MemoryDependenceAnalysis.cpp revision 616613d7a4ddc7cefce53b2bfe3fdcdec6b032c2
1//===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation --*- C++ -*-===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements an analysis that determines, for a given memory 11// operation, what preceding memory operations it depends on. It builds on 12// alias analysis information, and tries to provide a lazy, caching interface to 13// a common kind of alias information query. 14// 15//===----------------------------------------------------------------------===// 16 17#define DEBUG_TYPE "memdep" 18#include "llvm/Analysis/MemoryDependenceAnalysis.h" 19#include "llvm/Instructions.h" 20#include "llvm/IntrinsicInst.h" 21#include "llvm/Function.h" 22#include "llvm/Analysis/AliasAnalysis.h" 23#include "llvm/Analysis/InstructionSimplify.h" 24#include "llvm/Analysis/MemoryBuiltins.h" 25#include "llvm/ADT/Statistic.h" 26#include "llvm/ADT/STLExtras.h" 27#include "llvm/Support/PredIteratorCache.h" 28#include "llvm/Support/Debug.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"); 34 35STATISTIC(NumCacheNonLocalPtr, 36 "Number of fully cached non-local ptr responses"); 37STATISTIC(NumCacheDirtyNonLocalPtr, 38 "Number of cached, but dirty, non-local ptr responses"); 39STATISTIC(NumUncacheNonLocalPtr, 40 "Number of uncached non-local ptr responses"); 41STATISTIC(NumCacheCompleteNonLocalPtr, 42 "Number of block queries that were completely cached"); 43 44char MemoryDependenceAnalysis::ID = 0; 45 46// Register this pass... 47static RegisterPass<MemoryDependenceAnalysis> X("memdep", 48 "Memory Dependence Analysis", false, true); 49 50MemoryDependenceAnalysis::MemoryDependenceAnalysis() 51: FunctionPass(&ID), PredCache(0) { 52} 53MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 54} 55 56/// Clean up memory in between runs 57void MemoryDependenceAnalysis::releaseMemory() { 58 LocalDeps.clear(); 59 NonLocalDeps.clear(); 60 NonLocalPointerDeps.clear(); 61 ReverseLocalDeps.clear(); 62 ReverseNonLocalDeps.clear(); 63 ReverseNonLocalPtrDeps.clear(); 64 PredCache->clear(); 65} 66 67 68 69/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 70/// 71void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 72 AU.setPreservesAll(); 73 AU.addRequiredTransitive<AliasAnalysis>(); 74} 75 76bool MemoryDependenceAnalysis::runOnFunction(Function &) { 77 AA = &getAnalysis<AliasAnalysis>(); 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 = AA->getTypeStoreSize(S->getOperand(0)->getType()); 114 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 115 Pointer = V->getOperand(0); 116 PointerSize = AA->getTypeStoreSize(V->getType()); 117 } else if (isFreeCall(Inst)) { 118 Pointer = Inst->getOperand(1); 119 // calls to free() erase the entire structure 120 PointerSize = ~0ULL; 121 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) { 122 // Debug intrinsics don't cause dependences. 123 if (isa<DbgInfoIntrinsic>(Inst)) continue; 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 Value *invariantTag = 0; 175 176 // Walk backwards through the basic block, looking for dependencies. 177 while (ScanIt != BB->begin()) { 178 Instruction *Inst = --ScanIt; 179 180 // If we're in an invariant region, no dependencies can be found before 181 // we pass an invariant-begin marker. 182 if (invariantTag == Inst) { 183 invariantTag = 0; 184 continue; 185 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 186 // If we pass an invariant-end marker, then we've just entered an 187 // invariant region and can start ignoring dependencies. 188 if (II->getIntrinsicID() == Intrinsic::invariant_end) { 189 uint64_t invariantSize = ~0ULL; 190 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getOperand(2))) 191 invariantSize = CI->getZExtValue(); 192 193 AliasAnalysis::AliasResult R = 194 AA->alias(II->getOperand(3), invariantSize, MemPtr, MemSize); 195 if (R == AliasAnalysis::MustAlias) { 196 invariantTag = II->getOperand(1); 197 continue; 198 } 199 200 // If we reach a lifetime begin or end marker, then the query ends here 201 // because the value is undefined. 202 } else if (II->getIntrinsicID() == Intrinsic::lifetime_start || 203 II->getIntrinsicID() == Intrinsic::lifetime_end) { 204 uint64_t invariantSize = ~0ULL; 205 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getOperand(1))) 206 invariantSize = CI->getZExtValue(); 207 208 AliasAnalysis::AliasResult R = 209 AA->alias(II->getOperand(2), invariantSize, MemPtr, MemSize); 210 if (R == AliasAnalysis::MustAlias) 211 return MemDepResult::getDef(II); 212 } 213 } 214 215 // If we're querying on a load and we're in an invariant region, we're done 216 // at this point. Nothing a load depends on can live in an invariant region. 217 if (isLoad && invariantTag) continue; 218 219 // Debug intrinsics don't cause dependences. 220 if (isa<DbgInfoIntrinsic>(Inst)) continue; 221 222 // Values depend on loads if the pointers are must aliased. This means that 223 // a load depends on another must aliased load from the same value. 224 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 225 Value *Pointer = LI->getPointerOperand(); 226 uint64_t PointerSize = AA->getTypeStoreSize(LI->getType()); 227 228 // If we found a pointer, check if it could be the same as our pointer. 