MemoryDependenceAnalysis.cpp revision b51deb929ca95ce62e622b0475a05d83f26ab04d
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 fully cached non-local responses"); 32STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 33STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 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 55/// getCallSiteDependency - Private helper for finding the local dependencies 56/// of a call site. 57MemDepResult MemoryDependenceAnalysis:: 58getCallSiteDependency(CallSite CS, BasicBlock::iterator ScanIt, 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 CallSite InstCS = CallSite::get(Inst); 79 // If these two calls do not interfere, look past it. 80 if (AA->getModRefInfo(CS, InstCS) == AliasAnalysis::NoModRef) 81 continue; 82 83 // FIXME: If this is a ref/ref result, we should ignore it! 84 // X = strlen(P); 85 // Y = strlen(Q); 86 // Z = strlen(P); // Z = X 87 88 // If they interfere, we generally return clobber. However, if they are 89 // calls to the same read-only functions we return Def. 90 if (!AA->onlyReadsMemory(CS) || CS.getCalledFunction() == 0 || 91 CS.getCalledFunction() != InstCS.getCalledFunction()) 92 return MemDepResult::getClobber(Inst); 93 return MemDepResult::getDef(Inst); 94 } else { 95 // Non-memory instruction. 96 continue; 97 } 98 99 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef) 100 return MemDepResult::getClobber(Inst); 101 } 102 103 // No dependence found. 104 return MemDepResult::getNonLocal(); 105} 106 107/// getDependencyFrom - Return the instruction on which a memory operation 108/// depends. 109MemDepResult MemoryDependenceAnalysis:: 110getDependencyFrom(Instruction *QueryInst, BasicBlock::iterator ScanIt, 111 BasicBlock *BB) { 112 // Get the pointer value for which dependence will be determined 113 Value *MemPtr = 0; 114 uint64_t MemSize = 0; 115 bool MemVolatile = false; 116 117 if (StoreInst* S = dyn_cast<StoreInst>(QueryInst)) { 118 MemPtr = S->getPointerOperand(); 119 MemSize = TD->getTypeStoreSize(S->getOperand(0)->getType()); 120 MemVolatile = S->isVolatile(); 121 } else if (LoadInst* LI = dyn_cast<LoadInst>(QueryInst)) { 122 MemPtr = LI->getPointerOperand(); 123 MemSize = TD->getTypeStoreSize(LI->getType()); 124 MemVolatile = LI->isVolatile(); 125 } else if (VAArgInst* V = dyn_cast<VAArgInst>(QueryInst)) { 126 MemPtr = V->getOperand(0); 127 MemSize = TD->getTypeStoreSize(V->getType()); 128 } else if (FreeInst* F = dyn_cast<FreeInst>(QueryInst)) { 129 MemPtr = F->getPointerOperand(); 130 // FreeInsts erase the entire structure, not just a field. 131 MemSize = ~0UL; 132 } else { 133 assert((isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) && 134 "Can only get dependency info for memory instructions!"); 135 return getCallSiteDependency(CallSite::get(QueryInst), ScanIt, BB); 136 } 137 138 // Walk backwards through the basic block, looking for dependencies 139 while (ScanIt != BB->begin()) { 140 Instruction *Inst = --ScanIt; 141 142 // Values depend on loads if the pointers are must aliased. This means that 143 // a load depends on another must aliased load from the same value. 144 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 145 // If the access is volatile and this is volatile, return a dependence. 146 if (MemVolatile && LI->isVolatile()) 147 return MemDepResult::getClobber(LI); 148 149 Value *Pointer = LI->getPointerOperand(); 150 uint64_t PointerSize = TD->getTypeStoreSize(LI->getType()); 151 152 // If we found a pointer, check if it could be the same as our pointer. 153 AliasAnalysis::AliasResult R = 154 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 155 if (R == AliasAnalysis::NoAlias) 156 continue; 157 158 // May-alias loads don't depend on each other without a dependence. 159 if (isa<LoadInst>(QueryInst) && R == AliasAnalysis::MayAlias) 160 continue; 161 return MemDepResult::getDef(Inst); 162 } 163 164 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 165 // If the access is volatile and this is volatile, return a dependence. 166 if (MemVolatile && SI->isVolatile()) 167 return MemDepResult::getClobber(SI); 168 169 Value *Pointer = SI->getPointerOperand(); 170 uint64_t PointerSize = TD->getTypeStoreSize(SI->getOperand(0)->getType()); 171 172 // If we found a pointer, check if it could be the same as our pointer. 