MemoryDependenceAnalysis.cpp revision 09981982f13085d5e7fd5fb8f87b4a626d770c5f
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/Analysis/AliasAnalysis.h" 23#include "llvm/Analysis/Dominators.h" 24#include "llvm/Analysis/InstructionSimplify.h" 25#include "llvm/Analysis/MemoryBuiltins.h" 26#include "llvm/Analysis/PHITransAddr.h" 27#include "llvm/ADT/Statistic.h" 28#include "llvm/ADT/STLExtras.h" 29#include "llvm/Support/PredIteratorCache.h" 30#include "llvm/Support/Debug.h" 31using namespace llvm; 32 33STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 34STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 35STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 36 37STATISTIC(NumCacheNonLocalPtr, 38 "Number of fully cached non-local ptr responses"); 39STATISTIC(NumCacheDirtyNonLocalPtr, 40 "Number of cached, but dirty, non-local ptr responses"); 41STATISTIC(NumUncacheNonLocalPtr, 42 "Number of uncached non-local ptr responses"); 43STATISTIC(NumCacheCompleteNonLocalPtr, 44 "Number of block queries that were completely cached"); 45 46char MemoryDependenceAnalysis::ID = 0; 47 48// Register this pass... 49INITIALIZE_PASS(MemoryDependenceAnalysis, "memdep", 50 "Memory Dependence Analysis", false, true); 51 52MemoryDependenceAnalysis::MemoryDependenceAnalysis() 53: FunctionPass(ID), PredCache(0) { 54} 55MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 56} 57 58/// Clean up memory in between runs 59void MemoryDependenceAnalysis::releaseMemory() { 60 LocalDeps.clear(); 61 NonLocalDeps.clear(); 62 NonLocalPointerDeps.clear(); 63 ReverseLocalDeps.clear(); 64 ReverseNonLocalDeps.clear(); 65 ReverseNonLocalPtrDeps.clear(); 66 PredCache->clear(); 67} 68 69 70 71/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 72/// 73void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 74 AU.setPreservesAll(); 75 AU.addRequiredTransitive<AliasAnalysis>(); 76} 77 78bool MemoryDependenceAnalysis::runOnFunction(Function &) { 79 AA = &getAnalysis<AliasAnalysis>(); 80 if (PredCache == 0) 81 PredCache.reset(new PredIteratorCache()); 82 return false; 83} 84 85/// RemoveFromReverseMap - This is a helper function that removes Val from 86/// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry. 87template <typename KeyTy> 88static void RemoveFromReverseMap(DenseMap<Instruction*, 89 SmallPtrSet<KeyTy, 4> > &ReverseMap, 90 Instruction *Inst, KeyTy Val) { 91 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator 92 InstIt = ReverseMap.find(Inst); 93 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); 94 bool Found = InstIt->second.erase(Val); 95 assert(Found && "Invalid reverse map!"); Found=Found; 96 if (InstIt->second.empty()) 97 ReverseMap.erase(InstIt); 98} 99 100 101/// getCallSiteDependencyFrom - Private helper for finding the local 102/// dependencies of a call site. 103MemDepResult MemoryDependenceAnalysis:: 104getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 105 BasicBlock::iterator ScanIt, BasicBlock *BB) { 106 // Walk backwards through the block, looking for dependencies 107 while (ScanIt != BB->begin()) { 108 Instruction *Inst = --ScanIt; 109 110 // If this inst is a memory op, get the pointer it accessed 111 Value *Pointer = 0; 112 uint64_t PointerSize = 0; 113 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) { 114 Pointer = S->getPointerOperand(); 115 PointerSize = AA->getTypeStoreSize(S->getOperand(0)->getType()); 116 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 117 Pointer = V->getOperand(0); 118 PointerSize = AA->getTypeStoreSize(V->getType()); 119 } else if (const CallInst *CI = isFreeCall(Inst)) { 120 Pointer = CI->getArgOperand(0); 121 // calls to free() erase the entire structure 122 PointerSize = ~0ULL; 123 } else if (CallSite InstCS = cast<Value>(Inst)) { 124 // Debug intrinsics don't cause dependences. 125 if (isa<DbgInfoIntrinsic>(Inst)) continue; 126 // If these two calls do not interfere, look past it. 127 switch (AA->getModRefInfo(CS, InstCS)) { 128 case AliasAnalysis::NoModRef: 129 // If the two calls are the same, return InstCS as a Def, so that 130 // CS can be found redundant and eliminated. 131 if (isReadOnlyCall && InstCS.onlyReadsMemory() && 132 CS.getInstruction()->isIdenticalToWhenDefined(Inst)) 133 return MemDepResult::getDef(Inst); 134 135 // Otherwise if the two calls don't interact (e.g. InstCS is readnone) 136 // keep scanning. 137 continue; 138 default: 139 return MemDepResult::getClobber(Inst); 140 } 141 } else { 142 // Non-memory instruction. 143 continue; 144 } 145 146 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef) 147 return MemDepResult::getClobber(Inst); 148 } 149 150 // No dependence found. If this is the entry block of the function, it is a 151 // clobber, otherwise it is non-local. 152 if (BB != &BB->getParent()->getEntryBlock()) 153 return MemDepResult::getNonLocal(); 154 return MemDepResult::getClobber(ScanIt); 155} 156 157/// getPointerDependencyFrom - Return the instruction on which a memory 158/// location depends. If isLoad is true, this routine ignore may-aliases with 159/// read-only operations. 160MemDepResult MemoryDependenceAnalysis:: 161getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, 162 BasicBlock::iterator ScanIt, BasicBlock *BB) { 163 164 Value *InvariantTag = 0; 165 166 // Walk backwards through the basic block, looking for dependencies. 167 while (ScanIt != BB->begin()) { 168 Instruction *Inst = --ScanIt; 169 170 // If we're in an invariant region, no dependencies can be found before 171 // we pass an invariant-begin marker. 172 if (InvariantTag == Inst) { 173 InvariantTag = 0; 174 continue; 175 } 176 177 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 178 // Debug intrinsics don't (and can't) cause dependences. 179 if (isa<DbgInfoIntrinsic>(II)) continue; 180 181 // If we pass an invariant-end marker, then we've just entered an 182 // invariant region and can start ignoring dependencies. 