MemoryDependenceAnalysis.cpp revision 9e59c64c14cfe55e7cc9086c6bff8cfeecac361e
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/ADT/STLExtras.h" 25#include "llvm/Support/PredIteratorCache.h" 26#include "llvm/Support/Debug.h" 27#include "llvm/Target/TargetData.h" 28using namespace llvm; 29 30STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 31STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 32STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 33 34STATISTIC(NumCacheNonLocalPtr, 35 "Number of fully cached non-local ptr responses"); 36STATISTIC(NumCacheDirtyNonLocalPtr, 37 "Number of cached, but dirty, non-local ptr responses"); 38STATISTIC(NumUncacheNonLocalPtr, 39 "Number of uncached non-local ptr responses"); 40STATISTIC(NumCacheCompleteNonLocalPtr, 41 "Number of block queries that were completely cached"); 42 43char MemoryDependenceAnalysis::ID = 0; 44 45// Register this pass... 46static RegisterPass<MemoryDependenceAnalysis> X("memdep", 47 "Memory Dependence Analysis", false, true); 48 49MemoryDependenceAnalysis::MemoryDependenceAnalysis() 50: FunctionPass(&ID), PredCache(0) { 51} 52MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 53} 54 55/// Clean up memory in between runs 56void MemoryDependenceAnalysis::releaseMemory() { 57 LocalDeps.clear(); 58 NonLocalDeps.clear(); 59 NonLocalPointerDeps.clear(); 60 ReverseLocalDeps.clear(); 61 ReverseNonLocalDeps.clear(); 62 ReverseNonLocalPtrDeps.clear(); 63 PredCache->clear(); 64} 65 66 67 68/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 69/// 70void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 71 AU.setPreservesAll(); 72 AU.addRequiredTransitive<AliasAnalysis>(); 73 AU.addRequiredTransitive<TargetData>(); 74} 75 76bool MemoryDependenceAnalysis::runOnFunction(Function &) { 77 AA = &getAnalysis<AliasAnalysis>(); 78 TD = &getAnalysis<TargetData>(); 79 if (PredCache == 0) 80 PredCache.reset(new PredIteratorCache()); 81 return false; 82} 83 84/// RemoveFromReverseMap - This is a helper function that removes Val from 85/// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry. 86template <typename KeyTy> 87static void RemoveFromReverseMap(DenseMap<Instruction*, 88 SmallPtrSet<KeyTy*, 4> > &ReverseMap, 89 Instruction *Inst, KeyTy *Val) { 90 typename DenseMap<Instruction*, SmallPtrSet<KeyTy*, 4> >::iterator 91 InstIt = ReverseMap.find(Inst); 92 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); 93 bool Found = InstIt->second.erase(Val); 94 assert(Found && "Invalid reverse map!"); Found=Found; 95 if (InstIt->second.empty()) 96 ReverseMap.erase(InstIt); 97} 98 99 100/// getCallSiteDependencyFrom - Private helper for finding the local 101/// dependencies of a call site. 102MemDepResult MemoryDependenceAnalysis:: 103getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 104 BasicBlock::iterator ScanIt, BasicBlock *BB) { 105 // Walk backwards through the block, looking for dependencies 106 while (ScanIt != BB->begin()) { 107 Instruction *Inst = --ScanIt; 108 109 // If this inst is a memory op, get the pointer it accessed 110 Value *Pointer = 0; 111 uint64_t PointerSize = 0; 112 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) { 113 Pointer = S->getPointerOperand(); 114 PointerSize = TD->getTypeStoreSize(S->getOperand(0)->getType()); 115 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 116 Pointer = V->getOperand(0); 117 PointerSize = TD->getTypeStoreSize(V->getType()); 118 } else if (FreeInst *F = dyn_cast<FreeInst>(Inst)) { 119 Pointer = F->getPointerOperand(); 120 121 // FreeInsts erase the entire structure 122 PointerSize = ~0ULL; 123 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) { 124 CallSite InstCS = CallSite::get(Inst); 125 // If these two calls do not interfere, look past it. 126 switch (AA->getModRefInfo(CS, InstCS)) { 127 case AliasAnalysis::NoModRef: 128 // If the two calls don't interact (e.g. InstCS is readnone) keep 129 // scanning. 130 continue; 131 case AliasAnalysis::Ref: 132 // If the two calls read the same memory locations and CS is a readonly 133 // function, then we have two cases: 1) the calls may not interfere with 134 // each other at all. 2) the calls may produce the same value. In case 135 // #1 we want to ignore the values, in case #2, we want to return Inst 136 // as a Def dependence. This allows us to CSE in cases like: 137 // X = strlen(P); 138 // memchr(...); 139 // Y = strlen(P); // Y = X 140 if (isReadOnlyCall) { 141 if (CS.getCalledFunction() != 0 && 142 CS.getCalledFunction() == InstCS.getCalledFunction()) 143 return MemDepResult::getDef(Inst); 144 // Ignore unrelated read/read call dependences. 