; ++DI) 1109 if (Instruction *Inst = DI->getResult().getInst()) 1110 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1111 NonLocalDeps.erase(NLDI); 1112 } 1113 1114 // If we have a cached local dependence query for this instruction, remove it. 1115 // 1116 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1117 if (LocalDepEntry != LocalDeps.end()) { 1118 // Remove us from DepInst's reverse set now that the local dep info is gone. 1119 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1120 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1121 1122 // Remove this local dependency info. 1123 LocalDeps.erase(LocalDepEntry); 1124 } 1125 1126 // If we have any cached pointer dependencies on this instruction, remove 1127 // them. If the instruction has non-pointer type, then it can't be a pointer 1128 // base. 1129 1130 // Remove it from both the load info and the store info. The instruction 1131 // can't be in either of these maps if it is non-pointer. 1132 if (RemInst->getType()->isPointerTy()) { 1133 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1134 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1135 } 1136 1137 // Loop over all of the things that depend on the instruction we're removing. 1138 // 1139 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1140 1141 // If we find RemInst as a clobber or Def in any of the maps for other values, 1142 // we need to replace its entry with a dirty version of the instruction after 1143 // it. If RemInst is a terminator, we use a null dirty value. 1144 // 1145 // Using a dirty version of the instruction after RemInst saves having to scan 1146 // the entire block to get to this point. 1147 MemDepResult NewDirtyVal; 1148 if (!RemInst->isTerminator()) 1149 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1150 1151 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1152 if (ReverseDepIt != ReverseLocalDeps.end()) { 1153 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1154 // RemInst can't be the terminator if it has local stuff depending on it. 1155 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1156 "Nothing can locally depend on a terminator"); 1157 1158 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1159 E = ReverseDeps.end(); I != E; ++I) { 1160 Instruction *InstDependingOnRemInst = *I; 1161 assert(InstDependingOnRemInst != RemInst && 1162 "Already removed our local dep info"); 1163 1164 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1165 1166 // Make sure to remember that new things depend on NewDepInst. 1167 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1168 "a local dep on this if it is a terminator!"); 1169 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1170 InstDependingOnRemInst)); 1171 } 1172 1173 ReverseLocalDeps.erase(ReverseDepIt); 1174 1175 // Add new reverse deps after scanning the set, to avoid invalidating the 1176 // 'ReverseDeps' reference. 1177 while (!ReverseDepsToAdd.empty()) { 1178 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1179 .insert(ReverseDepsToAdd.back().second); 1180 ReverseDepsToAdd.pop_back(); 1181 } 1182 } 1183 1184 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1185 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1186 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1187 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1188 I != E; ++I) { 1189 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1190 1191 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1192 // The information is now dirty! 1193 INLD.second = true; 1194 1195 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1196 DE = INLD.first.end(); DI != DE; ++DI) { 1197 if (DI->getResult().getInst() != RemInst) continue; 1198 1199 // Convert to a dirty entry for the subsequent instruction. 1200 DI->setResult(NewDirtyVal); 1201 1202 if (Instruction *NextI = NewDirtyVal.getInst()) 1203 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1204 } 1205 } 1206 1207 ReverseNonLocalDeps.erase(ReverseDepIt); 1208 1209 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1210 while (!ReverseDepsToAdd.empty()) { 1211 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1212 .insert(ReverseDepsToAdd.back().second); 1213 ReverseDepsToAdd.pop_back(); 1214 } 1215 } 1216 1217 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1218 // value in the NonLocalPointerDeps info. 1219 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1220 ReverseNonLocalPtrDeps.find(RemInst); 1221 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1222 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1223 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1224 1225 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1226 E = Set.end(); I != E; ++I) { 1227 ValueIsLoadPair P = *I; 1228 assert(P.getPointer() != RemInst && 1229 "Already removed NonLocalPointerDeps info for RemInst"); 1230 1231 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps; 1232 1233 // The cache is not valid for any specific block anymore. 1234 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair(); 1235 1236 // Update any entries for RemInst to use the instruction after it. 1237 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1238 DI != DE; ++DI) { 1239 if (DI->getResult().getInst() != RemInst) continue; 1240 1241 // Convert to a dirty entry for the subsequent instruction. 1242 DI->setResult(NewDirtyVal); 1243 1244 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1245 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1246 } 1247 1248 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1249 // subsequent value may invalidate the sortedness. 1250 std::sort(NLPDI.begin(), NLPDI.end()); 1251 } 1252 1253 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1254 1255 while (!ReversePtrDepsToAdd.empty()) { 1256 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1257 .insert(ReversePtrDepsToAdd.back().second); 1258 ReversePtrDepsToAdd.pop_back(); 1259 } 1260 } 1261 1262 1263 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1264 AA->deleteValue(RemInst); 1265 DEBUG(verifyRemoved(RemInst)); 1266} 1267/// verifyRemoved - Verify that the specified instruction does not occur 1268/// in our internal data structures. 1269void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1270 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1271 E = LocalDeps.end(); I != E; ++I) { 1272 assert(I->first != D && "Inst occurs in data structures"); 1273 assert(I->second.getInst() != D && 1274 "Inst occurs in data structures"); 1275 } 1276 1277 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1278 E = NonLocalPointerDeps.end(); I != E; ++I) { 1279 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1280 const NonLocalDepInfo &Val = I->second.NonLocalDeps; 1281 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1282 II != E; ++II) 1283 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1284 } 1285 1286 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1287 E = NonLocalDeps.end(); I != E; ++I) { 1288 assert(I->first != D && "Inst occurs in data structures"); 1289 const PerInstNLInfo &INLD = I->second; 1290 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1291 EE = INLD.first.end(); II != EE; ++II) 1292 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1293 } 1294 1295 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1296 E = ReverseLocalDeps.end(); I != E; ++I) { 1297 assert(I->first != D && "Inst occurs in data structures"); 1298 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1299 EE = I->second.end(); II != EE; ++II) 1300 assert(*II != D && "Inst occurs in data structures"); 1301 } 1302 1303 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1304 E = ReverseNonLocalDeps.end(); 1305 I != E; ++I) { 1306 assert(I->first != D && "Inst occurs in data structures"); 1307 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1308 EE = I->second.end(); II != EE; ++II) 1309 assert(*II != D && "Inst occurs in data structures"); 1310 } 1311 1312 for (ReverseNonLocalPtrDepTy::const_iterator 1313 I = ReverseNonLocalPtrDeps.begin(), 1314 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1315 assert(I->first != D && "Inst occurs in rev NLPD map"); 1316 1317 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1318 E = I->second.end(); II != E; ++II) 1319 assert(*II != ValueIsLoadPair(D, false) && 1320 *II != ValueIsLoadPair(D, true) && 1321 "Inst occurs in ReverseNonLocalPtrDeps map"); 1322 } 1323 1324} 1325