173 AliasAnalysis::AliasResult R = 174 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 175 176 if (R == AliasAnalysis::NoAlias) 177 continue; 178 if (R == AliasAnalysis::MayAlias) 179 return MemDepResult::getClobber(Inst); 180 return MemDepResult::getDef(Inst); 181 } 182 183 // If this is an allocation, and if we know that the accessed pointer is to 184 // the allocation, return Def. This means that there is no dependence and 185 // the access can be optimized based on that. For example, a load could 186 // turn into undef. 187 if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) { 188 Value *AccessPtr = MemPtr->getUnderlyingObject(); 189 190 if (AccessPtr == AI || 191 AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 192 return MemDepResult::getDef(AI); 193 continue; 194 } 195 196 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 197 if (AA->getModRefInfo(Inst, MemPtr, MemSize) == AliasAnalysis::NoModRef) 198 continue; 199 200 // Otherwise, there is a dependence. 201 return MemDepResult::getClobber(Inst); 202 } 203 204 // If we found nothing, return the non-local flag. 205 return MemDepResult::getNonLocal(); 206} 207 208/// getDependency - Return the instruction on which a memory operation 209/// depends. 210MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 211 Instruction *ScanPos = QueryInst; 212 213 // Check for a cached result 214 MemDepResult &LocalCache = LocalDeps[QueryInst]; 215 216 // If the cached entry is non-dirty, just return it. Note that this depends 217 // on MemDepResult's default constructing to 'dirty'. 218 if (!LocalCache.isDirty()) 219 return LocalCache; 220 221 // Otherwise, if we have a dirty entry, we know we can start the scan at that 222 // instruction, which may save us some work. 223 if (Instruction *Inst = LocalCache.getInst()) { 224 ScanPos = Inst; 225 226 SmallPtrSet<Instruction*, 4> &InstMap = ReverseLocalDeps[Inst]; 227 InstMap.erase(QueryInst); 228 if (InstMap.empty()) 229 ReverseLocalDeps.erase(Inst); 230 } 231 232 // Do the scan. 233 LocalCache = getDependencyFrom(QueryInst, ScanPos, QueryInst->getParent()); 234 235 // Remember the result! 236 if (Instruction *I = LocalCache.getInst()) 237 ReverseLocalDeps[I].insert(QueryInst); 238 239 return LocalCache; 240} 241 242/// getNonLocalDependency - Perform a full dependency query for the 243/// specified instruction, returning the set of blocks that the value is 244/// potentially live across. The returned set of results will include a 245/// "NonLocal" result for all blocks where the value is live across. 246/// 247/// This method assumes the instruction returns a "nonlocal" dependency 248/// within its own block. 249/// 250const MemoryDependenceAnalysis::NonLocalDepInfo & 251MemoryDependenceAnalysis::getNonLocalDependency(Instruction *QueryInst) { 252 assert(getDependency(QueryInst).isNonLocal() && 253 "getNonLocalDependency should only be used on insts with non-local deps!"); 254 PerInstNLInfo &CacheP = NonLocalDeps[QueryInst]; 255 256 NonLocalDepInfo &Cache = CacheP.first; 257 258 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 259 /// the cached case, this can happen due to instructions being deleted etc. In 260 /// the uncached case, this starts out as the set of predecessors we care 261 /// about. 262 SmallVector<BasicBlock*, 32> DirtyBlocks; 263 264 if (!Cache.empty()) { 265 // Okay, we have a cache entry. If we know it is not dirty, just return it 266 // with no computation. 267 if (!CacheP.second) { 268 NumCacheNonLocal++; 269 return Cache; 270 } 271 272 // If we already have a partially computed set of results, scan them to 273 // determine what is dirty, seeding our initial DirtyBlocks worklist. 274 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 275 I != E; ++I) 276 if (I->second.isDirty()) 277 DirtyBlocks.push_back(I->first); 278 279 // Sort the cache so that we can do fast binary search lookups below. 280 std::sort(Cache.begin(), Cache.end()); 281 282 ++NumCacheDirtyNonLocal; 283 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 284 // << Cache.size() << " cached: " << *QueryInst; 285 } else { 286 // Seed DirtyBlocks with each of the preds of QueryInst's block. 