145 continue; 146 } 147 // FALL THROUGH 148 default: 149 return MemDepResult::getClobber(Inst); 150 } 151 } else { 152 // Non-memory instruction. 153 continue; 154 } 155 156 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef) 157 return MemDepResult::getClobber(Inst); 158 } 159 160 // No dependence found. If this is the entry block of the function, it is a 161 // clobber, otherwise it is non-local. 162 if (BB != &BB->getParent()->getEntryBlock()) 163 return MemDepResult::getNonLocal(); 164 return MemDepResult::getClobber(ScanIt); 165} 166 167/// getPointerDependencyFrom - Return the instruction on which a memory 168/// location depends. If isLoad is true, this routine ignore may-aliases with 169/// read-only operations. 170MemDepResult MemoryDependenceAnalysis:: 171getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, 172 BasicBlock::iterator ScanIt, BasicBlock *BB) { 173 174 // Walk backwards through the basic block, looking for dependencies. 175 while (ScanIt != BB->begin()) { 176 Instruction *Inst = --ScanIt; 177 178 // Values depend on loads if the pointers are must aliased. This means that 179 // a load depends on another must aliased load from the same value. 180 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 181 Value *Pointer = LI->getPointerOperand(); 182 uint64_t PointerSize = TD->getTypeStoreSize(LI->getType()); 183 184 // If we found a pointer, check if it could be the same as our pointer. 185 AliasAnalysis::AliasResult R = 186 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 187 if (R == AliasAnalysis::NoAlias) 188 continue; 189 190 // May-alias loads don't depend on each other without a dependence. 191 if (isLoad && R == AliasAnalysis::MayAlias) 192 continue; 193 // Stores depend on may and must aliased loads, loads depend on must-alias 194 // loads. 195 return MemDepResult::getDef(Inst); 196 } 197 198 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 199 Value *Pointer = SI->getPointerOperand(); 200 uint64_t PointerSize = TD->getTypeStoreSize(SI->getOperand(0)->getType()); 201 202 // If we found a pointer, check if it could be the same as our pointer. 203 AliasAnalysis::AliasResult R = 204 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 205 206 if (R == AliasAnalysis::NoAlias) 207 continue; 208 if (R == AliasAnalysis::MayAlias) 209 return MemDepResult::getClobber(Inst); 210 return MemDepResult::getDef(Inst); 211 } 212 213 // If this is an allocation, and if we know that the accessed pointer is to 214 // the allocation, return Def. This means that there is no dependence and 215 // the access can be optimized based on that. For example, a load could 216 // turn into undef. 217 if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) { 218 Value *AccessPtr = MemPtr->getUnderlyingObject(); 219 220 if (AccessPtr == AI || 221 AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 222 return MemDepResult::getDef(AI); 223 continue; 224 } 225 226 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 227 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) { 228 case AliasAnalysis::NoModRef: 229 // If the call has no effect on the queried pointer, just ignore it. 230 continue; 231 case AliasAnalysis::Ref: 232 // If the call is known to never store to the pointer, and if this is a 233 // load query, we can safely ignore it (scan past it). 234 if (isLoad) 235 continue; 236 // FALL THROUGH. 237 default: 238 // Otherwise, there is a potential dependence. Return a clobber. 239 return MemDepResult::getClobber(Inst); 240 } 241 } 242 243 // No dependence found. If this is the entry block of the function, it is a 244 // clobber, otherwise it is non-local. 245 if (BB != &BB->getParent()->getEntryBlock()) 246 return MemDepResult::getNonLocal(); 247 return MemDepResult::getClobber(ScanIt); 248} 249 250/// getDependency - Return the instruction on which a memory operation 251/// depends. 252MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 253 Instruction *ScanPos = QueryInst; 254 255 // Check for a cached result 256 MemDepResult &LocalCache = LocalDeps[QueryInst]; 257 258 // If the cached entry is non-dirty, just return it. Note that this depends 259 // on MemDepResult's default constructing to 'dirty'. 260 if (!LocalCache.isDirty()) 261 return LocalCache; 262 263 // Otherwise, if we have a dirty entry, we know we can start the scan at that 264 // instruction, which may save us some work. 265 if (Instruction *Inst = LocalCache.getInst()) { 266 ScanPos = Inst; 267 268 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 269 } 270 271 BasicBlock *QueryParent = QueryInst->getParent(); 272 273 Value *MemPtr = 0; 274 uint64_t MemSize = 0; 275 276 // Do the scan. 277 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 278 // No dependence found. If this is the entry block of the function, it is a 279 // clobber, otherwise it is non-local. 