183 if (II->getIntrinsicID() == Intrinsic::invariant_end) { 184 // FIXME: This only considers queries directly on the invariant-tagged 185 // pointer, not on query pointers that are indexed off of them. It'd 186 // be nice to handle that at some point. 187 AliasAnalysis::AliasResult R = AA->alias(II->getArgOperand(2), MemPtr); 188 if (R == AliasAnalysis::MustAlias) 189 InvariantTag = II->getArgOperand(0); 190 191 continue; 192 } 193 194 // If we reach a lifetime begin or end marker, then the query ends here 195 // because the value is undefined. 196 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 197 // FIXME: This only considers queries directly on the invariant-tagged 198 // pointer, not on query pointers that are indexed off of them. It'd 199 // be nice to handle that at some point. 200 AliasAnalysis::AliasResult R = AA->alias(II->getArgOperand(1), MemPtr); 201 if (R == AliasAnalysis::MustAlias) 202 return MemDepResult::getDef(II); 203 continue; 204 } 205 } 206 207 // If we're querying on a load and we're in an invariant region, we're done 208 // at this point. Nothing a load depends on can live in an invariant region. 209 // 210 // FIXME: this will prevent us from returning load/load must-aliases, so GVN 211 // won't remove redundant loads. 212 if (isLoad && InvariantTag) continue; 213 214 // Values depend on loads if the pointers are must aliased. This means that 215 // a load depends on another must aliased load from the same value. 216 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 217 Value *Pointer = LI->getPointerOperand(); 218 uint64_t PointerSize = AA->getTypeStoreSize(LI->getType()); 219 220 // If we found a pointer, check if it could be the same as our pointer. 221 AliasAnalysis::AliasResult R = 222 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 223 if (R == AliasAnalysis::NoAlias) 224 continue; 225 226 // May-alias loads don't depend on each other without a dependence. 227 if (isLoad && R == AliasAnalysis::MayAlias) 228 continue; 229 // Stores depend on may and must aliased loads, loads depend on must-alias 230 // loads. 231 return MemDepResult::getDef(Inst); 232 } 233 234 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 235 // There can't be stores to the value we care about inside an 236 // invariant region. 237 if (InvariantTag) continue; 238 239 // If alias analysis can tell that this store is guaranteed to not modify 240 // the query pointer, ignore it. Use getModRefInfo to handle cases where 241 // the query pointer points to constant memory etc. 242 if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef) 243 continue; 244 245 // Ok, this store might clobber the query pointer. Check to see if it is 246 // a must alias: in this case, we want to return this as a def. 247 Value *Pointer = SI->getPointerOperand(); 248 uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); 249 250 // If we found a pointer, check if it could be the same as our pointer. 251 AliasAnalysis::AliasResult R = 252 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 253 254 if (R == AliasAnalysis::NoAlias) 255 continue; 256 if (R == AliasAnalysis::MayAlias) 257 return MemDepResult::getClobber(Inst); 258 return MemDepResult::getDef(Inst); 259 } 260 261 // If this is an allocation, and if we know that the accessed pointer is to 262 // the allocation, return Def. This means that there is no dependence and 263 // the access can be optimized based on that. For example, a load could 264 // turn into undef. 265 // Note: Only determine this to be a malloc if Inst is the malloc call, not 266 // a subsequent bitcast of the malloc call result. There can be stores to 267 // the malloced memory between the malloc call and its bitcast uses, and we 268 // need to continue scanning until the malloc call. 269 if (isa<AllocaInst>(Inst) || 270 (isa<CallInst>(Inst) && extractMallocCall(Inst))) { 271 Value *AccessPtr = MemPtr->getUnderlyingObject(); 272 273 if (AccessPtr == Inst || 274 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 275 return MemDepResult::getDef(Inst); 276 continue; 277 } 278 279 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 280 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) { 281 case AliasAnalysis::NoModRef: 282 // If the call has no effect on the queried pointer, just ignore it. 283 continue; 284 case AliasAnalysis::Mod: 285 // If we're in an invariant region, we can ignore calls that ONLY 286 // modify the pointer. 287 if (InvariantTag) continue; 288 return MemDepResult::getClobber(Inst); 289 case AliasAnalysis::Ref: 290 // If the call is known to never store to the pointer, and if this is a 291 // load query, we can safely ignore it (scan past it). 292 if (isLoad) 293 continue; 294 default: 295 // Otherwise, there is a potential dependence. Return a clobber. 296 return MemDepResult::getClobber(Inst); 297 } 298 } 299 300 // No dependence found. If this is the entry block of the function, it is a 301 // clobber, otherwise it is non-local. 302 if (BB != &BB->getParent()->getEntryBlock()) 303 return MemDepResult::getNonLocal(); 304 return MemDepResult::getClobber(ScanIt); 305} 306 307/// getDependency - Return the instruction on which a memory operation 308/// depends. 309MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 310 Instruction *ScanPos = QueryInst; 311 312 // Check for a cached result 313 MemDepResult &LocalCache = LocalDeps[QueryInst]; 314 315 // If the cached entry is non-dirty, just return it. Note that this depends 316 // on MemDepResult's default constructing to 'dirty'. 317 if (!LocalCache.isDirty()) 318 return LocalCache; 319 320 // Otherwise, if we have a dirty entry, we know we can start the scan at that 321 // instruction, which may save us some work. 322 if (Instruction *Inst = LocalCache.getInst()) { 323 ScanPos = Inst; 324 325 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 326 } 327 328 BasicBlock *QueryParent = QueryInst->getParent(); 329 330 Value *MemPtr = 0; 331 uint64_t MemSize = 0; 332 333 // Do the scan. 334 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 335 // No dependence found. If this is the entry block of the function, it is a 336 // clobber, otherwise it is non-local. 