MemoryDependenceAnalysis.cpp revision 20d6f0982ad33818cfa141f80157ac13e36d5550
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/Constants.h" 20#include "llvm/Instructions.h" 21#include "llvm/Function.h" 22#include "llvm/Analysis/AliasAnalysis.h" 23#include "llvm/ADT/Statistic.h" 24#include "llvm/Support/PredIteratorCache.h" 25#include "llvm/Support/Debug.h" 26#include "llvm/Target/TargetData.h" 27using namespace llvm; 28 29STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 30STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 31STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 32 33STATISTIC(NumCacheNonLocalPtr, 34 "Number of fully cached non-local ptr responses"); 35STATISTIC(NumCacheDirtyNonLocalPtr, 36 "Number of cached, but dirty, non-local ptr responses"); 37STATISTIC(NumUncacheNonLocalPtr, 38 "Number of uncached non-local ptr responses"); 39STATISTIC(NumCacheCompleteNonLocalPtr, 40 "Number of block queries that were completely cached"); 41 42char MemoryDependenceAnalysis::ID = 0; 43 44// Register this pass... 45static RegisterPass<MemoryDependenceAnalysis> X("memdep", 46 "Memory Dependence Analysis", false, true); 47 48MemoryDependenceAnalysis::MemoryDependenceAnalysis() 49: FunctionPass(&ID), PredCache(0) { 50} 51MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 52} 53 54/// Clean up memory in between runs 55void MemoryDependenceAnalysis::releaseMemory() { 56 LocalDeps.clear(); 57 NonLocalDeps.clear(); 58 NonLocalPointerDeps.clear(); 59 ReverseLocalDeps.clear(); 60 ReverseNonLocalDeps.clear(); 61 ReverseNonLocalPtrDeps.clear(); 62 PredCache->clear(); 63} 64 65 66 67/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 68/// 69void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 70 AU.setPreservesAll(); 71 AU.addRequiredTransitive<AliasAnalysis>(); 72 AU.addRequiredTransitive<TargetData>(); 73} 74 75bool MemoryDependenceAnalysis::runOnFunction(Function &) { 76 AA = &getAnalysis<AliasAnalysis>(); 77 TD = &getAnalysis<TargetData>(); 78 if (PredCache == 0) 79 PredCache.reset(new PredIteratorCache()); 80 return false; 81} 82 83/// RemoveFromReverseMap - This is a helper function that removes Val from 84/// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry. 85template <typename KeyTy> 86static void RemoveFromReverseMap(DenseMap<Instruction*, 87 SmallPtrSet<KeyTy*, 4> > &ReverseMap, 88 Instruction *Inst, KeyTy *Val) { 89 typename DenseMap<Instruction*, SmallPtrSet<KeyTy*, 4> >::iterator 90 InstIt = ReverseMap.find(Inst); 91 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); 92 bool Found = InstIt->second.erase(Val); 93 assert(Found && "Invalid reverse map!"); Found=Found; 94 if (InstIt->second.empty()) 95 ReverseMap.erase(InstIt); 96} 97 98 99/// getCallSiteDependencyFrom - Private helper for finding the local 100/// dependencies of a call site. 101MemDepResult MemoryDependenceAnalysis:: 102getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 103 BasicBlock::iterator ScanIt, BasicBlock *BB) { 104 // Walk backwards through the block, looking for dependencies 105 while (ScanIt != BB->begin()) { 106 Instruction *Inst = --ScanIt; 107 108 // If this inst is a memory op, get the pointer it accessed 109 Value *Pointer = 0; 110 uint64_t PointerSize = 0; 111 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) { 112 Pointer = S->getPointerOperand(); 113 PointerSize = TD->getTypeStoreSize(S->getOperand(0)->getType()); 114 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 115 Pointer = V->getOperand(0); 116 PointerSize = TD->getTypeStoreSize(V->getType()); 117 } else if (FreeInst *F = dyn_cast<FreeInst>(Inst)) { 118 Pointer = F->getPointerOperand(); 119 120 // FreeInsts erase the entire structure 121 PointerSize = ~0ULL; 122 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) { 123 CallSite InstCS = CallSite::get(Inst); 124 // If these two calls do not interfere, look past it. 125 switch (AA->getModRefInfo(CS, InstCS)) { 126 case AliasAnalysis::NoModRef: 127 // If the two calls don't interact (e.g. InstCS is readnone) keep 128 // scanning. 