280 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 281 LocalCache = MemDepResult::getNonLocal(); 282 else 283 LocalCache = MemDepResult::getClobber(QueryInst); 284 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) { 285 // If this is a volatile store, don't mess around with it. Just return the 286 // previous instruction as a clobber. 287 if (SI->isVolatile()) 288 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 289 else { 290 MemPtr = SI->getPointerOperand(); 291 MemSize = TD->getTypeStoreSize(SI->getOperand(0)->getType()); 292 } 293 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) { 294 // If this is a volatile load, don't mess around with it. Just return the 295 // previous instruction as a clobber. 296 if (LI->isVolatile()) 297 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 298 else { 299 MemPtr = LI->getPointerOperand(); 300 MemSize = TD->getTypeStoreSize(LI->getType()); 301 } 302 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 303 CallSite QueryCS = CallSite::get(QueryInst); 304 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 305 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 306 QueryParent); 307 } else if (FreeInst *FI = dyn_cast<FreeInst>(QueryInst)) { 308 MemPtr = FI->getPointerOperand(); 309 // FreeInsts erase the entire structure, not just a field. 310 MemSize = ~0UL; 311 } else { 312 // Non-memory instruction. 313 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 314 } 315 316 // If we need to do a pointer scan, make it happen. 317 if (MemPtr) 318 LocalCache = getPointerDependencyFrom(MemPtr, MemSize, 319 isa<LoadInst>(QueryInst), 320 ScanPos, QueryParent); 321 322 // Remember the result! 323 if (Instruction *I = LocalCache.getInst()) 324 ReverseLocalDeps[I].insert(QueryInst); 325 326 return LocalCache; 327} 328 329/// getNonLocalCallDependency - Perform a full dependency query for the 330/// specified call, returning the set of blocks that the value is 331/// potentially live across. The returned set of results will include a 332/// "NonLocal" result for all blocks where the value is live across. 333/// 334/// This method assumes the instruction returns a "NonLocal" dependency 335/// within its own block. 336/// 337/// This returns a reference to an internal data structure that may be 338/// invalidated on the next non-local query or when an instruction is 339/// removed. Clients must copy this data if they want it around longer than 340/// that. 341const MemoryDependenceAnalysis::NonLocalDepInfo & 342MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 343 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 344 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 345 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 346 NonLocalDepInfo &Cache = CacheP.first; 347 348 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 349 /// the cached case, this can happen due to instructions being deleted etc. In 350 /// the uncached case, this starts out as the set of predecessors we care 351 /// about. 352 SmallVector<BasicBlock*, 32> DirtyBlocks; 353 354 if (!Cache.empty()) { 355 // Okay, we have a cache entry. If we know it is not dirty, just return it 356 // with no computation. 357 if (!CacheP.second) { 358 NumCacheNonLocal++; 359 return Cache; 360 } 361 362 // If we already have a partially computed set of results, scan them to 363 // determine what is dirty, seeding our initial DirtyBlocks worklist. 364 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 365 I != E; ++I) 366 if (I->second.isDirty()) 367 DirtyBlocks.push_back(I->first); 368 369 // Sort the cache so that we can do fast binary search lookups below. 370 std::sort(Cache.begin(), Cache.end()); 371 372 ++NumCacheDirtyNonLocal; 373 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 374 // << Cache.size() << " cached: " << *QueryInst; 375 } else { 376 // Seed DirtyBlocks with each of the preds of QueryInst's block. 377 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 378 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 379 DirtyBlocks.push_back(*PI); 380 NumUncacheNonLocal++; 381 } 382 383 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 384 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 385 386 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() && prior(Entry)->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 // This is the set of blocks we've inspected, and the pointer we consider in 490 // each block. Because of critical edges, we currently bail out if querying 491 // a block with multiple different pointers. This can happen during PHI 492 // translation. 