229 AliasAnalysis::AliasResult R = 230 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 231 if (R == AliasAnalysis::NoAlias) 232 continue; 233 234 // May-alias loads don't depend on each other without a dependence. 235 if (isLoad && R == AliasAnalysis::MayAlias) 236 continue; 237 // Stores depend on may and must aliased loads, loads depend on must-alias 238 // loads. 239 return MemDepResult::getDef(Inst); 240 } 241 242 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 243 // There can't be stores to the value we care about inside an 244 // invariant region. 245 if (invariantTag) continue; 246 247 // If alias analysis can tell that this store is guaranteed to not modify 248 // the query pointer, ignore it. Use getModRefInfo to handle cases where 249 // the query pointer points to constant memory etc. 250 if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef) 251 continue; 252 253 // Ok, this store might clobber the query pointer. Check to see if it is 254 // a must alias: in this case, we want to return this as a def. 255 Value *Pointer = SI->getPointerOperand(); 256 uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); 257 258 // If we found a pointer, check if it could be the same as our pointer. 259 AliasAnalysis::AliasResult R = 260 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 261 262 if (R == AliasAnalysis::NoAlias) 263 continue; 264 if (R == AliasAnalysis::MayAlias) 265 return MemDepResult::getClobber(Inst); 266 return MemDepResult::getDef(Inst); 267 } 268 269 // If this is an allocation, and if we know that the accessed pointer is to 270 // the allocation, return Def. This means that there is no dependence and 271 // the access can be optimized based on that. For example, a load could 272 // turn into undef. 273 // Note: Only determine this to be a malloc if Inst is the malloc call, not 274 // a subsequent bitcast of the malloc call result. There can be stores to 275 // the malloced memory between the malloc call and its bitcast uses, and we 276 // need to continue scanning until the malloc call. 277 if (isa<AllocaInst>(Inst) || extractMallocCall(Inst)) { 278 Value *AccessPtr = MemPtr->getUnderlyingObject(); 279 280 if (AccessPtr == Inst || 281 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 282 return MemDepResult::getDef(Inst); 283 continue; 284 } 285 286 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 287 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) { 288 case AliasAnalysis::NoModRef: 289 // If the call has no effect on the queried pointer, just ignore it. 290 continue; 291 case AliasAnalysis::Mod: 292 // If we're in an invariant region, we can ignore calls that ONLY 293 // modify the pointer. 294 if (invariantTag) continue; 295 return MemDepResult::getClobber(Inst); 296 case AliasAnalysis::Ref: 297 // If the call is known to never store to the pointer, and if this is a 298 // load query, we can safely ignore it (scan past it). 299 if (isLoad) 300 continue; 301 default: 302 // Otherwise, there is a potential dependence. Return a clobber. 303 return MemDepResult::getClobber(Inst); 304 } 305 } 306 307 // No dependence found. If this is the entry block of the function, it is a 308 // clobber, otherwise it is non-local. 309 if (BB != &BB->getParent()->getEntryBlock()) 310 return MemDepResult::getNonLocal(); 311 return MemDepResult::getClobber(ScanIt); 312} 313 314/// getDependency - Return the instruction on which a memory operation 315/// depends. 316MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 317 Instruction *ScanPos = QueryInst; 318 319 // Check for a cached result 320 MemDepResult &LocalCache = LocalDeps[QueryInst]; 321 322 // If the cached entry is non-dirty, just return it. Note that this depends 323 // on MemDepResult's default constructing to 'dirty'. 324 if (!LocalCache.isDirty()) 325 return LocalCache; 326 327 // Otherwise, if we have a dirty entry, we know we can start the scan at that 328 // instruction, which may save us some work. 329 if (Instruction *Inst = LocalCache.getInst()) { 330 ScanPos = Inst; 331 332 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 333 } 334 335 BasicBlock *QueryParent = QueryInst->getParent(); 336 337 Value *MemPtr = 0; 338 uint64_t MemSize = 0; 339 340 // Do the scan. 341 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 342 // No dependence found. If this is the entry block of the function, it is a 343 // clobber, otherwise it is non-local. 344 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 345 LocalCache = MemDepResult::getNonLocal(); 346 else 347 LocalCache = MemDepResult::getClobber(QueryInst); 348 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) { 349 // If this is a volatile store, don't mess around with it. Just return the 350 // previous instruction as a clobber. 351 if (SI->isVolatile()) 352 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 353 else { 354 MemPtr = SI->getPointerOperand(); 355 MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); 356 } 357 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) { 358 // If this is a volatile load, don't mess around with it. Just return the 359 // previous instruction as a clobber. 