129 continue; 130 case AliasAnalysis::Ref: 131 // If the two calls read the same memory locations and CS is a readonly 132 // function, then we have two cases: 1) the calls may not interfere with 133 // each other at all. 2) the calls may produce the same value. In case 134 // #1 we want to ignore the values, in case #2, we want to return Inst 135 // as a Def dependence. This allows us to CSE in cases like: 136 // X = strlen(P); 137 // memchr(...); 138 // Y = strlen(P); // Y = X 139 if (isReadOnlyCall) { 140 if (CS.getCalledFunction() != 0 && 141 CS.getCalledFunction() == InstCS.getCalledFunction()) 142 return MemDepResult::getDef(Inst); 143 // Ignore unrelated read/read call dependences. 144 continue; 145 } 146 // FALL THROUGH 147 default: 148 return MemDepResult::getClobber(Inst); 149 } 150 } else { 151 // Non-memory instruction. 152 continue; 153 } 154 155 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef) 156 return MemDepResult::getClobber(Inst); 157 } 158 159 // No dependence found. If this is the entry block of the function, it is a 160 // clobber, otherwise it is non-local. 161 if (BB != &BB->getParent()->getEntryBlock()) 162 return MemDepResult::getNonLocal(); 163 return MemDepResult::getClobber(ScanIt); 164} 165 166/// getPointerDependencyFrom - Return the instruction on which a memory 167/// location depends. If isLoad is true, this routine ignore may-aliases with 168/// read-only operations. 169MemDepResult MemoryDependenceAnalysis:: 170getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, 171 BasicBlock::iterator ScanIt, BasicBlock *BB) { 172 173 // Walk backwards through the basic block, looking for dependencies. 174 while (ScanIt != BB->begin()) { 175 Instruction *Inst = --ScanIt; 176 177 // Values depend on loads if the pointers are must aliased. This means that 178 // a load depends on another must aliased load from the same value. 179 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 180 Value *Pointer = LI->getPointerOperand(); 181 uint64_t PointerSize = TD->getTypeStoreSize(LI->getType()); 182 183 // If we found a pointer, check if it could be the same as our pointer. 184 AliasAnalysis::AliasResult R = 185 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 186 if (R == AliasAnalysis::NoAlias) 187 continue; 188 189 // May-alias loads don't depend on each other without a dependence. 190 if (isLoad && R == AliasAnalysis::MayAlias) 191 continue; 192 // Stores depend on may and must aliased loads, loads depend on must-alias 193 // loads. 194 return MemDepResult::getDef(Inst); 195 } 196 197 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 198 Value *Pointer = SI->getPointerOperand(); 199 uint64_t PointerSize = TD->getTypeStoreSize(SI->getOperand(0)->getType()); 200 201 // If we found a pointer, check if it could be the same as our pointer. 202 AliasAnalysis::AliasResult R = 203 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 204 205 if (R == AliasAnalysis::NoAlias) 206 continue; 207 if (R == AliasAnalysis::MayAlias) 208 return MemDepResult::getClobber(Inst); 209 return MemDepResult::getDef(Inst); 210 } 211 212 // If this is an allocation, and if we know that the accessed pointer is to 213 // the allocation, return Def. This means that there is no dependence and 214 // the access can be optimized based on that. For example, a load could 215 // turn into undef. 216 if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) { 217 Value *AccessPtr = MemPtr->getUnderlyingObject(); 218 219 if (AccessPtr == AI || 220 AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 221 return MemDepResult::getDef(AI); 222 continue; 223 } 224 225 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 226 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) { 227 case AliasAnalysis::NoModRef: 228 // If the call has no effect on the queried pointer, just ignore it. 229 continue; 230 case AliasAnalysis::Ref: 231 // If the call is known to never store to the pointer, and if this is a 232 // load query, we can safely ignore it (scan past it). 233 if (isLoad) 234 continue; 235 // FALL THROUGH. 236 default: 237 // Otherwise, there is a potential dependence. Return a clobber. 238 return MemDepResult::getClobber(Inst); 239 } 240 } 241 242 // No dependence found. If this is the entry block of the function, it is a 243 // clobber, otherwise it is non-local. 244 if (BB != &BB->getParent()->getEntryBlock()) 245 return MemDepResult::getNonLocal(); 246 return MemDepResult::getClobber(ScanIt); 247} 248 249/// getDependency - Return the instruction on which a memory operation 250/// depends. 251MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 252 Instruction *ScanPos = QueryInst; 253 254 // Check for a cached result 255 MemDepResult &LocalCache = LocalDeps[QueryInst]; 256 257 // If the cached entry is non-dirty, just return it. Note that this depends 258 // on MemDepResult's default constructing to 'dirty'. 259 if (!LocalCache.isDirty()) 260 return LocalCache; 261 262 // Otherwise, if we have a dirty entry, we know we can start the scan at that 263 // instruction, which may save us some work. 264 if (Instruction *Inst = LocalCache.getInst()) { 265 ScanPos = Inst; 266 267 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 268 } 269 270 BasicBlock *QueryParent = QueryInst->getParent(); 271 272 Value *MemPtr = 0; 273 uint64_t MemSize = 0; 274 275 // Do the scan. 276 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 277 // No dependence found. If this is the entry block of the function, it is a 278 // clobber, otherwise it is non-local. 279 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 280 LocalCache = MemDepResult::getNonLocal(); 281 else 282 LocalCache = MemDepResult::getClobber(QueryInst); 283 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) { 284 // If this is a volatile store, don't mess around with it. Just return the 285 // previous instruction as a clobber. 286 if (SI->isVolatile()) 287 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 288 else { 289 MemPtr = SI->getPointerOperand(); 290 MemSize = TD->getTypeStoreSize(SI->getOperand(0)->getType()); 291 } 292 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) { 293 // If this is a volatile load, don't mess around with it. Just return the 294 // previous instruction as a clobber. 295 if (LI->isVolatile()) 296 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 297 else { 298 MemPtr = LI->getPointerOperand(); 299 MemSize = TD->getTypeStoreSize(LI->getType()); 300 } 301 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 302 CallSite QueryCS = CallSite::get(QueryInst); 303 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 304 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 305 QueryParent); 306 } else if (FreeInst *FI = dyn_cast<FreeInst>(QueryInst)) { 307 MemPtr = FI->getPointerOperand(); 308 // FreeInsts erase the entire structure, not just a field. 309 MemSize = ~0UL; 310 } else { 311 // Non-memory instruction. 312 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 313 } 314 315 // If we need to do a pointer scan, make it happen. 316 if (MemPtr) 317 LocalCache = getPointerDependencyFrom(MemPtr, MemSize, 318 isa<LoadInst>(QueryInst), 319 ScanPos, QueryParent); 320 321 // Remember the result! 322 if (Instruction *I = LocalCache.getInst()) 323 ReverseLocalDeps[I].insert(QueryInst); 324 325 return LocalCache; 326} 327 328/// getNonLocalCallDependency - Perform a full dependency query for the 329/// specified call, returning the set of blocks that the value is 330/// potentially live across. The returned set of results will include a 331/// "NonLocal" result for all blocks where the value is live across. 332/// 333/// This method assumes the instruction returns a "NonLocal" dependency 334/// within its own block. 335/// 336/// This returns a reference to an internal data structure that may be 337/// invalidated on the next non-local query or when an instruction is 338/// removed. Clients must copy this data if they want it around longer than 339/// that. 340const MemoryDependenceAnalysis::NonLocalDepInfo & 341MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 342 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 343 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 344 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 345 NonLocalDepInfo &Cache = CacheP.first; 346 347 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 348 /// the cached case, this can happen due to instructions being deleted etc. In 349 /// the uncached case, this starts out as the set of predecessors we care 350 /// about. 351 SmallVector<BasicBlock*, 32> DirtyBlocks; 352 353 if (!Cache.empty()) { 354 // Okay, we have a cache entry. If we know it is not dirty, just return it 355 // with no computation. 