493 DenseMap<BasicBlock*, Value*> Visited; 494 if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB, 495 Result, Visited, true)) 496 return; 497 Result.push_back(std::make_pair(FromBB, 498 MemDepResult::getClobber(FromBB->begin()))); 499} 500 501/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 502/// Pointer/PointeeSize using either cached information in Cache or by doing a 503/// lookup (which may use dirty cache info if available). If we do a lookup, 504/// add the result to the cache. 505MemDepResult MemoryDependenceAnalysis:: 506GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, 507 bool isLoad, BasicBlock *BB, 508 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 509 510 // Do a binary search to see if we already have an entry for this block in 511 // the cache set. If so, find it. 512 NonLocalDepInfo::iterator Entry = 513 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 514 std::make_pair(BB, MemDepResult())); 515 if (Entry != Cache->begin() && prior(Entry)->first == BB) 516 --Entry; 517 518 MemDepResult *ExistingResult = 0; 519 if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB) 520 ExistingResult = &Entry->second; 521 522 // If we have a cached entry, and it is non-dirty, use it as the value for 523 // this dependency. 524 if (ExistingResult && !ExistingResult->isDirty()) { 525 ++NumCacheNonLocalPtr; 526 return *ExistingResult; 527 } 528 529 // Otherwise, we have to scan for the value. If we have a dirty cache 530 // entry, start scanning from its position, otherwise we scan from the end 531 // of the block. 532 BasicBlock::iterator ScanPos = BB->end(); 533 if (ExistingResult && ExistingResult->getInst()) { 534 assert(ExistingResult->getInst()->getParent() == BB && 535 "Instruction invalidated?"); 536 ++NumCacheDirtyNonLocalPtr; 537 ScanPos = ExistingResult->getInst(); 538 539 // Eliminating the dirty entry from 'Cache', so update the reverse info. 540 ValueIsLoadPair CacheKey(Pointer, isLoad); 541 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, 542 CacheKey.getOpaqueValue()); 543 } else { 544 ++NumUncacheNonLocalPtr; 545 } 546 547 // Scan the block for the dependency. 548 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, 549 ScanPos, BB); 550 551 // If we had a dirty entry for the block, update it. Otherwise, just add 552 // a new entry. 553 if (ExistingResult) 554 *ExistingResult = Dep; 555 else 556 Cache->push_back(std::make_pair(BB, Dep)); 557 558 // If the block has a dependency (i.e. it isn't completely transparent to 559 // the value), remember the reverse association because we just added it 560 // to Cache! 561 if (Dep.isNonLocal()) 562 return Dep; 563 564 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 565 // update MemDep when we remove instructions. 566 Instruction *Inst = Dep.getInst(); 567 assert(Inst && "Didn't depend on anything?"); 568 ValueIsLoadPair CacheKey(Pointer, isLoad); 569 ReverseNonLocalPtrDeps[Inst].insert(CacheKey.getOpaqueValue()); 570 return Dep; 571} 572 573 574/// getNonLocalPointerDepFromBB - Perform a dependency query based on 575/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 576/// results to the results vector and keep track of which blocks are visited in 577/// 'Visited'. 578/// 579/// This has special behavior for the first block queries (when SkipFirstBlock 580/// is true). In this special case, it ignores the contents of the specified 581/// block and starts returning dependence info for its predecessors. 582/// 583/// This function returns false on success, or true to indicate that it could 584/// not compute dependence information for some reason. This should be treated 585/// as a clobber dependence on the first instruction in the predecessor block. 586bool MemoryDependenceAnalysis:: 587getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize, 588 bool isLoad, BasicBlock *StartBB, 589 SmallVectorImpl<NonLocalDepEntry> &Result, 590 DenseMap<BasicBlock*, Value*> &Visited, 591 bool SkipFirstBlock) { 592 593 // Look up the cached info for Pointer. 594 ValueIsLoadPair CacheKey(Pointer, isLoad); 595 596 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo = 597 &NonLocalPointerDeps[CacheKey]; 598 NonLocalDepInfo *Cache = &CacheInfo->second; 599 600 // If we have valid cached information for exactly the block we are 601 // investigating, just return it with no recomputation. 