MemoryDependenceAnalysis.cpp revision d777d405cdda8d418ba8e8818e5c1272dfd999a0
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/CFG.h" 26#include "llvm/Support/CommandLine.h" 27#include "llvm/Support/Debug.h" 28#include "llvm/Target/TargetData.h" 29using namespace llvm; 30 31STATISTIC(NumCacheNonLocal, "Number of cached non-local responses"); 32STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 33 34char MemoryDependenceAnalysis::ID = 0; 35 36// Register this pass... 37static RegisterPass<MemoryDependenceAnalysis> X("memdep", 38 "Memory Dependence Analysis", false, true); 39 40/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 41/// 42void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 43 AU.setPreservesAll(); 44 AU.addRequiredTransitive<AliasAnalysis>(); 45 AU.addRequiredTransitive<TargetData>(); 46} 47 48bool MemoryDependenceAnalysis::runOnFunction(Function &) { 49 AA = &getAnalysis<AliasAnalysis>(); 50 TD = &getAnalysis<TargetData>(); 51 return false; 52} 53 54/// getCallSiteDependency - Private helper for finding the local dependencies 55/// of a call site. 56MemoryDependenceAnalysis::DepResultTy MemoryDependenceAnalysis:: 57getCallSiteDependency(CallSite C, BasicBlock::iterator ScanIt, 58 BasicBlock *BB) { 59 // Walk backwards through the block, looking for dependencies 60 while (ScanIt != BB->begin()) { 61 Instruction *Inst = --ScanIt; 62 63 // If this inst is a memory op, get the pointer it accessed 64 Value *Pointer = 0; 65 uint64_t PointerSize = 0; 66 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) { 67 Pointer = S->getPointerOperand(); 68 PointerSize = TD->getTypeStoreSize(S->getOperand(0)->getType()); 69 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 70 Pointer = V->getOperand(0); 71 PointerSize = TD->getTypeStoreSize(V->getType()); 72 } else if (FreeInst *F = dyn_cast<FreeInst>(Inst)) { 73 Pointer = F->getPointerOperand(); 74 75 // FreeInsts erase the entire structure 76 PointerSize = ~0UL; 77 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) { 78 if (AA->getModRefBehavior(CallSite::get(Inst)) == 79 AliasAnalysis::DoesNotAccessMemory) 80 continue; 81 return DepResultTy(Inst, Normal); 82 } else { 83 // Non-memory instruction. 84 continue; 85 } 86 87 if (AA->getModRefInfo(C, Pointer, PointerSize) != AliasAnalysis::NoModRef) 88 return DepResultTy(Inst, Normal); 89 } 90 91 // No dependence found. 92 return DepResultTy(0, NonLocal); 93} 94 95/// getDependency - Return the instruction on which a memory operation 96/// depends. The local parameter indicates if the query should only 97/// evaluate dependencies within the same basic block. 98MemoryDependenceAnalysis::DepResultTy MemoryDependenceAnalysis:: 99getDependencyFromInternal(Instruction *QueryInst, BasicBlock::iterator ScanIt, 100 BasicBlock *BB) { 101 // Get the pointer value for which dependence will be determined 102 Value *MemPtr = 0; 103 uint64_t MemSize = 0; 104 bool MemVolatile = false; 105 106 if (StoreInst* S = dyn_cast<StoreInst>(QueryInst)) { 107 MemPtr = S->getPointerOperand(); 108 MemSize = TD->getTypeStoreSize(S->getOperand(0)->getType()); 109 MemVolatile = S->isVolatile(); 110 } else if (LoadInst* L = dyn_cast<LoadInst>(QueryInst)) { 111 MemPtr = L->getPointerOperand(); 112 MemSize = TD->getTypeStoreSize(L->getType()); 113 MemVolatile = L->isVolatile(); 114 } else if (VAArgInst* V = dyn_cast<VAArgInst>(QueryInst)) { 115 MemPtr = V->getOperand(0); 116 MemSize = TD->getTypeStoreSize(V->getType()); 117 } else if (FreeInst* F = dyn_cast<FreeInst>(QueryInst)) { 118 MemPtr = F->getPointerOperand(); 119 // FreeInsts erase the entire structure, not just a field. 120 MemSize = ~0UL; 121 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) 122 return getCallSiteDependency(CallSite::get(QueryInst), ScanIt, BB); 123 else // Non-memory instructions depend on nothing. 124 return DepResultTy(0, None); 125 126 // Walk backwards through the basic block, looking for dependencies 127 while (ScanIt != BB->begin()) { 128 Instruction *Inst = --ScanIt; 129 130 // If the access is volatile and this is a volatile load/store, return a 131 // dependence. 