356 if (!CacheP.second) { 357 NumCacheNonLocal++; 358 return Cache; 359 } 360 361 // If we already have a partially computed set of results, scan them to 362 // determine what is dirty, seeding our initial DirtyBlocks worklist. 363 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 364 I != E; ++I) 365 if (I->second.isDirty()) 366 DirtyBlocks.push_back(I->first); 367 368 // Sort the cache so that we can do fast binary search lookups below. 369 std::sort(Cache.begin(), Cache.end()); 370 371 ++NumCacheDirtyNonLocal; 372 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 373 // << Cache.size() << " cached: " << *QueryInst; 374 } else { 375 // Seed DirtyBlocks with each of the preds of QueryInst's block. 376 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 377 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 378 DirtyBlocks.push_back(*PI); 379 NumUncacheNonLocal++; 380 } 381 382 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 383 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 384 385 // Visited checked first, vector in sorted order. 386 SmallPtrSet<BasicBlock*, 64> Visited; 387 388 unsigned NumSortedEntries = Cache.size(); 389 390 // Iterate while we still have blocks to update. 391 while (!DirtyBlocks.empty()) { 392 BasicBlock *DirtyBB = DirtyBlocks.back(); 393 DirtyBlocks.pop_back(); 394 395 // Already processed this block? 396 if (!Visited.insert(DirtyBB)) 397 continue; 398 399 // Do a binary search to see if we already have an entry for this block in 400 // the cache set. If so, find it. 401 NonLocalDepInfo::iterator Entry = 402 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 403 std::make_pair(DirtyBB, MemDepResult())); 404 if (Entry != Cache.begin() && (&*Entry)[-1].first == DirtyBB) 405 --Entry; 406 407 MemDepResult *ExistingResult = 0; 408 if (Entry != Cache.begin()+NumSortedEntries && 409 Entry->first == DirtyBB) { 410 // If we already have an entry, and if it isn't already dirty, the block 411 // is done. 412 if (!Entry->second.isDirty()) 413 continue; 414 415 // Otherwise, remember this slot so we can update the value. 416 ExistingResult = &Entry->second; 417 } 418 419 // If the dirty entry has a pointer, start scanning from it so we don't have 420 // to rescan the entire block. 421 BasicBlock::iterator ScanPos = DirtyBB->end(); 422 if (ExistingResult) { 423 if (Instruction *Inst = ExistingResult->getInst()) { 424 ScanPos = Inst; 425 // We're removing QueryInst's use of Inst. 426 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 427 QueryCS.getInstruction()); 428 } 429 } 430 431 // Find out if this block has a local dependency for QueryInst. 432 MemDepResult Dep; 433 434 if (ScanPos != DirtyBB->begin()) { 435 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 436 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 437 // No dependence found. If this is the entry block of the function, it is 438 // a clobber, otherwise it is non-local. 439 Dep = MemDepResult::getNonLocal(); 440 } else { 441 Dep = MemDepResult::getClobber(ScanPos); 442 } 443 444 // If we had a dirty entry for the block, update it. Otherwise, just add 445 // a new entry. 446 if (ExistingResult) 447 *ExistingResult = Dep; 448 else 449 Cache.push_back(std::make_pair(DirtyBB, Dep)); 450 451 // If the block has a dependency (i.e. it isn't completely transparent to 452 // the value), remember the association! 453 if (!Dep.isNonLocal()) { 454 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 455 // update this when we remove instructions. 456 if (Instruction *Inst = Dep.getInst()) 457 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 458 } else { 459 460 // If the block *is* completely transparent to the load, we need to check 461 // the predecessors of this block. Add them to our worklist. 462 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 463 DirtyBlocks.push_back(*PI); 464 } 465 } 466 467 return Cache; 468} 469 470/// getNonLocalPointerDependency - Perform a full dependency query for an 471/// access to the specified (non-volatile) memory location, returning the 472/// set of instructions that either define or clobber the value. 473/// 474/// This method assumes the pointer has a "NonLocal" dependency within its 475/// own block. 