287 BasicBlock *QueryBB = QueryInst->getParent(); 288 DirtyBlocks.append(pred_begin(QueryBB), pred_end(QueryBB)); 289 NumUncacheNonLocal++; 290 } 291 292 // Visited checked first, vector in sorted order. 293 SmallPtrSet<BasicBlock*, 64> Visited; 294 295 unsigned NumSortedEntries = Cache.size(); 296 297 // Iterate while we still have blocks to update. 298 while (!DirtyBlocks.empty()) { 299 BasicBlock *DirtyBB = DirtyBlocks.back(); 300 DirtyBlocks.pop_back(); 301 302 // Already processed this block? 303 if (!Visited.insert(DirtyBB)) 304 continue; 305 306 // Do a binary search to see if we already have an entry for this block in 307 // the cache set. If so, find it. 308 NonLocalDepInfo::iterator Entry = 309 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 310 std::make_pair(DirtyBB, MemDepResult())); 311 if (Entry != Cache.begin() && (&*Entry)[-1].first == DirtyBB) 312 --Entry; 313 314 MemDepResult *ExistingResult = 0; 315 if (Entry != Cache.begin()+NumSortedEntries && 316 Entry->first == DirtyBB) { 317 // If we already have an entry, and if it isn't already dirty, the block 318 // is done. 319 if (!Entry->second.isDirty()) 320 continue; 321 322 // Otherwise, remember this slot so we can update the value. 323 ExistingResult = &Entry->second; 324 } 325 326 // If the dirty entry has a pointer, start scanning from it so we don't have 327 // to rescan the entire block. 328 BasicBlock::iterator ScanPos = DirtyBB->end(); 329 if (ExistingResult) { 330 if (Instruction *Inst = ExistingResult->getInst()) { 331 ScanPos = Inst; 332 333 // We're removing QueryInst's use of Inst. 334 SmallPtrSet<Instruction*, 4> &InstMap = ReverseNonLocalDeps[Inst]; 335 InstMap.erase(QueryInst); 336 if (InstMap.empty()) ReverseNonLocalDeps.erase(Inst); 337 } 338 } 339 340 // Find out if this block has a local dependency for QueryInst. 341 MemDepResult Dep = getDependencyFrom(QueryInst, ScanPos, DirtyBB); 342 343 // If we had a dirty entry for the block, update it. Otherwise, just add 344 // a new entry. 345 if (ExistingResult) 346 *ExistingResult = Dep; 347 else 348 Cache.push_back(std::make_pair(DirtyBB, Dep)); 349 350 // If the block has a dependency (i.e. it isn't completely transparent to 351 // the value), remember the association! 352 if (!Dep.isNonLocal()) { 353 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 354 // update this when we remove instructions. 355 if (Instruction *Inst = Dep.getInst()) 356 ReverseNonLocalDeps[Inst].insert(QueryInst); 357 } else { 358 359 // If the block *is* completely transparent to the load, we need to check 360 // the predecessors of this block. Add them to our worklist. 361 DirtyBlocks.append(pred_begin(DirtyBB), pred_end(DirtyBB)); 362 } 363 } 364 365 return Cache; 366} 367 368/// removeInstruction - Remove an instruction from the dependence analysis, 369/// updating the dependence of instructions that previously depended on it. 370/// This method attempts to keep the cache coherent using the reverse map. 371void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 372 // Walk through the Non-local dependencies, removing this one as the value 373 // for any cached queries. 374 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 375 if (NLDI != NonLocalDeps.end()) { 376 NonLocalDepInfo &BlockMap = NLDI->second.first; 377 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 378 DI != DE; ++DI) 379 if (Instruction *Inst = DI->second.getInst()) 380 ReverseNonLocalDeps[Inst].erase(RemInst); 381 NonLocalDeps.erase(NLDI); 382 } 383 384 // If we have a cached local dependence query for this instruction, remove it. 385 // 386 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 387 if (LocalDepEntry != LocalDeps.end()) { 388 // Remove us from DepInst's reverse set now that the local dep info is gone. 389 if (Instruction *Inst = LocalDepEntry->second.getInst()) { 390 SmallPtrSet<Instruction*, 4> &RLD = ReverseLocalDeps[Inst]; 391 RLD.erase(RemInst); 392 if (RLD.empty()) 393 ReverseLocalDeps.erase(Inst); 394 } 395 396 // Remove this local dependency info. 397 LocalDeps.erase(LocalDepEntry); 398 } 399 400 // Loop over all of the things that depend on the instruction we're removing. 