602 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 603 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 604 I != E; ++I) 605 if (!I->second.isNonLocal()) 606 Result.push_back(*I); 607 ++NumCacheCompleteNonLocalPtr; 608 return false; 609 } 610 611 // Otherwise, either this is a new block, a block with an invalid cache 612 // pointer or one that we're about to invalidate by putting more info into it 613 // than its valid cache info. If empty, the result will be valid cache info, 614 // otherwise it isn't. 615 if (Cache->empty()) 616 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 617 else 618 CacheInfo->first = BBSkipFirstBlockPair(); 619 620 SmallVector<BasicBlock*, 32> Worklist; 621 Worklist.push_back(StartBB); 622 623 // Keep track of the entries that we know are sorted. Previously cached 624 // entries will all be sorted. The entries we add we only sort on demand (we 625 // don't insert every element into its sorted position). We know that we 626 // won't get any reuse from currently inserted values, because we don't 627 // revisit blocks after we insert info for them. 628 unsigned NumSortedEntries = Cache->size(); 629 630 while (!Worklist.empty()) { 631 BasicBlock *BB = Worklist.pop_back_val(); 632 633 // Skip the first block if we have it. 634 if (!SkipFirstBlock) { 635 // Analyze the dependency of *Pointer in FromBB. See if we already have 636 // been here. 637 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 638 639 // Get the dependency info for Pointer in BB. If we have cached 640 // information, we will use it, otherwise we compute it. 641 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad, 642 BB, Cache, NumSortedEntries); 643 644 // If we got a Def or Clobber, add this to the list of results. 645 if (!Dep.isNonLocal()) { 646 Result.push_back(NonLocalDepEntry(BB, Dep)); 647 continue; 648 } 649 } 650 651 // If 'Pointer' is an instruction defined in this block, then we need to do 652 // phi translation to change it into a value live in the predecessor block. 653 // If phi translation fails, then we can't continue dependence analysis. 654 Instruction *PtrInst = dyn_cast<Instruction>(Pointer); 655 bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB; 656 657 // If no PHI translation is needed, just add all the predecessors of this 658 // block to scan them as well. 659 if (!NeedsPHITranslation) { 660 SkipFirstBlock = false; 661 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 662 // Verify that we haven't looked at this block yet. 663 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 664 InsertRes = Visited.insert(std::make_pair(*PI, Pointer)); 665 if (InsertRes.second) { 666 // First time we've looked at *PI. 667 Worklist.push_back(*PI); 668 continue; 669 } 670 671 // If we have seen this block before, but it was with a different 672 // pointer then we have a phi translation failure and we have to treat 673 // this as a clobber. 674 if (InsertRes.first->second != Pointer) 675 goto PredTranslationFailure; 676 } 677 continue; 678 } 679 680 // If we do need to do phi translation, then there are a bunch of different 681 // cases, because we have to find a Value* live in the predecessor block. We 682 // know that PtrInst is defined in this block at least. 683 684 // If this is directly a PHI node, just use the incoming values for each 685 // pred as the phi translated version. 686 if (PHINode *PtrPHI = dyn_cast<PHINode>(PtrInst)) { 687 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI){ 688 BasicBlock *Pred = *PI; 689 Value *PredPtr = PtrPHI->getIncomingValueForBlock(Pred); 690 691 // Check to see if we have already visited this pred block with another 692 // pointer. If so, we can't do this lookup. This failure can occur 693 // with PHI translation when a critical edge exists and the PHI node in 694 // the successor translates to a pointer value different than the 695 // pointer the block was first analyzed with. 696 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 697 InsertRes = Visited.insert(std::make_pair(Pred, PredPtr)); 698 699 if (!InsertRes.second) { 700 // If the predecessor was visited with PredPtr, then we already did 701 // the analysis and can ignore it. 702 if (InsertRes.first->second == PredPtr) 703 continue; 704 705 // Otherwise, the block was previously analyzed with a different 706 // pointer. We can't represent the result of this case, so we just 707 // treat this as a phi translation failure. 708 goto PredTranslationFailure; 709 } 710 711 // If we have a problem phi translating, fall through to the code below 712 // to handle the failure condition. 