476/// 477void MemoryDependenceAnalysis:: 478getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB, 479 SmallVectorImpl<NonLocalDepEntry> &Result) { 480 assert(isa<PointerType>(Pointer->getType()) && 481 "Can't get pointer deps of a non-pointer!"); 482 Result.clear(); 483 484 // We know that the pointer value is live into FromBB find the def/clobbers 485 // from presecessors. 486 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType(); 487 uint64_t PointeeSize = TD->getTypeStoreSize(EltTy); 488 489 // While we have blocks to analyze, get their values. 490 SmallPtrSet<BasicBlock*, 64> Visited; 491 getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB, 492 Result, Visited); 493} 494 495/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 496/// Pointer/PointeeSize using either cached information in Cache or by doing a 497/// lookup (which may use dirty cache info if available). If we do a lookup, 498/// add the result to the cache. 499MemDepResult MemoryDependenceAnalysis:: 500GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, 501 bool isLoad, BasicBlock *BB, 502 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 503 504 // Do a binary search to see if we already have an entry for this block in 505 // the cache set. If so, find it. 506 NonLocalDepInfo::iterator Entry = 507 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 508 std::make_pair(BB, MemDepResult())); 509 if (Entry != Cache->begin() && (&*Entry)[-1].first == BB) 510 --Entry; 511 512 MemDepResult *ExistingResult = 0; 513 if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB) 514 ExistingResult = &Entry->second; 515 516 // If we have a cached entry, and it is non-dirty, use it as the value for 517 // this dependency. 518 if (ExistingResult && !ExistingResult->isDirty()) { 519 ++NumCacheNonLocalPtr; 520 return *ExistingResult; 521 } 522 523 // Otherwise, we have to scan for the value. If we have a dirty cache 524 // entry, start scanning from its position, otherwise we scan from the end 525 // of the block. 526 BasicBlock::iterator ScanPos = BB->end(); 527 if (ExistingResult && ExistingResult->getInst()) { 528 assert(ExistingResult->getInst()->getParent() == BB && 529 "Instruction invalidated?"); 530 ++NumCacheDirtyNonLocalPtr; 531 ScanPos = ExistingResult->getInst(); 532 533 // Eliminating the dirty entry from 'Cache', so update the reverse info. 534 ValueIsLoadPair CacheKey(Pointer, isLoad); 535 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, 536 CacheKey.getOpaqueValue()); 537 } else { 538 ++NumUncacheNonLocalPtr; 539 } 540 541 // Scan the block for the dependency. 542 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, 543 ScanPos, BB); 544 545 // If we had a dirty entry for the block, update it. Otherwise, just add 546 // a new entry. 547 if (ExistingResult) 548 *ExistingResult = Dep; 549 else 550 Cache->push_back(std::make_pair(BB, Dep)); 551 552 // If the block has a dependency (i.e. it isn't completely transparent to 553 // the value), remember the reverse association because we just added it 554 // to Cache! 555 if (Dep.isNonLocal()) 556 return Dep; 557 558 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 559 // update MemDep when we remove instructions. 560 Instruction *Inst = Dep.getInst(); 561 assert(Inst && "Didn't depend on anything?"); 562 ValueIsLoadPair CacheKey(Pointer, isLoad); 563 ReverseNonLocalPtrDeps[Inst].insert(CacheKey.getOpaqueValue()); 564 return Dep; 565} 566 567 568/// getNonLocalPointerDepFromBB - 569void MemoryDependenceAnalysis:: 570getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize, 571 bool isLoad, BasicBlock *StartBB, 572 SmallVectorImpl<NonLocalDepEntry> &Result, 573 SmallPtrSet<BasicBlock*, 64> &Visited) { 574 // Look up the cached info for Pointer. 575 ValueIsLoadPair CacheKey(Pointer, isLoad); 576 577 std::pair<BasicBlock*, NonLocalDepInfo> &CacheInfo = 578 NonLocalPointerDeps[CacheKey]; 579 NonLocalDepInfo *Cache = &CacheInfo.second; 580 581 // If we have valid cached information for exactly the block we are 582 // investigating, just return it with no recomputation. 