337 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 338 LocalCache = MemDepResult::getNonLocal(); 339 else 340 LocalCache = MemDepResult::getClobber(QueryInst); 341 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) { 342 // If this is a volatile store, don't mess around with it. Just return the 343 // previous instruction as a clobber. 344 if (SI->isVolatile()) 345 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 346 else { 347 MemPtr = SI->getPointerOperand(); 348 MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); 349 } 350 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) { 351 // If this is a volatile load, don't mess around with it. Just return the 352 // previous instruction as a clobber. 353 if (LI->isVolatile()) 354 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 355 else { 356 MemPtr = LI->getPointerOperand(); 357 MemSize = AA->getTypeStoreSize(LI->getType()); 358 } 359 } else if (const CallInst *CI = isFreeCall(QueryInst)) { 360 MemPtr = CI->getArgOperand(0); 361 // calls to free() erase the entire structure, not just a field. 362 MemSize = ~0UL; 363 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 364 int IntrinsicID = 0; // Intrinsic IDs start at 1. 365 IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst); 366 if (II) 367 IntrinsicID = II->getIntrinsicID(); 368 369 switch (IntrinsicID) { 370 case Intrinsic::lifetime_start: 371 case Intrinsic::lifetime_end: 372 case Intrinsic::invariant_start: 373 MemPtr = II->getArgOperand(1); 374 MemSize = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue(); 375 break; 376 case Intrinsic::invariant_end: 377 MemPtr = II->getArgOperand(2); 378 MemSize = cast<ConstantInt>(II->getArgOperand(1))->getZExtValue(); 379 break; 380 default: 381 CallSite QueryCS(QueryInst); 382 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 383 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 384 QueryParent); 385 break; 386 } 387 } else { 388 // Non-memory instruction. 389 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 390 } 391 392 // If we need to do a pointer scan, make it happen. 393 if (MemPtr) { 394 bool isLoad = !QueryInst->mayWriteToMemory(); 395 if (IntrinsicInst *II = dyn_cast<MemoryUseIntrinsic>(QueryInst)) { 396 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end; 397 } 398 LocalCache = getPointerDependencyFrom(MemPtr, MemSize, isLoad, ScanPos, 399 QueryParent); 400 } 401 402 // Remember the result! 403 if (Instruction *I = LocalCache.getInst()) 404 ReverseLocalDeps[I].insert(QueryInst); 405 406 return LocalCache; 407} 408 409#ifndef NDEBUG 410/// AssertSorted - This method is used when -debug is specified to verify that 411/// cache arrays are properly kept sorted. 412static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 413 int Count = -1) { 414 if (Count == -1) Count = Cache.size(); 415 if (Count == 0) return; 416 417 for (unsigned i = 1; i != unsigned(Count); ++i) 418 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 419} 420#endif 421 422/// getNonLocalCallDependency - Perform a full dependency query for the 423/// specified call, returning the set of blocks that the value is 424/// potentially live across. The returned set of results will include a 425/// "NonLocal" result for all blocks where the value is live across. 426/// 427/// This method assumes the instruction returns a "NonLocal" dependency 428/// within its own block. 429/// 430/// This returns a reference to an internal data structure that may be 431/// invalidated on the next non-local query or when an instruction is 432/// removed. Clients must copy this data if they want it around longer than 433/// that. 434const MemoryDependenceAnalysis::NonLocalDepInfo & 435MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 436 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 437 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 438 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 439 NonLocalDepInfo &Cache = CacheP.first; 440 441 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 442 /// the cached case, this can happen due to instructions being deleted etc. In 443 /// the uncached case, this starts out as the set of predecessors we care 444 /// about. 445 SmallVector<BasicBlock*, 32> DirtyBlocks; 446 447 if (!Cache.empty()) { 448 // Okay, we have a cache entry. If we know it is not dirty, just return it 449 // with no computation. 450 if (!CacheP.second) { 451 ++NumCacheNonLocal; 452 return Cache; 453 } 454 455 // If we already have a partially computed set of results, scan them to 456 // determine what is dirty, seeding our initial DirtyBlocks worklist. 457 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 458 I != E; ++I) 459 if (I->getResult().isDirty()) 460 DirtyBlocks.push_back(I->getBB()); 461 462 // Sort the cache so that we can do fast binary search lookups below. 463 std::sort(Cache.begin(), Cache.end()); 464 465 ++NumCacheDirtyNonLocal; 466 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 467 // << Cache.size() << " cached: " << *QueryInst; 468 } else { 469 // Seed DirtyBlocks with each of the preds of QueryInst's block. 470 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 471 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 472 DirtyBlocks.push_back(*PI); 473 ++NumUncacheNonLocal; 474 } 475 476 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 477 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 478 479 SmallPtrSet<BasicBlock*, 64> Visited; 480 481 unsigned NumSortedEntries = Cache.size(); 482 DEBUG(AssertSorted(Cache)); 483 484 // Iterate while we still have blocks to update. 485 while (!DirtyBlocks.empty()) { 486 BasicBlock *DirtyBB = DirtyBlocks.back(); 487 DirtyBlocks.pop_back(); 488 489 // Already processed this block? 