360 if (LI->isVolatile()) 361 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 362 else { 363 MemPtr = LI->getPointerOperand(); 364 MemSize = AA->getTypeStoreSize(LI->getType()); 365 } 366 } else if (isFreeCall(QueryInst)) { 367 MemPtr = QueryInst->getOperand(1); 368 // calls to free() erase the entire structure, not just a field. 369 MemSize = ~0UL; 370 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 371 CallSite QueryCS = CallSite::get(QueryInst); 372 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 373 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 374 QueryParent); 375 } else { 376 // Non-memory instruction. 377 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 378 } 379 380 // If we need to do a pointer scan, make it happen. 381 if (MemPtr) 382 LocalCache = getPointerDependencyFrom(MemPtr, MemSize, 383 isa<LoadInst>(QueryInst), 384 ScanPos, QueryParent); 385 386 // Remember the result! 387 if (Instruction *I = LocalCache.getInst()) 388 ReverseLocalDeps[I].insert(QueryInst); 389 390 return LocalCache; 391} 392 393#ifndef NDEBUG 394/// AssertSorted - This method is used when -debug is specified to verify that 395/// cache arrays are properly kept sorted. 396static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 397 int Count = -1) { 398 if (Count == -1) Count = Cache.size(); 399 if (Count == 0) return; 400 401 for (unsigned i = 1; i != unsigned(Count); ++i) 402 assert(Cache[i-1] <= Cache[i] && "Cache isn't sorted!"); 403} 404#endif 405 406/// getNonLocalCallDependency - Perform a full dependency query for the 407/// specified call, returning the set of blocks that the value is 408/// potentially live across. The returned set of results will include a 409/// "NonLocal" result for all blocks where the value is live across. 410/// 411/// This method assumes the instruction returns a "NonLocal" dependency 412/// within its own block. 413/// 414/// This returns a reference to an internal data structure that may be 415/// invalidated on the next non-local query or when an instruction is 416/// removed. Clients must copy this data if they want it around longer than 417/// that. 418const MemoryDependenceAnalysis::NonLocalDepInfo & 419MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 420 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 421 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 422 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 423 NonLocalDepInfo &Cache = CacheP.first; 424 425 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 426 /// the cached case, this can happen due to instructions being deleted etc. In 427 /// the uncached case, this starts out as the set of predecessors we care 428 /// about. 429 SmallVector<BasicBlock*, 32> DirtyBlocks; 430 431 if (!Cache.empty()) { 432 // Okay, we have a cache entry. If we know it is not dirty, just return it 433 // with no computation. 434 if (!CacheP.second) { 435 NumCacheNonLocal++; 436 return Cache; 437 } 438 439 // If we already have a partially computed set of results, scan them to 440 // determine what is dirty, seeding our initial DirtyBlocks worklist. 441 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 442 I != E; ++I) 443 if (I->second.isDirty()) 444 DirtyBlocks.push_back(I->first); 445 446 // Sort the cache so that we can do fast binary search lookups below. 447 std::sort(Cache.begin(), Cache.end()); 448 449 ++NumCacheDirtyNonLocal; 450 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 451 // << Cache.size() << " cached: " << *QueryInst; 452 } else { 453 // Seed DirtyBlocks with each of the preds of QueryInst's block. 454 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 455 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 456 DirtyBlocks.push_back(*PI); 457 NumUncacheNonLocal++; 458 } 459 460 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 461 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 462 463 SmallPtrSet<BasicBlock*, 64> Visited; 464 465 unsigned NumSortedEntries = Cache.size(); 466 DEBUG(AssertSorted(Cache)); 467 468 // Iterate while we still have blocks to update. 469 while (!DirtyBlocks.empty()) { 470 BasicBlock *DirtyBB = DirtyBlocks.back(); 471 DirtyBlocks.pop_back(); 472 473 // Already processed this block? 474 if (!Visited.insert(DirtyBB)) 475 continue; 476 477 // Do a binary search to see if we already have an entry for this block in 478 // the cache set. If so, find it. 479 DEBUG(AssertSorted(Cache, NumSortedEntries)); 480 NonLocalDepInfo::iterator Entry = 481 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 482 std::make_pair(DirtyBB, MemDepResult())); 483 if (Entry != Cache.begin() && prior(Entry)->first == DirtyBB) 484 --Entry; 485 486 MemDepResult *ExistingResult = 0; 487 if (Entry != Cache.begin()+NumSortedEntries && 488 Entry->first == DirtyBB) { 489 // If we already have an entry, and if it isn't already dirty, the block 490 // is done. 491 if (!Entry->second.isDirty()) 492 continue; 493 494 // Otherwise, remember this slot so we can update the value. 