583 if (CacheInfo.first == StartBB) { 584 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 585 I != E; ++I) 586 if (!I->second.isNonLocal()) 587 Result.push_back(*I); 588 ++NumCacheCompleteNonLocalPtr; 589 return; 590 } 591 592 // Otherwise, either this is a new block, a block with an invalid cache 593 // pointer or one that we're about to invalidate by putting more info into it 594 // than its valid cache info. If empty, the result will be valid cache info, 595 // otherwise it isn't. 596 CacheInfo.first = Cache->empty() ? StartBB : 0; 597 598 SmallVector<BasicBlock*, 32> Worklist; 599 Worklist.push_back(StartBB); 600 601 // Keep track of the entries that we know are sorted. Previously cached 602 // entries will all be sorted. The entries we add we only sort on demand (we 603 // don't insert every element into its sorted position). We know that we 604 // won't get any reuse from currently inserted values, because we don't 605 // revisit blocks after we insert info for them. 606 unsigned NumSortedEntries = Cache->size(); 607 608 // SkipFirstBlock - If this is the very first block that we're processing, we 609 // don't want to scan or think about its body, because the client was supposed 610 // to do a local dependence query. Instead, just start processing it by 611 // adding its predecessors to the worklist and iterating. 612 bool SkipFirstBlock = Visited.empty(); 613 614 while (!Worklist.empty()) { 615 BasicBlock *BB = Worklist.pop_back_val(); 616 617 // Skip the first block if we have it. 618 if (SkipFirstBlock) { 619 SkipFirstBlock = false; 620 } else { 621 // Analyze the dependency of *Pointer in FromBB. See if we already have 622 // been here. 623 if (!Visited.insert(BB)) 624 continue; 625 626 // Get the dependency info for Pointer in BB. If we have cached 627 // information, we will use it, otherwise we compute it. 628 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad, 629 BB, Cache, NumSortedEntries); 630 631 // If we got a Def or Clobber, add this to the list of results. 632 if (!Dep.isNonLocal()) { 633 Result.push_back(NonLocalDepEntry(BB, Dep)); 634 continue; 635 } 636 } 637 638 // Otherwise, we have to process all the predecessors of this block to scan 639 // them as well. 640 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 641 // TODO: PHI TRANSLATE. 642 Worklist.push_back(*PI); 643 } 644 } 645 646 // Okay, we're done now. If we added new values to the cache, re-sort it. 647 switch (Cache->size()-NumSortedEntries) { 648 case 0: 649 // done, no new entries. 650 break; 651 case 2: { 652 // Two new entries, insert the last one into place. 653 NonLocalDepEntry Val = Cache->back(); 654 Cache->pop_back(); 655 NonLocalDepInfo::iterator Entry = 656 std::upper_bound(Cache->begin(), Cache->end()-1, Val); 657 Cache->insert(Entry, Val); 658 // FALL THROUGH. 659 } 660 case 1: { 661 // One new entry, Just insert the new value at the appropriate position. 662 NonLocalDepEntry Val = Cache->back(); 663 Cache->pop_back(); 664 NonLocalDepInfo::iterator Entry = 665 std::upper_bound(Cache->begin(), Cache->end(), Val); 666 Cache->insert(Entry, Val); 667 break; 668 } 669 default: 670 // Added many values, do a full scale sort. 671 std::sort(Cache->begin(), Cache->end()); 672 } 673} 674 675/// RemoveCachedNonLocalPointerDependencies - If P exists in 676/// CachedNonLocalPointerInfo, remove it. 677void MemoryDependenceAnalysis:: 678RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 679 CachedNonLocalPointerInfo::iterator It = 680 NonLocalPointerDeps.find(P); 681 if (It == NonLocalPointerDeps.end()) return; 682 683 // Remove all of the entries in the BB->val map. This involves removing 684 // instructions from the reverse map. 685 NonLocalDepInfo &PInfo = It->second.second; 686 687 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 688 Instruction *Target = PInfo[i].second.getInst(); 689 if (Target == 0) continue; // Ignore non-local dep results. 690 assert(Target->getParent() == PInfo[i].first && Target != P.getPointer()); 691 692 // Eliminating the dirty entry from 'Cache', so update the reverse info. 693 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P.getOpaqueValue()); 694 } 695 696 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 697 NonLocalPointerDeps.erase(It); 698} 699 700 701/// removeInstruction - Remove an instruction from the dependence analysis, 702/// updating the dependence of instructions that previously depended on it. 703/// This method attempts to keep the cache coherent using the reverse map. 704void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 705 // Walk through the Non-local dependencies, removing this one as the value 706 // for any cached queries. 