713 if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred, 714 Result, Visited)) 715 goto PredTranslationFailure; 716 } 717 718 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 719 CacheInfo = &NonLocalPointerDeps[CacheKey]; 720 Cache = &CacheInfo->second; 721 722 // Since we did phi translation, the "Cache" set won't contain all of the 723 // results for the query. This is ok (we can still use it to accelerate 724 // specific block queries) but we can't do the fastpath "return all 725 // results from the set" Clear out the indicator for this. 726 CacheInfo->first = BBSkipFirstBlockPair(); 727 SkipFirstBlock = false; 728 continue; 729 } 730 731 // TODO: BITCAST, GEP. 732 733 // cerr << "MEMDEP: Could not PHI translate: " << *Pointer; 734 // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst)) 735 // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0); 736 PredTranslationFailure: 737 738 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 739 CacheInfo = &NonLocalPointerDeps[CacheKey]; 740 Cache = &CacheInfo->second; 741 742 // Since we did phi translation, the "Cache" set won't contain all of the 743 // results for the query. This is ok (we can still use it to accelerate 744 // specific block queries) but we can't do the fastpath "return all 745 // results from the set" Clear out the indicator for this. 746 CacheInfo->first = BBSkipFirstBlockPair(); 747 748 // If *nothing* works, mark the pointer as being clobbered by the first 749 // instruction in this block. 750 // 751 // If this is the magic first block, return this as a clobber of the whole 752 // incoming value. Since we can't phi translate to one of the predecessors, 753 // we have to bail out. 754 if (SkipFirstBlock) 755 return true; 756 757 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 758 assert(I != Cache->rend() && "Didn't find current block??"); 759 if (I->first != BB) 760 continue; 761 762 assert(I->second.isNonLocal() && 763 "Should only be here with transparent block"); 764 I->second = MemDepResult::getClobber(BB->begin()); 765 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey.getOpaqueValue()); 766 Result.push_back(*I); 767 break; 768 } 769 } 770 771 // Okay, we're done now. If we added new values to the cache, re-sort it. 772 switch (Cache->size()-NumSortedEntries) { 773 case 0: 774 // done, no new entries. 775 break; 776 case 2: { 777 // Two new entries, insert the last one into place. 778 NonLocalDepEntry Val = Cache->back(); 779 Cache->pop_back(); 780 NonLocalDepInfo::iterator Entry = 781 std::upper_bound(Cache->begin(), Cache->end()-1, Val); 782 Cache->insert(Entry, Val); 783 // FALL THROUGH. 784 } 785 case 1: 786 // One new entry, Just insert the new value at the appropriate position. 787 if (Cache->size() != 1) { 788 NonLocalDepEntry Val = Cache->back(); 789 Cache->pop_back(); 790 NonLocalDepInfo::iterator Entry = 791 std::upper_bound(Cache->begin(), Cache->end(), Val); 792 Cache->insert(Entry, Val); 793 } 794 break; 795 default: 796 // Added many values, do a full scale sort. 797 std::sort(Cache->begin(), Cache->end()); 798 } 799 800 return false; 801} 802 803/// RemoveCachedNonLocalPointerDependencies - If P exists in 804/// CachedNonLocalPointerInfo, remove it. 805void MemoryDependenceAnalysis:: 806RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 807 CachedNonLocalPointerInfo::iterator It = 808 NonLocalPointerDeps.find(P); 809 if (It == NonLocalPointerDeps.end()) return; 810 811 // Remove all of the entries in the BB->val map. This involves removing 812 // instructions from the reverse map. 813 NonLocalDepInfo &PInfo = It->second.second; 814 815 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 816 Instruction *Target = PInfo[i].second.getInst(); 817 if (Target == 0) continue; // Ignore non-local dep results. 818 assert(Target->getParent() == PInfo[i].first); 819 820 // Eliminating the dirty entry from 'Cache', so update the reverse info. 821 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P.getOpaqueValue()); 822 } 823 824 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 825 NonLocalPointerDeps.erase(It); 826} 827 828 829/// invalidateCachedPointerInfo - This method is used to invalidate cached 830/// information about the specified pointer, because it may be too 831/// conservative in memdep. This is an optional call that can be used when 832/// the client detects an equivalence between the pointer and some other 833/// value and replaces the other value with ptr. This can make Ptr available 834/// in more places that cached info does not necessarily keep. 835void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 836 // If Ptr isn't really a pointer, just ignore it. 837 if (!isa<PointerType>(Ptr->getType())) return; 838 // Flush store info for the pointer. 