495 ExistingResult = &Entry->second; 496 } 497 498 // If the dirty entry has a pointer, start scanning from it so we don't have 499 // to rescan the entire block. 500 BasicBlock::iterator ScanPos = DirtyBB->end(); 501 if (ExistingResult) { 502 if (Instruction *Inst = ExistingResult->getInst()) { 503 ScanPos = Inst; 504 // We're removing QueryInst's use of Inst. 505 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 506 QueryCS.getInstruction()); 507 } 508 } 509 510 // Find out if this block has a local dependency for QueryInst. 511 MemDepResult Dep; 512 513 if (ScanPos != DirtyBB->begin()) { 514 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 515 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 516 // No dependence found. If this is the entry block of the function, it is 517 // a clobber, otherwise it is non-local. 518 Dep = MemDepResult::getNonLocal(); 519 } else { 520 Dep = MemDepResult::getClobber(ScanPos); 521 } 522 523 // If we had a dirty entry for the block, update it. Otherwise, just add 524 // a new entry. 525 if (ExistingResult) 526 *ExistingResult = Dep; 527 else 528 Cache.push_back(std::make_pair(DirtyBB, Dep)); 529 530 // If the block has a dependency (i.e. it isn't completely transparent to 531 // the value), remember the association! 532 if (!Dep.isNonLocal()) { 533 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 534 // update this when we remove instructions. 535 if (Instruction *Inst = Dep.getInst()) 536 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 537 } else { 538 539 // If the block *is* completely transparent to the load, we need to check 540 // the predecessors of this block. Add them to our worklist. 541 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 542 DirtyBlocks.push_back(*PI); 543 } 544 } 545 546 return Cache; 547} 548 549/// getNonLocalPointerDependency - Perform a full dependency query for an 550/// access to the specified (non-volatile) memory location, returning the 551/// set of instructions that either define or clobber the value. 552/// 553/// This method assumes the pointer has a "NonLocal" dependency within its 554/// own block. 555/// 556void MemoryDependenceAnalysis:: 557getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB, 558 SmallVectorImpl<NonLocalDepEntry> &Result) { 559 assert(isa<PointerType>(Pointer->getType()) && 560 "Can't get pointer deps of a non-pointer!"); 561 Result.clear(); 562 563 // We know that the pointer value is live into FromBB find the def/clobbers 564 // from presecessors. 565 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType(); 566 uint64_t PointeeSize = AA->getTypeStoreSize(EltTy); 567 568 // This is the set of blocks we've inspected, and the pointer we consider in 569 // each block. Because of critical edges, we currently bail out if querying 570 // a block with multiple different pointers. This can happen during PHI 571 // translation. 572 DenseMap<BasicBlock*, Value*> Visited; 573 if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB, 574 Result, Visited, true)) 575 return; 576 Result.clear(); 577 Result.push_back(std::make_pair(FromBB, 578 MemDepResult::getClobber(FromBB->begin()))); 579} 580 581/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 582/// Pointer/PointeeSize using either cached information in Cache or by doing a 583/// lookup (which may use dirty cache info if available). If we do a lookup, 584/// add the result to the cache. 585MemDepResult MemoryDependenceAnalysis:: 586GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, 587 bool isLoad, BasicBlock *BB, 588 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 589 590 // Do a binary search to see if we already have an entry for this block in 591 // the cache set. If so, find it. 592 NonLocalDepInfo::iterator Entry = 593 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 594 std::make_pair(BB, MemDepResult())); 595 if (Entry != Cache->begin() && prior(Entry)->first == BB) 596 --Entry; 597 598 MemDepResult *ExistingResult = 0; 599 if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB) 600 ExistingResult = &Entry->second; 601 602 // If we have a cached entry, and it is non-dirty, use it as the value for 603 // this dependency. 604 if (ExistingResult && !ExistingResult->isDirty()) { 605 ++NumCacheNonLocalPtr; 606 return *ExistingResult; 607 } 608 609 // Otherwise, we have to scan for the value. If we have a dirty cache 610 // entry, start scanning from its position, otherwise we scan from the end 611 // of the block. 612 BasicBlock::iterator ScanPos = BB->end(); 613 if (ExistingResult && ExistingResult->getInst()) { 614 assert(ExistingResult->getInst()->getParent() == BB && 615 "Instruction invalidated?"); 616 ++NumCacheDirtyNonLocalPtr; 617 ScanPos = ExistingResult->getInst(); 618 619 // Eliminating the dirty entry from 'Cache', so update the reverse info. 620 ValueIsLoadPair CacheKey(Pointer, isLoad); 621 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 622 } else { 623 ++NumUncacheNonLocalPtr; 624 } 625 626 // Scan the block for the dependency. 627 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, 628 ScanPos, BB); 629 630 // If we had a dirty entry for the block, update it. Otherwise, just add 631 // a new entry. 