707 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 708 if (NLDI != NonLocalDeps.end()) { 709 NonLocalDepInfo &BlockMap = NLDI->second.first; 710 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 711 DI != DE; ++DI) 712 if (Instruction *Inst = DI->second.getInst()) 713 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 714 NonLocalDeps.erase(NLDI); 715 } 716 717 // If we have a cached local dependence query for this instruction, remove it. 718 // 719 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 720 if (LocalDepEntry != LocalDeps.end()) { 721 // Remove us from DepInst's reverse set now that the local dep info is gone. 722 if (Instruction *Inst = LocalDepEntry->second.getInst()) 723 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 724 725 // Remove this local dependency info. 726 LocalDeps.erase(LocalDepEntry); 727 } 728 729 // If we have any cached pointer dependencies on this instruction, remove 730 // them. If the instruction has non-pointer type, then it can't be a pointer 731 // base. 732 733 // Remove it from both the load info and the store info. The instruction 734 // can't be in either of these maps if it is non-pointer. 735 if (isa<PointerType>(RemInst->getType())) { 736 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 737 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 738 } 739 740 // Loop over all of the things that depend on the instruction we're removing. 741 // 742 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 743 744 // If we find RemInst as a clobber or Def in any of the maps for other values, 745 // we need to replace its entry with a dirty version of the instruction after 746 // it. If RemInst is a terminator, we use a null dirty value. 747 // 748 // Using a dirty version of the instruction after RemInst saves having to scan 749 // the entire block to get to this point. 750 MemDepResult NewDirtyVal; 751 if (!RemInst->isTerminator()) 752 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 753 754 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 755 if (ReverseDepIt != ReverseLocalDeps.end()) { 756 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 757 // RemInst can't be the terminator if it has local stuff depending on it. 758 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 759 "Nothing can locally depend on a terminator"); 760 761 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 762 E = ReverseDeps.end(); I != E; ++I) { 763 Instruction *InstDependingOnRemInst = *I; 764 assert(InstDependingOnRemInst != RemInst && 765 "Already removed our local dep info"); 766 767 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 768 769 // Make sure to remember that new things depend on NewDepInst. 770 assert(NewDirtyVal.getInst() && "There is no way something else can have " 771 "a local dep on this if it is a terminator!"); 772 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 773 InstDependingOnRemInst)); 774 } 775 776 ReverseLocalDeps.erase(ReverseDepIt); 777 778 // Add new reverse deps after scanning the set, to avoid invalidating the 779 // 'ReverseDeps' reference. 780 while (!ReverseDepsToAdd.empty()) { 781 ReverseLocalDeps[ReverseDepsToAdd.back().first] 782 .insert(ReverseDepsToAdd.back().second); 783 ReverseDepsToAdd.pop_back(); 784 } 785 } 786 787 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 788 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 789 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 790 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 791 I != E; ++I) { 792 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 793 794 PerInstNLInfo &INLD = NonLocalDeps[*I]; 795 // The information is now dirty! 796 INLD.second = true; 797 798 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 799 DE = INLD.first.end(); DI != DE; ++DI) { 800 if (DI->second.getInst() != RemInst) continue; 801 802 // Convert to a dirty entry for the subsequent instruction. 