132 if (MemVolatile && 133 ((isa<LoadInst>(Inst) && cast<LoadInst>(Inst)->isVolatile()) || 134 (isa<StoreInst>(Inst) && cast<StoreInst>(Inst)->isVolatile()))) 135 return DepResultTy(Inst, Normal); 136 137 // Values depend on loads if the pointers are must aliased. This means that 138 // a load depends on another must aliased load from the same value. 139 if (LoadInst *L = dyn_cast<LoadInst>(Inst)) { 140 Value *Pointer = L->getPointerOperand(); 141 uint64_t PointerSize = TD->getTypeStoreSize(L->getType()); 142 143 // If we found a pointer, check if it could be the same as our pointer 144 AliasAnalysis::AliasResult R = 145 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 146 147 if (R == AliasAnalysis::NoAlias) 148 continue; 149 150 // May-alias loads don't depend on each other without a dependence. 151 if (isa<LoadInst>(QueryInst) && R == AliasAnalysis::MayAlias) 152 continue; 153 return DepResultTy(Inst, Normal); 154 } 155 156 // If this is an allocation, and if we know that the accessed pointer is to 157 // the allocation, return None. This means that there is no dependence and 158 // the access can be optimized based on that. For example, a load could 159 // turn into undef. 160 if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) { 161 Value *AccessPtr = MemPtr->getUnderlyingObject(); 162 163 if (AccessPtr == AI || 164 AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 165 return DepResultTy(0, None); 166 continue; 167 } 168 169 // See if this instruction mod/ref's the pointer. 170 AliasAnalysis::ModRefResult MRR = AA->getModRefInfo(Inst, MemPtr, MemSize); 171 172 if (MRR == AliasAnalysis::NoModRef) 173 continue; 174 175 // Loads don't depend on read-only instructions. 176 if (isa<LoadInst>(QueryInst) && MRR == AliasAnalysis::Ref) 177 continue; 178 179 // Otherwise, there is a dependence. 180 return DepResultTy(Inst, Normal); 181 } 182 183 // If we found nothing, return the non-local flag. 184 return DepResultTy(0, NonLocal); 185} 186 187/// getDependency - Return the instruction on which a memory operation 188/// depends. 189MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 190 Instruction *ScanPos = QueryInst; 191 192 // Check for a cached result 193 DepResultTy &LocalCache = LocalDeps[QueryInst]; 194 195 // If the cached entry is non-dirty, just return it. Note that this depends 196 // on DepResultTy's default constructing to 'dirty'. 197 if (LocalCache.getInt() != Dirty) 198 return ConvToResult(LocalCache); 199 200 // Otherwise, if we have a dirty entry, we know we can start the scan at that 201 // instruction, which may save us some work. 202 if (Instruction *Inst = LocalCache.getPointer()) { 203 ScanPos = Inst; 204 205 SmallPtrSet<Instruction*, 4> &InstMap = ReverseLocalDeps[Inst]; 206 InstMap.erase(QueryInst); 207 if (InstMap.empty()) 208 ReverseLocalDeps.erase(Inst); 209 } 210 211 // Do the scan. 212 LocalCache = getDependencyFromInternal(QueryInst, ScanPos, 213 QueryInst->getParent()); 214 215 // Remember the result! 216 if (Instruction *I = LocalCache.getPointer()) 217 ReverseLocalDeps[I].insert(QueryInst); 218 219 return ConvToResult(LocalCache); 220} 221 222/// getNonLocalDependency - Perform a full dependency query for the 223/// specified instruction, returning the set of blocks that the value is 224/// potentially live across. The returned set of results will include a 225/// "NonLocal" result for all blocks where the value is live across. 226/// 227/// This method assumes the instruction returns a "nonlocal" dependency 228/// within its own block. 229/// 230void MemoryDependenceAnalysis:: 231getNonLocalDependency(Instruction *QueryInst, 232 SmallVectorImpl<std::pair<BasicBlock*, 233 MemDepResult> > &Result) { 234 assert(getDependency(QueryInst).isNonLocal() && 235 "getNonLocalDependency should only be used on insts with non-local deps!"); 236 PerInstNLInfo &CacheP = NonLocalDeps[QueryInst]; 237 if (CacheP.getPointer() == 0) CacheP.setPointer(new NonLocalDepInfo()); 238 239 NonLocalDepInfo &Cache = *CacheP.getPointer(); 240 241 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 242 /// the cached case, this can happen due to instructions being deleted etc. In 243 /// the uncached case, this starts out as the set of predecessors we care 244 /// about. 245 SmallVector<BasicBlock*, 32> DirtyBlocks; 246 247 if (!Cache.empty()) { 248 // If we already have a partially computed set of results, scan them to 249 // determine what is dirty, seeding our initial DirtyBlocks worklist. The 250 // Int bit of CacheP tells us if we have anything dirty. 