401 // 402 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 403 404 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 405 if (ReverseDepIt != ReverseLocalDeps.end()) { 406 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 407 // RemInst can't be the terminator if it has stuff depending on it. 408 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 409 "Nothing can locally depend on a terminator"); 410 411 // Anything that was locally dependent on RemInst is now going to be 412 // dependent on the instruction after RemInst. It will have the dirty flag 413 // set so it will rescan. This saves having to scan the entire block to get 414 // to this point. 415 Instruction *NewDepInst = next(BasicBlock::iterator(RemInst)); 416 417 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 418 E = ReverseDeps.end(); I != E; ++I) { 419 Instruction *InstDependingOnRemInst = *I; 420 assert(InstDependingOnRemInst != RemInst && 421 "Already removed our local dep info"); 422 423 LocalDeps[InstDependingOnRemInst] = MemDepResult::getDirty(NewDepInst); 424 425 // Make sure to remember that new things depend on NewDepInst. 426 ReverseDepsToAdd.push_back(std::make_pair(NewDepInst, 427 InstDependingOnRemInst)); 428 } 429 430 ReverseLocalDeps.erase(ReverseDepIt); 431 432 // Add new reverse deps after scanning the set, to avoid invalidating the 433 // 'ReverseDeps' reference. 434 while (!ReverseDepsToAdd.empty()) { 435 ReverseLocalDeps[ReverseDepsToAdd.back().first] 436 .insert(ReverseDepsToAdd.back().second); 437 ReverseDepsToAdd.pop_back(); 438 } 439 } 440 441 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 442 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 443 SmallPtrSet<Instruction*, 4>& set = ReverseDepIt->second; 444 for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end(); 445 I != E; ++I) { 446 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 447 448 PerInstNLInfo &INLD = NonLocalDeps[*I]; 449 // The information is now dirty! 450 INLD.second = true; 451 452 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 453 DE = INLD.first.end(); DI != DE; ++DI) { 454 if (DI->second.getInst() != RemInst) continue; 455 456 // Convert to a dirty entry for the subsequent instruction. 457 Instruction *NextI = 0; 458 if (!RemInst->isTerminator()) { 459 NextI = next(BasicBlock::iterator(RemInst)); 460 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 461 } 462 DI->second = MemDepResult::getDirty(NextI); 463 } 464 } 465 466 ReverseNonLocalDeps.erase(ReverseDepIt); 467 468 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 469 while (!ReverseDepsToAdd.empty()) { 470 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 471 .insert(ReverseDepsToAdd.back().second); 472 ReverseDepsToAdd.pop_back(); 473 } 474 } 475 476 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 477 AA->deleteValue(RemInst); 478 DEBUG(verifyRemoved(RemInst)); 479} 480 481/// verifyRemoved - Verify that the specified instruction does not occur 482/// in our internal data structures. 483void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 484 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 485 E = LocalDeps.end(); I != E; ++I) { 486 assert(I->first != D && "Inst occurs in data structures"); 487 assert(I->second.getInst() != D && 488 "Inst occurs in data structures"); 489 } 490 491 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 492 E = NonLocalDeps.end(); I != E; ++I) { 493 assert(I->first != D && "Inst occurs in data structures"); 494 const PerInstNLInfo &INLD = I->second; 495 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 496 EE = INLD.first.end(); II != EE; ++II) 497 assert(II->second.getInst() != D && "Inst occurs in data structures"); 498 } 499 500 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 501 E = ReverseLocalDeps.end(); I != E; ++I) { 502 assert(I->first != D && "Inst occurs in data structures"); 503 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 504 EE = I->second.end(); II != EE; ++II) 505 assert(*II != D && "Inst occurs in data structures"); 506 } 507 508 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 509 E = ReverseNonLocalDeps.end(); 510 I != E; ++I) { 511 assert(I->first != D && "Inst occurs in data structures"); 512 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 513 EE = I->second.end(); II != EE; ++II) 514 assert(*II != D && "Inst occurs in data structures"); 515 } 516} 517