490 if (!Visited.insert(DirtyBB)) 491 continue; 492 493 // Do a binary search to see if we already have an entry for this block in 494 // the cache set. If so, find it. 495 DEBUG(AssertSorted(Cache, NumSortedEntries)); 496 NonLocalDepInfo::iterator Entry = 497 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 498 NonLocalDepEntry(DirtyBB)); 499 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB) 500 --Entry; 501 502 NonLocalDepEntry *ExistingResult = 0; 503 if (Entry != Cache.begin()+NumSortedEntries && 504 Entry->getBB() == DirtyBB) { 505 // If we already have an entry, and if it isn't already dirty, the block 506 // is done. 507 if (!Entry->getResult().isDirty()) 508 continue; 509 510 // Otherwise, remember this slot so we can update the value. 511 ExistingResult = &*Entry; 512 } 513 514 // If the dirty entry has a pointer, start scanning from it so we don't have 515 // to rescan the entire block. 516 BasicBlock::iterator ScanPos = DirtyBB->end(); 517 if (ExistingResult) { 518 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 519 ScanPos = Inst; 520 // We're removing QueryInst's use of Inst. 521 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 522 QueryCS.getInstruction()); 523 } 524 } 525 526 // Find out if this block has a local dependency for QueryInst. 527 MemDepResult Dep; 528 529 if (ScanPos != DirtyBB->begin()) { 530 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 531 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 532 // No dependence found. If this is the entry block of the function, it is 533 // a clobber, otherwise it is non-local. 534 Dep = MemDepResult::getNonLocal(); 535 } else { 536 Dep = MemDepResult::getClobber(ScanPos); 537 } 538 539 // If we had a dirty entry for the block, update it. Otherwise, just add 540 // a new entry. 541 if (ExistingResult) 542 ExistingResult->setResult(Dep); 543 else 544 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 545 546 // If the block has a dependency (i.e. it isn't completely transparent to 547 // the value), remember the association! 548 if (!Dep.isNonLocal()) { 549 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 550 // update this when we remove instructions. 551 if (Instruction *Inst = Dep.getInst()) 552 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 553 } else { 554 555 // If the block *is* completely transparent to the load, we need to check 556 // the predecessors of this block. Add them to our worklist. 557 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 558 DirtyBlocks.push_back(*PI); 559 } 560 } 561 562 return Cache; 563} 564 565/// getNonLocalPointerDependency - Perform a full dependency query for an 566/// access to the specified (non-volatile) memory location, returning the 567/// set of instructions that either define or clobber the value. 568/// 569/// This method assumes the pointer has a "NonLocal" dependency within its 570/// own block. 571/// 572void MemoryDependenceAnalysis:: 573getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB, 574 SmallVectorImpl<NonLocalDepResult> &Result) { 575 assert(Pointer->getType()->isPointerTy() && 576 "Can't get pointer deps of a non-pointer!"); 577 Result.clear(); 578 579 // We know that the pointer value is live into FromBB find the def/clobbers 580 // from presecessors. 581 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType(); 582 uint64_t PointeeSize = AA->getTypeStoreSize(EltTy); 583 584 PHITransAddr Address(Pointer, TD); 585 586 // This is the set of blocks we've inspected, and the pointer we consider in 587 // each block. Because of critical edges, we currently bail out if querying 588 // a block with multiple different pointers. This can happen during PHI 589 // translation. 590 DenseMap<BasicBlock*, Value*> Visited; 591 if (!getNonLocalPointerDepFromBB(Address, PointeeSize, isLoad, FromBB, 592 Result, Visited, true)) 593 return; 594 Result.clear(); 595 Result.push_back(NonLocalDepResult(FromBB, 596 MemDepResult::getClobber(FromBB->begin()), 597 Pointer)); 598} 599 600/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 601/// Pointer/PointeeSize using either cached information in Cache or by doing a 602/// lookup (which may use dirty cache info if available). If we do a lookup, 603/// add the result to the cache. 604MemDepResult MemoryDependenceAnalysis:: 605GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, 606 bool isLoad, BasicBlock *BB, 607 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 608 609 // Do a binary search to see if we already have an entry for this block in 610 // the cache set. If so, find it. 611 NonLocalDepInfo::iterator Entry = 612 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 613 NonLocalDepEntry(BB)); 614 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 615 --Entry; 616 617 NonLocalDepEntry *ExistingResult = 0; 618 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 619 ExistingResult = &*Entry; 620 621 // If we have a cached entry, and it is non-dirty, use it as the value for 622 // this dependency. 623 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 624 ++NumCacheNonLocalPtr; 625 return ExistingResult->getResult(); 626 } 627 628 // Otherwise, we have to scan for the value. If we have a dirty cache 629 // entry, start scanning from its position, otherwise we scan from the end 630 // of the block. 631 BasicBlock::iterator ScanPos = BB->end(); 632 if (ExistingResult && ExistingResult->getResult().getInst()) { 633 assert(ExistingResult->getResult().getInst()->getParent() == BB && 634 "Instruction invalidated?"); 635 ++NumCacheDirtyNonLocalPtr; 636 ScanPos = ExistingResult->getResult().getInst(); 637 638 // Eliminating the dirty entry from 'Cache', so update the reverse info. 639 ValueIsLoadPair CacheKey(Pointer, isLoad); 640 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 641 } else { 642 ++NumUncacheNonLocalPtr; 643 } 644 645 // Scan the block for the dependency. 646 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, 647 ScanPos, BB); 648 649 // If we had a dirty entry for the block, update it. Otherwise, just add 650 // a new entry. 