632 if (ExistingResult) 633 *ExistingResult = Dep; 634 else 635 Cache->push_back(std::make_pair(BB, Dep)); 636 637 // If the block has a dependency (i.e. it isn't completely transparent to 638 // the value), remember the reverse association because we just added it 639 // to Cache! 640 if (Dep.isNonLocal()) 641 return Dep; 642 643 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 644 // update MemDep when we remove instructions. 645 Instruction *Inst = Dep.getInst(); 646 assert(Inst && "Didn't depend on anything?"); 647 ValueIsLoadPair CacheKey(Pointer, isLoad); 648 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 649 return Dep; 650} 651 652/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 653/// number of elements in the array that are already properly ordered. This is 654/// optimized for the case when only a few entries are added. 655static void 656SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 657 unsigned NumSortedEntries) { 658 switch (Cache.size() - NumSortedEntries) { 659 case 0: 660 // done, no new entries. 661 break; 662 case 2: { 663 // Two new entries, insert the last one into place. 664 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back(); 665 Cache.pop_back(); 666 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 667 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 668 Cache.insert(Entry, Val); 669 // FALL THROUGH. 670 } 671 case 1: 672 // One new entry, Just insert the new value at the appropriate position. 673 if (Cache.size() != 1) { 674 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back(); 675 Cache.pop_back(); 676 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 677 std::upper_bound(Cache.begin(), Cache.end(), Val); 678 Cache.insert(Entry, Val); 679 } 680 break; 681 default: 682 // Added many values, do a full scale sort. 683 std::sort(Cache.begin(), Cache.end()); 684 break; 685 } 686} 687 688/// isPHITranslatable - Return true if the specified computation is derived from 689/// a PHI node in the current block and if it is simple enough for us to handle. 690static bool isPHITranslatable(Instruction *Inst) { 691 if (isa<PHINode>(Inst)) 692 return true; 693 694 // We can handle bitcast of a PHI, but the PHI needs to be in the same block 695 // as the bitcast. 696 if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) 697 if (PHINode *PN = dyn_cast<PHINode>(BC->getOperand(0))) 698 if (PN->getParent() == BC->getParent()) 699 return true; 700 701 // We can translate a GEP that uses a PHI in the current block for at least 702 // one of its operands. 703 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) { 704 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) 705 if (PHINode *PN = dyn_cast<PHINode>(GEP->getOperand(i))) 706 if (PN->getParent() == GEP->getParent()) 707 return true; 708 } 709 710 // cerr << "MEMDEP: Could not PHI translate: " << *Pointer; 711 // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst)) 712 // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0); 713 714 return false; 715} 716 717/// PHITranslateForPred - Given a computation that satisfied the 718/// isPHITranslatable predicate, see if we can translate the computation into 719/// the specified predecessor block. If so, return that value. 720Value *MemoryDependenceAnalysis:: 721PHITranslatePointer(Value *InVal, BasicBlock *CurBB, BasicBlock *Pred, 722 const TargetData *TD) const { 723 // If the input value is not an instruction, or if it is not defined in CurBB, 724 // then we don't need to phi translate it. 725 Instruction *Inst = dyn_cast<Instruction>(InVal); 726 if (Inst == 0 || Inst->getParent() != CurBB) 727 return InVal; 728 729 if (PHINode *PN = dyn_cast<PHINode>(Inst)) 730 return PN->getIncomingValueForBlock(Pred); 731 732 // Handle bitcast of PHI. 733 if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) { 734 PHINode *BCPN = cast<PHINode>(BC->getOperand(0)); 735 Value *PHIIn = BCPN->getIncomingValueForBlock(Pred); 736 737 // Constants are trivial to phi translate. 738 if (Constant *C = dyn_cast<Constant>(PHIIn)) 739 return ConstantExpr::getBitCast(C, BC->getType()); 740 741 // Otherwise we have to see if a bitcasted version of the incoming pointer 742 // is available. If so, we can use it, otherwise we have to fail. 743 for (Value::use_iterator UI = PHIIn->use_begin(), E = PHIIn->use_end(); 744 UI != E; ++UI) { 745 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) 746 if (BCI->getType() == BC->getType()) 747 return BCI; 748 } 749 return 0; 750 } 751 752 // Handle getelementptr with at least one PHI operand. 753 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) { 754 SmallVector<Value*, 8> GEPOps; 755 BasicBlock *CurBB = GEP->getParent(); 756 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) { 757 Value *GEPOp = GEP->getOperand(i); 758 // No PHI translation is needed of operands whose values are live in to 759 // the predecessor block. 760 if (!isa<Instruction>(GEPOp) || 761 cast<Instruction>(GEPOp)->getParent() != CurBB) { 762 GEPOps.push_back(GEPOp); 763 continue; 764 } 765 766 // If the operand is a phi node, do phi translation. 