251 if (CacheP.getInt()) 252 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 253 I != E; ++I) 254 if (I->second.getInt() == Dirty) 255 DirtyBlocks.push_back(I->first); 256 257 NumCacheNonLocal++; 258 259 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 260 // << Cache.size() << " cached: " << *QueryInst; 261 } else { 262 // Seed DirtyBlocks with each of the preds of QueryInst's block. 263 BasicBlock *QueryBB = QueryInst->getParent(); 264 DirtyBlocks.append(pred_begin(QueryBB), pred_end(QueryBB)); 265 NumUncacheNonLocal++; 266 } 267 268 // Iterate while we still have blocks to update. 269 while (!DirtyBlocks.empty()) { 270 BasicBlock *DirtyBB = DirtyBlocks.back(); 271 DirtyBlocks.pop_back(); 272 273 // Get the entry for this block. Note that this relies on DepResultTy 274 // default initializing to Dirty. 275 DepResultTy &DirtyBBEntry = Cache[DirtyBB]; 276 277 // If DirtyBBEntry isn't dirty, it ended up on the worklist multiple times. 278 if (DirtyBBEntry.getInt() != Dirty) continue; 279 280 // If the dirty entry has a pointer, start scanning from it so we don't have 281 // to rescan the entire block. 282 BasicBlock::iterator ScanPos = DirtyBB->end(); 283 if (Instruction *Inst = DirtyBBEntry.getPointer()) { 284 ScanPos = Inst; 285 286 // We're removing QueryInst's dependence on Inst. 287 SmallPtrSet<Instruction*, 4> &InstMap = ReverseNonLocalDeps[Inst]; 288 InstMap.erase(QueryInst); 289 if (InstMap.empty()) ReverseNonLocalDeps.erase(Inst); 290 } 291 292 // Find out if this block has a local dependency for QueryInst. 293 DirtyBBEntry = getDependencyFromInternal(QueryInst, ScanPos, DirtyBB); 294 295 // If the block has a dependency (i.e. it isn't completely transparent to 296 // the value), remember it! 297 if (DirtyBBEntry.getInt() != NonLocal) { 298 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 299 // update this when we remove instructions. 300 if (Instruction *Inst = DirtyBBEntry.getPointer()) 301 ReverseNonLocalDeps[Inst].insert(QueryInst); 302 continue; 303 } 304 305 // If the block *is* completely transparent to the load, we need to check 306 // the predecessors of this block. Add them to our worklist. 307 DirtyBlocks.append(pred_begin(DirtyBB), pred_end(DirtyBB)); 308 } 309 310 311 // Copy the result into the output set. 312 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); I != E;++I) 313 Result.push_back(std::make_pair(I->first, ConvToResult(I->second))); 314} 315 316/// removeInstruction - Remove an instruction from the dependence analysis, 317/// updating the dependence of instructions that previously depended on it. 318/// This method attempts to keep the cache coherent using the reverse map. 319void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 320 // Walk through the Non-local dependencies, removing this one as the value 321 // for any cached queries. 322 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 323 if (NLDI != NonLocalDeps.end()) { 324 NonLocalDepInfo &BlockMap = *NLDI->second.getPointer(); 325 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 326 DI != DE; ++DI) 327 if (Instruction *Inst = DI->second.getPointer()) 328 ReverseNonLocalDeps[Inst].erase(RemInst); 329 delete &BlockMap; 330 NonLocalDeps.erase(NLDI); 331 } 332 333 // If we have a cached local dependence query for this instruction, remove it. 334 // 335 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 336 if (LocalDepEntry != LocalDeps.end()) { 337 // Remove us from DepInst's reverse set now that the local dep info is gone. 338 if (Instruction *Inst = LocalDepEntry->second.getPointer()) { 339 SmallPtrSet<Instruction*, 4> &RLD = ReverseLocalDeps[Inst]; 340 RLD.erase(RemInst); 341 if (RLD.empty()) 342 ReverseLocalDeps.erase(Inst); 343 } 344 345 // Remove this local dependency info. 346 LocalDeps.erase(LocalDepEntry); 347 } 348 349 // Loop over all of the things that depend on the instruction we're removing. 