651 if (ExistingResult) 652 ExistingResult->setResult(Dep); 653 else 654 Cache->push_back(NonLocalDepEntry(BB, Dep)); 655 656 // If the block has a dependency (i.e. it isn't completely transparent to 657 // the value), remember the reverse association because we just added it 658 // to Cache! 659 if (Dep.isNonLocal()) 660 return Dep; 661 662 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 663 // update MemDep when we remove instructions. 664 Instruction *Inst = Dep.getInst(); 665 assert(Inst && "Didn't depend on anything?"); 666 ValueIsLoadPair CacheKey(Pointer, isLoad); 667 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 668 return Dep; 669} 670 671/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 672/// number of elements in the array that are already properly ordered. This is 673/// optimized for the case when only a few entries are added. 674static void 675SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 676 unsigned NumSortedEntries) { 677 switch (Cache.size() - NumSortedEntries) { 678 case 0: 679 // done, no new entries. 680 break; 681 case 2: { 682 // Two new entries, insert the last one into place. 683 NonLocalDepEntry Val = Cache.back(); 684 Cache.pop_back(); 685 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 686 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 687 Cache.insert(Entry, Val); 688 // FALL THROUGH. 689 } 690 case 1: 691 // One new entry, Just insert the new value at the appropriate position. 692 if (Cache.size() != 1) { 693 NonLocalDepEntry Val = Cache.back(); 694 Cache.pop_back(); 695 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 696 std::upper_bound(Cache.begin(), Cache.end(), Val); 697 Cache.insert(Entry, Val); 698 } 699 break; 700 default: 701 // Added many values, do a full scale sort. 702 std::sort(Cache.begin(), Cache.end()); 703 break; 704 } 705} 706 707/// getNonLocalPointerDepFromBB - Perform a dependency query based on 708/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 709/// results to the results vector and keep track of which blocks are visited in 710/// 'Visited'. 711/// 712/// This has special behavior for the first block queries (when SkipFirstBlock 713/// is true). In this special case, it ignores the contents of the specified 714/// block and starts returning dependence info for its predecessors. 715/// 716/// This function returns false on success, or true to indicate that it could 717/// not compute dependence information for some reason. This should be treated 718/// as a clobber dependence on the first instruction in the predecessor block. 719bool MemoryDependenceAnalysis:: 720getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, 721 bool isLoad, BasicBlock *StartBB, 722 SmallVectorImpl<NonLocalDepResult> &Result, 723 DenseMap<BasicBlock*, Value*> &Visited, 724 bool SkipFirstBlock) { 725 726 // Look up the cached info for Pointer. 727 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 728 729 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo = 730 &NonLocalPointerDeps[CacheKey]; 731 NonLocalDepInfo *Cache = &CacheInfo->second; 732 733 // If we have valid cached information for exactly the block we are 734 // investigating, just return it with no recomputation. 735 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 736 // We have a fully cached result for this query then we can just return the 737 // cached results and populate the visited set. However, we have to verify 738 // that we don't already have conflicting results for these blocks. Check 739 // to ensure that if a block in the results set is in the visited set that 740 // it was for the same pointer query. 741 if (!Visited.empty()) { 742 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 743 I != E; ++I) { 744 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 745 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 746 continue; 747 748 // We have a pointer mismatch in a block. Just return clobber, saying 749 // that something was clobbered in this result. We could also do a 750 // non-fully cached query, but there is little point in doing this. 751 return true; 752 } 753 } 754 755 Value *Addr = Pointer.getAddr(); 756 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 757 I != E; ++I) { 758 Visited.insert(std::make_pair(I->getBB(), Addr)); 759 if (!I->getResult().isNonLocal()) 760 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 761 } 762 ++NumCacheCompleteNonLocalPtr; 763 return false; 764 } 765 766 // Otherwise, either this is a new block, a block with an invalid cache 767 // pointer or one that we're about to invalidate by putting more info into it 768 // than its valid cache info. If empty, the result will be valid cache info, 769 // otherwise it isn't. 770 if (Cache->empty()) 771 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 772 else 773 CacheInfo->first = BBSkipFirstBlockPair(); 774 775 SmallVector<BasicBlock*, 32> Worklist; 776 Worklist.push_back(StartBB); 777 778 // Keep track of the entries that we know are sorted. Previously cached 779 // entries will all be sorted. The entries we add we only sort on demand (we 780 // don't insert every element into its sorted position). We know that we 781 // won't get any reuse from currently inserted values, because we don't 782 // revisit blocks after we insert info for them. 783 unsigned NumSortedEntries = Cache->size(); 784 DEBUG(AssertSorted(*Cache)); 785 786 while (!Worklist.empty()) { 787 BasicBlock *BB = Worklist.pop_back_val(); 788 789 // Skip the first block if we have it. 790 if (!SkipFirstBlock) { 791 // Analyze the dependency of *Pointer in FromBB. See if we already have 792 // been here. 793 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 794 795 // Get the dependency info for Pointer in BB. If we have cached 796 // information, we will use it, otherwise we compute it. 797 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 798 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer.getAddr(), PointeeSize, 799 isLoad, BB, Cache, 800 NumSortedEntries); 801 802 // If we got a Def or Clobber, add this to the list of results. 