767 if (PHINode *PN = dyn_cast<PHINode>(GEPOp)) { 768 GEPOps.push_back(PN->getIncomingValueForBlock(Pred)); 769 continue; 770 } 771 772 // Otherwise, we can't PHI translate this random value defined in this 773 // block. 774 return 0; 775 } 776 777 // Simplify the GEP to handle 'gep x, 0' -> x etc. 778 if (Value *V = SimplifyGEPInst(&GEPOps[0], GEPOps.size(), TD)) 779 return V; 780 781 782 // Scan to see if we have this GEP available. 783 Value *APHIOp = GEPOps[0]; 784 for (Value::use_iterator UI = APHIOp->use_begin(), E = APHIOp->use_end(); 785 UI != E; ++UI) { 786 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) 787 if (GEPI->getType() == GEP->getType() && 788 GEPI->getNumOperands() == GEPOps.size() && 789 GEPI->getParent()->getParent() == CurBB->getParent()) { 790 bool Mismatch = false; 791 for (unsigned i = 0, e = GEPOps.size(); i != e; ++i) 792 if (GEPI->getOperand(i) != GEPOps[i]) { 793 Mismatch = true; 794 break; 795 } 796 if (!Mismatch) 797 return GEPI; 798 } 799 } 800 return 0; 801 } 802 803 return 0; 804} 805 806/// InsertPHITranslatedPointer - Insert a computation of the PHI translated 807/// version of 'V' for the edge PredBB->CurBB into the end of the PredBB 808/// block. 809/// 810/// This is only called when PHITranslatePointer returns a value that doesn't 811/// dominate the block, so we don't need to handle the trivial cases here. 812Value *MemoryDependenceAnalysis:: 813InsertPHITranslatedPointer(Value *InVal, BasicBlock *CurBB, 814 BasicBlock *PredBB, const TargetData *TD) const { 815 // If the input value isn't an instruction in CurBB, it doesn't need phi 816 // translation. 817 Instruction *Inst = cast<Instruction>(InVal); 818 assert(Inst->getParent() == CurBB && "Doesn't need phi trans"); 819 820 // Handle bitcast of PHI. 821 if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) { 822 PHINode *BCPN = cast<PHINode>(BC->getOperand(0)); 823 Value *PHIIn = BCPN->getIncomingValueForBlock(PredBB); 824 825 // Otherwise insert a bitcast at the end of PredBB. 826 return new BitCastInst(PHIIn, InVal->getType(), 827 InVal->getName()+".phi.trans.insert", 828 PredBB->getTerminator()); 829 } 830 831 // Handle getelementptr with at least one PHI operand. 832 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) { 833 SmallVector<Value*, 8> GEPOps; 834 Value *APHIOp = 0; 835 BasicBlock *CurBB = GEP->getParent(); 836 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) { 837 GEPOps.push_back(GEP->getOperand(i)->DoPHITranslation(CurBB, PredBB)); 838 if (!isa<Constant>(GEPOps.back())) 839 APHIOp = GEPOps.back(); 840 } 841 842 GetElementPtrInst *Result = 843 GetElementPtrInst::Create(GEPOps[0], GEPOps.begin()+1, GEPOps.end(), 844 InVal->getName()+".phi.trans.insert", 845 PredBB->getTerminator()); 846 Result->setIsInBounds(GEP->isInBounds()); 847 return Result; 848 } 849 850 return 0; 851} 852 853/// getNonLocalPointerDepFromBB - Perform a dependency query based on 854/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 855/// results to the results vector and keep track of which blocks are visited in 856/// 'Visited'. 857/// 858/// This has special behavior for the first block queries (when SkipFirstBlock 859/// is true). In this special case, it ignores the contents of the specified 860/// block and starts returning dependence info for its predecessors. 861/// 862/// This function returns false on success, or true to indicate that it could 863/// not compute dependence information for some reason. This should be treated 864/// as a clobber dependence on the first instruction in the predecessor block. 865bool MemoryDependenceAnalysis:: 866getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize, 867 bool isLoad, BasicBlock *StartBB, 868 SmallVectorImpl<NonLocalDepEntry> &Result, 869 DenseMap<BasicBlock*, Value*> &Visited, 870 bool SkipFirstBlock) { 871 872 // Look up the cached info for Pointer. 873 ValueIsLoadPair CacheKey(Pointer, isLoad); 874 875 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo = 876 &NonLocalPointerDeps[CacheKey]; 877 NonLocalDepInfo *Cache = &CacheInfo->second; 878 879 // If we have valid cached information for exactly the block we are 880 // investigating, just return it with no recomputation. 881 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 882 // We have a fully cached result for this query then we can just return the 883 // cached results and populate the visited set. However, we have to verify 884 // that we don't already have conflicting results for these blocks. Check 885 // to ensure that if a block in the results set is in the visited set that 886 // it was for the same pointer query. 