350 // 351 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 352 353 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 354 if (ReverseDepIt != ReverseLocalDeps.end()) { 355 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 356 // RemInst can't be the terminator if it has stuff depending on it. 357 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 358 "Nothing can locally depend on a terminator"); 359 360 // Anything that was locally dependent on RemInst is now going to be 361 // dependent on the instruction after RemInst. It will have the dirty flag 362 // set so it will rescan. This saves having to scan the entire block to get 363 // to this point. 364 Instruction *NewDepInst = next(BasicBlock::iterator(RemInst)); 365 366 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 367 E = ReverseDeps.end(); I != E; ++I) { 368 Instruction *InstDependingOnRemInst = *I; 369 assert(InstDependingOnRemInst != RemInst && 370 "Already removed our local dep info"); 371 372 LocalDeps[InstDependingOnRemInst] = DepResultTy(NewDepInst, Dirty); 373 374 // Make sure to remember that new things depend on NewDepInst. 375 ReverseDepsToAdd.push_back(std::make_pair(NewDepInst, 376 InstDependingOnRemInst)); 377 } 378 379 ReverseLocalDeps.erase(ReverseDepIt); 380 381 // Add new reverse deps after scanning the set, to avoid invalidating the 382 // 'ReverseDeps' reference. 383 while (!ReverseDepsToAdd.empty()) { 384 ReverseLocalDeps[ReverseDepsToAdd.back().first] 385 .insert(ReverseDepsToAdd.back().second); 386 ReverseDepsToAdd.pop_back(); 387 } 388 } 389 390 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 391 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 392 SmallPtrSet<Instruction*, 4>& set = ReverseDepIt->second; 393 for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end(); 394 I != E; ++I) { 395 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 396 397 PerInstNLInfo &INLD = NonLocalDeps[*I]; 398 assert(INLD.getPointer() != 0 && "Reverse mapping out of date?"); 399 // The information is now dirty! 400 INLD.setInt(true); 401 402 for (NonLocalDepInfo::iterator DI = INLD.getPointer()->begin(), 403 DE = INLD.getPointer()->end(); DI != DE; ++DI) { 404 if (DI->second.getPointer() != RemInst) continue; 405 406 // Convert to a dirty entry for the subsequent instruction. 407 DI->second.setInt(Dirty); 408 if (RemInst->isTerminator()) 409 DI->second.setPointer(0); 410 else { 411 Instruction *NextI = next(BasicBlock::iterator(RemInst)); 412 DI->second.setPointer(NextI); 413 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 414 } 415 } 416 } 417 418 ReverseNonLocalDeps.erase(ReverseDepIt); 419 420 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 421 while (!ReverseDepsToAdd.empty()) { 422 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 423 .insert(ReverseDepsToAdd.back().second); 424 ReverseDepsToAdd.pop_back(); 425 } 426 } 427 428 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 429 AA->deleteValue(RemInst); 430 DEBUG(verifyRemoved(RemInst)); 431} 432 433/// verifyRemoved - Verify that the specified instruction does not occur 434/// in our internal data structures. 435void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 436 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 437 E = LocalDeps.end(); I != E; ++I) { 438 assert(I->first != D && "Inst occurs in data structures"); 439 assert(I->second.getPointer() != D && 440 "Inst occurs in data structures"); 441 } 442 443 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 444 E = NonLocalDeps.end(); I != E; ++I) { 445 assert(I->first != D && "Inst occurs in data structures"); 446 const PerInstNLInfo &INLD = I->second; 447 for (NonLocalDepInfo::iterator II = INLD.getPointer()->begin(), 448 EE = INLD.getPointer()->end(); II != EE; ++II) 449 assert(II->second.getPointer() != D && "Inst occurs in data structures"); 450 } 451 452 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 453 E = ReverseLocalDeps.end(); I != E; ++I) { 454 assert(I->first != D && "Inst occurs in data structures"); 455 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 456 EE = I->second.end(); II != EE; ++II) 457 assert(*II != D && "Inst occurs in data structures"); 458 } 459 460 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 461 E = ReverseNonLocalDeps.end(); 462 I != E; ++I) { 463 assert(I->first != D && "Inst occurs in data structures"); 464 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 465 EE = I->second.end(); II != EE; ++II) 466 assert(*II != D && "Inst occurs in data structures"); 467 } 468} 469