803 if (!Dep.isNonLocal()) { 804 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 805 continue; 806 } 807 } 808 809 // If 'Pointer' is an instruction defined in this block, then we need to do 810 // phi translation to change it into a value live in the predecessor block. 811 // If not, we just add the predecessors to the worklist and scan them with 812 // the same Pointer. 813 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 814 SkipFirstBlock = false; 815 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 816 // Verify that we haven't looked at this block yet. 817 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 818 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 819 if (InsertRes.second) { 820 // First time we've looked at *PI. 821 Worklist.push_back(*PI); 822 continue; 823 } 824 825 // If we have seen this block before, but it was with a different 826 // pointer then we have a phi translation failure and we have to treat 827 // this as a clobber. 828 if (InsertRes.first->second != Pointer.getAddr()) 829 goto PredTranslationFailure; 830 } 831 continue; 832 } 833 834 // We do need to do phi translation, if we know ahead of time we can't phi 835 // translate this value, don't even try. 836 if (!Pointer.IsPotentiallyPHITranslatable()) 837 goto PredTranslationFailure; 838 839 // We may have added values to the cache list before this PHI translation. 840 // If so, we haven't done anything to ensure that the cache remains sorted. 841 // Sort it now (if needed) so that recursive invocations of 842 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 843 // value will only see properly sorted cache arrays. 844 if (Cache && NumSortedEntries != Cache->size()) { 845 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 846 NumSortedEntries = Cache->size(); 847 } 848 Cache = 0; 849 850 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 851 BasicBlock *Pred = *PI; 852 853 // Get the PHI translated pointer in this predecessor. This can fail if 854 // not translatable, in which case the getAddr() returns null. 855 PHITransAddr PredPointer(Pointer); 856 PredPointer.PHITranslateValue(BB, Pred, 0); 857 858 Value *PredPtrVal = PredPointer.getAddr(); 859 860 // Check to see if we have already visited this pred block with another 861 // pointer. If so, we can't do this lookup. This failure can occur 862 // with PHI translation when a critical edge exists and the PHI node in 863 // the successor translates to a pointer value different than the 864 // pointer the block was first analyzed with. 865 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 866 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 867 868 if (!InsertRes.second) { 869 // If the predecessor was visited with PredPtr, then we already did 870 // the analysis and can ignore it. 871 if (InsertRes.first->second == PredPtrVal) 872 continue; 873 874 // Otherwise, the block was previously analyzed with a different 875 // pointer. We can't represent the result of this case, so we just 876 // treat this as a phi translation failure. 877 goto PredTranslationFailure; 878 } 879 880 // If PHI translation was unable to find an available pointer in this 881 // predecessor, then we have to assume that the pointer is clobbered in 882 // that predecessor. We can still do PRE of the load, which would insert 883 // a computation of the pointer in this predecessor. 884 if (PredPtrVal == 0) { 885 // Add the entry to the Result list. 886 NonLocalDepResult Entry(Pred, 887 MemDepResult::getClobber(Pred->getTerminator()), 888 PredPtrVal); 889 Result.push_back(Entry); 890 891 // Since we had a phi translation failure, the cache for CacheKey won't 892 // include all of the entries that we need to immediately satisfy future 893 // queries. Mark this in NonLocalPointerDeps by setting the 894 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 895 // cached value to do more work but not miss the phi trans failure. 896 NonLocalPointerDeps[CacheKey].first = BBSkipFirstBlockPair(); 897 continue; 898 } 899 900 // FIXME: it is entirely possible that PHI translating will end up with 901 // the same value. Consider PHI translating something like: 902 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 903 // to recurse here, pedantically speaking. 904 905 // If we have a problem phi translating, fall through to the code below 906 // to handle the failure condition. 907 if (getNonLocalPointerDepFromBB(PredPointer, PointeeSize, isLoad, Pred, 908 Result, Visited)) 909 goto PredTranslationFailure; 910 } 911 912 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 913 CacheInfo = &NonLocalPointerDeps[CacheKey]; 914 Cache = &CacheInfo->second; 915 NumSortedEntries = Cache->size(); 916 917 // Since we did phi translation, the "Cache" set won't contain all of the 918 // results for the query. This is ok (we can still use it to accelerate 919 // specific block queries) but we can't do the fastpath "return all 920 // results from the set" Clear out the indicator for this. 921 CacheInfo->first = BBSkipFirstBlockPair(); 922 SkipFirstBlock = false; 923 continue; 924 925 PredTranslationFailure: 926 927 if (Cache == 0) { 928 // Refresh the CacheInfo/Cache pointer if it got invalidated. 929 CacheInfo = &NonLocalPointerDeps[CacheKey]; 930 Cache = &CacheInfo->second; 931 NumSortedEntries = Cache->size(); 932 } 933 934 // Since we failed phi translation, the "Cache" set won't contain all of the 935 // results for the query. This is ok (we can still use it to accelerate 936 // specific block queries) but we can't do the fastpath "return all 937 // results from the set". Clear out the indicator for this. 938 CacheInfo->first = BBSkipFirstBlockPair(); 939 940 // If *nothing* works, mark the pointer as being clobbered by the first 941 // instruction in this block. 