887 if (!Visited.empty()) { 888 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 889 I != E; ++I) { 890 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->first); 891 if (VI == Visited.end() || VI->second == Pointer) continue; 892 893 // We have a pointer mismatch in a block. Just return clobber, saying 894 // that something was clobbered in this result. We could also do a 895 // non-fully cached query, but there is little point in doing this. 896 return true; 897 } 898 } 899 900 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 901 I != E; ++I) { 902 Visited.insert(std::make_pair(I->first, Pointer)); 903 if (!I->second.isNonLocal()) 904 Result.push_back(*I); 905 } 906 ++NumCacheCompleteNonLocalPtr; 907 return false; 908 } 909 910 // Otherwise, either this is a new block, a block with an invalid cache 911 // pointer or one that we're about to invalidate by putting more info into it 912 // than its valid cache info. If empty, the result will be valid cache info, 913 // otherwise it isn't. 914 if (Cache->empty()) 915 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 916 else 917 CacheInfo->first = BBSkipFirstBlockPair(); 918 919 SmallVector<BasicBlock*, 32> Worklist; 920 Worklist.push_back(StartBB); 921 922 // Keep track of the entries that we know are sorted. Previously cached 923 // entries will all be sorted. The entries we add we only sort on demand (we 924 // don't insert every element into its sorted position). We know that we 925 // won't get any reuse from currently inserted values, because we don't 926 // revisit blocks after we insert info for them. 927 unsigned NumSortedEntries = Cache->size(); 928 DEBUG(AssertSorted(*Cache)); 929 930 while (!Worklist.empty()) { 931 BasicBlock *BB = Worklist.pop_back_val(); 932 933 // Skip the first block if we have it. 934 if (!SkipFirstBlock) { 935 // Analyze the dependency of *Pointer in FromBB. See if we already have 936 // been here. 937 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 938 939 // Get the dependency info for Pointer in BB. If we have cached 940 // information, we will use it, otherwise we compute it. 941 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 942 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad, 943 BB, Cache, NumSortedEntries); 944 945 // If we got a Def or Clobber, add this to the list of results. 946 if (!Dep.isNonLocal()) { 947 Result.push_back(NonLocalDepEntry(BB, Dep)); 948 continue; 949 } 950 } 951 952 // If 'Pointer' is an instruction defined in this block, then we need to do 953 // phi translation to change it into a value live in the predecessor block. 954 // If phi translation fails, then we can't continue dependence analysis. 955 Instruction *PtrInst = dyn_cast<Instruction>(Pointer); 956 bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB; 957 958 // If no PHI translation is needed, just add all the predecessors of this 959 // block to scan them as well. 960 if (!NeedsPHITranslation) { 961 SkipFirstBlock = false; 962 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 963 // Verify that we haven't looked at this block yet. 964 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 965 InsertRes = Visited.insert(std::make_pair(*PI, Pointer)); 966 if (InsertRes.second) { 967 // First time we've looked at *PI. 968 Worklist.push_back(*PI); 969 continue; 970 } 971 972 // If we have seen this block before, but it was with a different 973 // pointer then we have a phi translation failure and we have to treat 974 // this as a clobber. 975 if (InsertRes.first->second != Pointer) 976 goto PredTranslationFailure; 977 } 978 continue; 979 } 980 981 // If we do need to do phi translation, then there are a bunch of different 982 // cases, because we have to find a Value* live in the predecessor block. We 983 // know that PtrInst is defined in this block at least. 984 985 // We may have added values to the cache list before this PHI translation. 986 // If so, we haven't done anything to ensure that the cache remains sorted. 987 // Sort it now (if needed) so that recursive invocations of 988 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 989 // value will only see properly sorted cache arrays. 990 if (Cache && NumSortedEntries != Cache->size()) { 991 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 992 NumSortedEntries = Cache->size(); 993 } 994 995 // If this is a computation derived from a PHI node, use the suitably 996 // translated incoming values for each pred as the phi translated version. 997 if (isPHITranslatable(PtrInst)) { 998 Cache = 0; 999 1000 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1001 BasicBlock *Pred = *PI; 1002 Value *PredPtr = PHITranslatePointer(PtrInst, BB, Pred, TD); 1003 1004 // If PHI translation fails, bail out. 1005 if (PredPtr == 0) 1006 goto PredTranslationFailure; 1007 1008 // Check to see if we have already visited this pred block with another 1009 // pointer. If so, we can't do this lookup. This failure can occur 1010 // with PHI translation when a critical edge exists and the PHI node in 1011 // the successor translates to a pointer value different than the 1012 // pointer the block was first analyzed with. 1013 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1014 InsertRes = Visited.insert(std::make_pair(Pred, PredPtr)); 1015 1016 if (!InsertRes.second) { 1017 // If the predecessor was visited with PredPtr, then we already did 1018 // the analysis and can ignore it. 1019 if (InsertRes.first->second == PredPtr) 1020 continue; 1021 1022 // Otherwise, the block was previously analyzed with a different 1023 // pointer. We can't represent the result of this case, so we just 1024 // treat this as a phi translation failure. 