839 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 840 // Flush load info for the pointer. 841 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 842} 843 844/// removeInstruction - Remove an instruction from the dependence analysis, 845/// updating the dependence of instructions that previously depended on it. 846/// This method attempts to keep the cache coherent using the reverse map. 847void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 848 // Walk through the Non-local dependencies, removing this one as the value 849 // for any cached queries. 850 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 851 if (NLDI != NonLocalDeps.end()) { 852 NonLocalDepInfo &BlockMap = NLDI->second.first; 853 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 854 DI != DE; ++DI) 855 if (Instruction *Inst = DI->second.getInst()) 856 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 857 NonLocalDeps.erase(NLDI); 858 } 859 860 // If we have a cached local dependence query for this instruction, remove it. 861 // 862 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 863 if (LocalDepEntry != LocalDeps.end()) { 864 // Remove us from DepInst's reverse set now that the local dep info is gone. 865 if (Instruction *Inst = LocalDepEntry->second.getInst()) 866 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 867 868 // Remove this local dependency info. 869 LocalDeps.erase(LocalDepEntry); 870 } 871 872 // If we have any cached pointer dependencies on this instruction, remove 873 // them. If the instruction has non-pointer type, then it can't be a pointer 874 // base. 875 876 // Remove it from both the load info and the store info. The instruction 877 // can't be in either of these maps if it is non-pointer. 878 if (isa<PointerType>(RemInst->getType())) { 879 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 880 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 881 } 882 883 // Loop over all of the things that depend on the instruction we're removing. 884 // 885 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 886 887 // If we find RemInst as a clobber or Def in any of the maps for other values, 888 // we need to replace its entry with a dirty version of the instruction after 889 // it. If RemInst is a terminator, we use a null dirty value. 890 // 891 // Using a dirty version of the instruction after RemInst saves having to scan 892 // the entire block to get to this point. 893 MemDepResult NewDirtyVal; 894 if (!RemInst->isTerminator()) 895 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 896 897 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 898 if (ReverseDepIt != ReverseLocalDeps.end()) { 899 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 900 // RemInst can't be the terminator if it has local stuff depending on it. 901 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 902 "Nothing can locally depend on a terminator"); 903 904 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 905 E = ReverseDeps.end(); I != E; ++I) { 906 Instruction *InstDependingOnRemInst = *I; 907 assert(InstDependingOnRemInst != RemInst && 908 "Already removed our local dep info"); 909 910 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 911 912 // Make sure to remember that new things depend on NewDepInst. 913 assert(NewDirtyVal.getInst() && "There is no way something else can have " 914 "a local dep on this if it is a terminator!"); 915 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 916 InstDependingOnRemInst)); 917 } 918 919 ReverseLocalDeps.erase(ReverseDepIt); 920 921 // Add new reverse deps after scanning the set, to avoid invalidating the 922 // 'ReverseDeps' reference. 923 while (!ReverseDepsToAdd.empty()) { 924 ReverseLocalDeps[ReverseDepsToAdd.back().first] 925 .insert(ReverseDepsToAdd.back().second); 926 ReverseDepsToAdd.pop_back(); 927 } 928 } 929 930 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 931 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 932 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 933 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 934 I != E; ++I) { 935 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 936 937 PerInstNLInfo &INLD = NonLocalDeps[*I]; 938 // The information is now dirty! 939 INLD.second = true; 940 941 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 942 DE = INLD.first.end(); DI != DE; ++DI) { 943 if (DI->second.getInst() != RemInst) continue; 944 945 // Convert to a dirty entry for the subsequent instruction. 