942 // 943 // If this is the magic first block, return this as a clobber of the whole 944 // incoming value. Since we can't phi translate to one of the predecessors, 945 // we have to bail out. 946 if (SkipFirstBlock) 947 return true; 948 949 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 950 assert(I != Cache->rend() && "Didn't find current block??"); 951 if (I->getBB() != BB) 952 continue; 953 954 assert(I->getResult().isNonLocal() && 955 "Should only be here with transparent block"); 956 I->setResult(MemDepResult::getClobber(BB->begin())); 957 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey); 958 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 959 Pointer.getAddr())); 960 break; 961 } 962 } 963 964 // Okay, we're done now. If we added new values to the cache, re-sort it. 965 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 966 DEBUG(AssertSorted(*Cache)); 967 return false; 968} 969 970/// RemoveCachedNonLocalPointerDependencies - If P exists in 971/// CachedNonLocalPointerInfo, remove it. 972void MemoryDependenceAnalysis:: 973RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 974 CachedNonLocalPointerInfo::iterator It = 975 NonLocalPointerDeps.find(P); 976 if (It == NonLocalPointerDeps.end()) return; 977 978 // Remove all of the entries in the BB->val map. This involves removing 979 // instructions from the reverse map. 980 NonLocalDepInfo &PInfo = It->second.second; 981 982 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 983 Instruction *Target = PInfo[i].getResult().getInst(); 984 if (Target == 0) continue; // Ignore non-local dep results. 985 assert(Target->getParent() == PInfo[i].getBB()); 986 987 // Eliminating the dirty entry from 'Cache', so update the reverse info. 988 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 989 } 990 991 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 992 NonLocalPointerDeps.erase(It); 993} 994 995 996/// invalidateCachedPointerInfo - This method is used to invalidate cached 997/// information about the specified pointer, because it may be too 998/// conservative in memdep. This is an optional call that can be used when 999/// the client detects an equivalence between the pointer and some other 1000/// value and replaces the other value with ptr. This can make Ptr available 1001/// in more places that cached info does not necessarily keep. 1002void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1003 // If Ptr isn't really a pointer, just ignore it. 1004 if (!Ptr->getType()->isPointerTy()) return; 1005 // Flush store info for the pointer. 1006 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1007 // Flush load info for the pointer. 1008 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1009} 1010 1011/// invalidateCachedPredecessors - Clear the PredIteratorCache info. 1012/// This needs to be done when the CFG changes, e.g., due to splitting 1013/// critical edges. 1014void MemoryDependenceAnalysis::invalidateCachedPredecessors() { 1015 PredCache->clear(); 1016} 1017 1018/// removeInstruction - Remove an instruction from the dependence analysis, 1019/// updating the dependence of instructions that previously depended on it. 1020/// This method attempts to keep the cache coherent using the reverse map. 1021void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1022 // Walk through the Non-local dependencies, removing this one as the value 1023 // for any cached queries. 1024 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1025 if (NLDI != NonLocalDeps.end()) { 1026 NonLocalDepInfo &BlockMap = NLDI->second.first; 1027 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1028 DI != DE; ++DI) 1029 if (Instruction *Inst = DI->getResult().getInst()) 1030 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1031 NonLocalDeps.erase(NLDI); 1032 } 1033 1034 // If we have a cached local dependence query for this instruction, remove it. 1035 // 1036 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1037 if (LocalDepEntry != LocalDeps.end()) { 1038 // Remove us from DepInst's reverse set now that the local dep info is gone. 1039 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1040 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1041 1042 // Remove this local dependency info. 1043 LocalDeps.erase(LocalDepEntry); 1044 } 1045 1046 // If we have any cached pointer dependencies on this instruction, remove 1047 // them. If the instruction has non-pointer type, then it can't be a pointer 1048 // base. 1049 1050 // Remove it from both the load info and the store info. The instruction 1051 // can't be in either of these maps if it is non-pointer. 1052 if (RemInst->getType()->isPointerTy()) { 1053 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1054 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1055 } 1056 1057 // Loop over all of the things that depend on the instruction we're removing. 1058 // 1059 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1060 1061 // If we find RemInst as a clobber or Def in any of the maps for other values, 1062 // we need to replace its entry with a dirty version of the instruction after 1063 // it. If RemInst is a terminator, we use a null dirty value. 1064 // 1065 // Using a dirty version of the instruction after RemInst saves having to scan 1066 // the entire block to get to this point. 1067 MemDepResult NewDirtyVal; 1068 if (!RemInst->isTerminator()) 1069 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1070 1071 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1072 if (ReverseDepIt != ReverseLocalDeps.end()) { 1073 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1074 // RemInst can't be the terminator if it has local stuff depending on it. 1075 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1076 "Nothing can locally depend on a terminator"); 1077 1078 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1079 E = ReverseDeps.end(); I != E; ++I) { 1080 Instruction *InstDependingOnRemInst = *I; 1081 assert(InstDependingOnRemInst != RemInst && 1082 "Already removed our local dep info"); 1083 1084 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1085 1086 // Make sure to remember that new things depend on NewDepInst. 