803 DI->second = NewDirtyVal; 804 805 if (Instruction *NextI = NewDirtyVal.getInst()) 806 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 807 } 808 } 809 810 ReverseNonLocalDeps.erase(ReverseDepIt); 811 812 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 813 while (!ReverseDepsToAdd.empty()) { 814 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 815 .insert(ReverseDepsToAdd.back().second); 816 ReverseDepsToAdd.pop_back(); 817 } 818 } 819 820 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 821 // value in the NonLocalPointerDeps info. 822 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 823 ReverseNonLocalPtrDeps.find(RemInst); 824 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 825 SmallPtrSet<void*, 4> &Set = ReversePtrDepIt->second; 826 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 827 828 for (SmallPtrSet<void*, 4>::iterator I = Set.begin(), E = Set.end(); 829 I != E; ++I) { 830 ValueIsLoadPair P; 831 P.setFromOpaqueValue(*I); 832 assert(P.getPointer() != RemInst && 833 "Already removed NonLocalPointerDeps info for RemInst"); 834 835 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second; 836 837 // The cache is not valid for any specific block anymore. 838 NonLocalPointerDeps[P].first = 0; 839 840 // Update any entries for RemInst to use the instruction after it. 841 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 842 DI != DE; ++DI) { 843 if (DI->second.getInst() != RemInst) continue; 844 845 // Convert to a dirty entry for the subsequent instruction. 846 DI->second = NewDirtyVal; 847 848 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 849 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 850 } 851 } 852 853 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 854 855 while (!ReversePtrDepsToAdd.empty()) { 856 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 857 .insert(ReversePtrDepsToAdd.back().second.getOpaqueValue()); 858 ReversePtrDepsToAdd.pop_back(); 859 } 860 } 861 862 863 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 864 AA->deleteValue(RemInst); 865 DEBUG(verifyRemoved(RemInst)); 866} 867 868/// verifyRemoved - Verify that the specified instruction does not occur 869/// in our internal data structures. 870void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 871 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 872 E = LocalDeps.end(); I != E; ++I) { 873 assert(I->first != D && "Inst occurs in data structures"); 874 assert(I->second.getInst() != D && 875 "Inst occurs in data structures"); 876 } 877 878 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 879 E = NonLocalPointerDeps.end(); I != E; ++I) { 880 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 881 const NonLocalDepInfo &Val = I->second.second; 882 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 883 II != E; ++II) 884 assert(II->second.getInst() != D && "Inst occurs as NLPD value"); 885 } 886 887 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 888 E = NonLocalDeps.end(); I != E; ++I) { 889 assert(I->first != D && "Inst occurs in data structures"); 890 const PerInstNLInfo &INLD = I->second; 891 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 892 EE = INLD.first.end(); II != EE; ++II) 893 assert(II->second.getInst() != D && "Inst occurs in data structures"); 894 } 895 896 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 897 E = ReverseLocalDeps.end(); I != E; ++I) { 898 assert(I->first != D && "Inst occurs in data structures"); 899 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 900 EE = I->second.end(); II != EE; ++II) 901 assert(*II != D && "Inst occurs in data structures"); 902 } 903 904 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 905 E = ReverseNonLocalDeps.end(); 906 I != E; ++I) { 907 assert(I->first != D && "Inst occurs in data structures"); 908 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 909 EE = I->second.end(); II != EE; ++II) 910 assert(*II != D && "Inst occurs in data structures"); 911 } 912 913 for (ReverseNonLocalPtrDepTy::const_iterator 914 I = ReverseNonLocalPtrDeps.begin(), 915 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 916 assert(I->first != D && "Inst occurs in rev NLPD map"); 917 918 for (SmallPtrSet<void*, 4>::const_iterator II = I->second.begin(), 919 E = I->second.end(); II != E; ++II) 920 assert(*II != ValueIsLoadPair(D, false).getOpaqueValue() && 921 *II != ValueIsLoadPair(D, true).getOpaqueValue() && 922 "Inst occurs in ReverseNonLocalPtrDeps map"); 923 } 924 925} 926