946 DI->second = NewDirtyVal; 947 948 if (Instruction *NextI = NewDirtyVal.getInst()) 949 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 950 } 951 } 952 953 ReverseNonLocalDeps.erase(ReverseDepIt); 954 955 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 956 while (!ReverseDepsToAdd.empty()) { 957 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 958 .insert(ReverseDepsToAdd.back().second); 959 ReverseDepsToAdd.pop_back(); 960 } 961 } 962 963 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 964 // value in the NonLocalPointerDeps info. 965 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 966 ReverseNonLocalPtrDeps.find(RemInst); 967 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 968 SmallPtrSet<void*, 4> &Set = ReversePtrDepIt->second; 969 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 970 971 for (SmallPtrSet<void*, 4>::iterator I = Set.begin(), E = Set.end(); 972 I != E; ++I) { 973 ValueIsLoadPair P; 974 P.setFromOpaqueValue(*I); 975 assert(P.getPointer() != RemInst && 976 "Already removed NonLocalPointerDeps info for RemInst"); 977 978 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second; 979 980 // The cache is not valid for any specific block anymore. 981 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair(); 982 983 // Update any entries for RemInst to use the instruction after it. 984 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 985 DI != DE; ++DI) { 986 if (DI->second.getInst() != RemInst) continue; 987 988 // Convert to a dirty entry for the subsequent instruction. 989 DI->second = NewDirtyVal; 990 991 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 992 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 993 } 994 } 995 996 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 997 998 while (!ReversePtrDepsToAdd.empty()) { 999 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1000 .insert(ReversePtrDepsToAdd.back().second.getOpaqueValue()); 1001 ReversePtrDepsToAdd.pop_back(); 1002 } 1003 } 1004 1005 1006 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1007 AA->deleteValue(RemInst); 1008 DEBUG(verifyRemoved(RemInst)); 1009} 1010/// verifyRemoved - Verify that the specified instruction does not occur 1011/// in our internal data structures. 1012void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1013 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1014 E = LocalDeps.end(); I != E; ++I) { 1015 assert(I->first != D && "Inst occurs in data structures"); 1016 assert(I->second.getInst() != D && 1017 "Inst occurs in data structures"); 1018 } 1019 1020 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1021 E = NonLocalPointerDeps.end(); I != E; ++I) { 1022 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1023 const NonLocalDepInfo &Val = I->second.second; 1024 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1025 II != E; ++II) 1026 assert(II->second.getInst() != D && "Inst occurs as NLPD value"); 1027 } 1028 1029 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1030 E = NonLocalDeps.end(); I != E; ++I) { 1031 assert(I->first != D && "Inst occurs in data structures"); 1032 const PerInstNLInfo &INLD = I->second; 1033 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1034 EE = INLD.first.end(); II != EE; ++II) 1035 assert(II->second.getInst() != D && "Inst occurs in data structures"); 1036 } 1037 1038 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1039 E = ReverseLocalDeps.end(); I != E; ++I) { 1040 assert(I->first != D && "Inst occurs in data structures"); 1041 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1042 EE = I->second.end(); II != EE; ++II) 1043 assert(*II != D && "Inst occurs in data structures"); 1044 } 1045 1046 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1047 E = ReverseNonLocalDeps.end(); 1048 I != E; ++I) { 1049 assert(I->first != D && "Inst occurs in data structures"); 1050 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1051 EE = I->second.end(); II != EE; ++II) 1052 assert(*II != D && "Inst occurs in data structures"); 1053 } 1054 1055 for (ReverseNonLocalPtrDepTy::const_iterator 1056 I = ReverseNonLocalPtrDeps.begin(), 1057 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1058 assert(I->first != D && "Inst occurs in rev NLPD map"); 1059 1060 for (SmallPtrSet<void*, 4>::const_iterator II = I->second.begin(), 1061 E = I->second.end(); II != E; ++II) 1062 assert(*II != ValueIsLoadPair(D, false).getOpaqueValue() && 1063 *II != ValueIsLoadPair(D, true).getOpaqueValue() && 1064 "Inst occurs in ReverseNonLocalPtrDeps map"); 1065 } 1066 1067} 1068