1087 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1088 "a local dep on this if it is a terminator!"); 1089 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1090 InstDependingOnRemInst)); 1091 } 1092 1093 ReverseLocalDeps.erase(ReverseDepIt); 1094 1095 // Add new reverse deps after scanning the set, to avoid invalidating the 1096 // 'ReverseDeps' reference. 1097 while (!ReverseDepsToAdd.empty()) { 1098 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1099 .insert(ReverseDepsToAdd.back().second); 1100 ReverseDepsToAdd.pop_back(); 1101 } 1102 } 1103 1104 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1105 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1106 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1107 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1108 I != E; ++I) { 1109 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1110 1111 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1112 // The information is now dirty! 1113 INLD.second = true; 1114 1115 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1116 DE = INLD.first.end(); DI != DE; ++DI) { 1117 if (DI->getResult().getInst() != RemInst) continue; 1118 1119 // Convert to a dirty entry for the subsequent instruction. 1120 DI->setResult(NewDirtyVal); 1121 1122 if (Instruction *NextI = NewDirtyVal.getInst()) 1123 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1124 } 1125 } 1126 1127 ReverseNonLocalDeps.erase(ReverseDepIt); 1128 1129 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1130 while (!ReverseDepsToAdd.empty()) { 1131 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1132 .insert(ReverseDepsToAdd.back().second); 1133 ReverseDepsToAdd.pop_back(); 1134 } 1135 } 1136 1137 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1138 // value in the NonLocalPointerDeps info. 1139 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1140 ReverseNonLocalPtrDeps.find(RemInst); 1141 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1142 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1143 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1144 1145 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1146 E = Set.end(); I != E; ++I) { 1147 ValueIsLoadPair P = *I; 1148 assert(P.getPointer() != RemInst && 1149 "Already removed NonLocalPointerDeps info for RemInst"); 1150 1151 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second; 1152 1153 // The cache is not valid for any specific block anymore. 1154 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair(); 1155 1156 // Update any entries for RemInst to use the instruction after it. 1157 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1158 DI != DE; ++DI) { 1159 if (DI->getResult().getInst() != RemInst) continue; 1160 1161 // Convert to a dirty entry for the subsequent instruction. 1162 DI->setResult(NewDirtyVal); 1163 1164 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1165 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1166 } 1167 1168 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1169 // subsequent value may invalidate the sortedness. 1170 std::sort(NLPDI.begin(), NLPDI.end()); 1171 } 1172 1173 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1174 1175 while (!ReversePtrDepsToAdd.empty()) { 1176 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1177 .insert(ReversePtrDepsToAdd.back().second); 1178 ReversePtrDepsToAdd.pop_back(); 1179 } 1180 } 1181 1182 1183 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1184 AA->deleteValue(RemInst); 1185 DEBUG(verifyRemoved(RemInst)); 1186} 1187/// verifyRemoved - Verify that the specified instruction does not occur 1188/// in our internal data structures. 1189void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1190 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1191 E = LocalDeps.end(); I != E; ++I) { 1192 assert(I->first != D && "Inst occurs in data structures"); 1193 assert(I->second.getInst() != D && 1194 "Inst occurs in data structures"); 1195 } 1196 1197 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1198 E = NonLocalPointerDeps.end(); I != E; ++I) { 1199 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1200 const NonLocalDepInfo &Val = I->second.second; 1201 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1202 II != E; ++II) 1203 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1204 } 1205 1206 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1207 E = NonLocalDeps.end(); I != E; ++I) { 1208 assert(I->first != D && "Inst occurs in data structures"); 1209 const PerInstNLInfo &INLD = I->second; 1210 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1211 EE = INLD.first.end(); II != EE; ++II) 1212 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1213 } 1214 1215 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1216 E = ReverseLocalDeps.end(); I != E; ++I) { 1217 assert(I->first != D && "Inst occurs in data structures"); 1218 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1219 EE = I->second.end(); II != EE; ++II) 1220 assert(*II != D && "Inst occurs in data structures"); 1221 } 1222 1223 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1224 E = ReverseNonLocalDeps.end(); 1225 I != E; ++I) { 1226 assert(I->first != D && "Inst occurs in data structures"); 1227 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1228 EE = I->second.end(); II != EE; ++II) 1229 assert(*II != D && "Inst occurs in data structures"); 1230 } 1231 1232 for (ReverseNonLocalPtrDepTy::const_iterator 1233 I = ReverseNonLocalPtrDeps.begin(), 1234 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1235 assert(I->first != D && "Inst occurs in rev NLPD map"); 1236 1237 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1238 E = I->second.end(); II != E; ++II) 1239 assert(*II != ValueIsLoadPair(D, false) && 1240 *II != ValueIsLoadPair(D, true) && 1241 "Inst occurs in ReverseNonLocalPtrDeps map"); 1242 } 1243 1244} 1245