1025 goto PredTranslationFailure; 1026 } 1027 1028 // FIXME: it is entirely possible that PHI translating will end up with 1029 // the same value. Consider PHI translating something like: 1030 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 1031 // to recurse here, pedantically speaking. 1032 1033 // If we have a problem phi translating, fall through to the code below 1034 // to handle the failure condition. 1035 if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred, 1036 Result, Visited)) 1037 goto PredTranslationFailure; 1038 } 1039 1040 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 1041 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1042 Cache = &CacheInfo->second; 1043 NumSortedEntries = Cache->size(); 1044 1045 // Since we did phi translation, the "Cache" set won't contain all of the 1046 // results for the query. This is ok (we can still use it to accelerate 1047 // specific block queries) but we can't do the fastpath "return all 1048 // results from the set" Clear out the indicator for this. 1049 CacheInfo->first = BBSkipFirstBlockPair(); 1050 SkipFirstBlock = false; 1051 continue; 1052 } 1053 1054 PredTranslationFailure: 1055 1056 if (Cache == 0) { 1057 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1058 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1059 Cache = &CacheInfo->second; 1060 NumSortedEntries = Cache->size(); 1061 } 1062 1063 // Since we did phi translation, the "Cache" set won't contain all of the 1064 // results for the query. This is ok (we can still use it to accelerate 1065 // specific block queries) but we can't do the fastpath "return all 1066 // results from the set" Clear out the indicator for this. 1067 CacheInfo->first = BBSkipFirstBlockPair(); 1068 1069 // If *nothing* works, mark the pointer as being clobbered by the first 1070 // instruction in this block. 1071 // 1072 // If this is the magic first block, return this as a clobber of the whole 1073 // incoming value. Since we can't phi translate to one of the predecessors, 1074 // we have to bail out. 1075 if (SkipFirstBlock) 1076 return true; 1077 1078 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 1079 assert(I != Cache->rend() && "Didn't find current block??"); 1080 if (I->first != BB) 1081 continue; 1082 1083 assert(I->second.isNonLocal() && 1084 "Should only be here with transparent block"); 1085 I->second = MemDepResult::getClobber(BB->begin()); 1086 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey); 1087 Result.push_back(*I); 1088 break; 1089 } 1090 } 1091 1092 // Okay, we're done now. If we added new values to the cache, re-sort it. 1093 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1094 DEBUG(AssertSorted(*Cache)); 1095 return false; 1096} 1097 1098/// RemoveCachedNonLocalPointerDependencies - If P exists in 1099/// CachedNonLocalPointerInfo, remove it. 1100void MemoryDependenceAnalysis:: 1101RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 1102 CachedNonLocalPointerInfo::iterator It = 1103 NonLocalPointerDeps.find(P); 1104 if (It == NonLocalPointerDeps.end()) return; 1105 1106 // Remove all of the entries in the BB->val map. This involves removing 1107 // instructions from the reverse map. 1108 NonLocalDepInfo &PInfo = It->second.second; 1109 1110 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 1111 Instruction *Target = PInfo[i].second.getInst(); 1112 if (Target == 0) continue; // Ignore non-local dep results. 1113 assert(Target->getParent() == PInfo[i].first); 1114 1115 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1116 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1117 } 1118 1119 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1120 NonLocalPointerDeps.erase(It); 1121} 1122 1123 1124/// invalidateCachedPointerInfo - This method is used to invalidate cached 1125/// information about the specified pointer, because it may be too 1126/// conservative in memdep. This is an optional call that can be used when 1127/// the client detects an equivalence between the pointer and some other 1128/// value and replaces the other value with ptr. This can make Ptr available 1129/// in more places that cached info does not necessarily keep. 1130void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1131 // If Ptr isn't really a pointer, just ignore it. 1132 if (!isa<PointerType>(Ptr->getType())) return; 1133 // Flush store info for the pointer. 1134 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1135 // Flush load info for the pointer. 1136 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1137} 1138 1139/// removeInstruction - Remove an instruction from the dependence analysis, 1140/// updating the dependence of instructions that previously depended on it. 1141/// This method attempts to keep the cache coherent using the reverse map. 1142void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1143 // Walk through the Non-local dependencies, removing this one as the value 1144 // for any cached queries. 1145 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1146 if (NLDI != NonLocalDeps.end()) { 1147 NonLocalDepInfo &BlockMap = NLDI->second.first; 1148 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1149 DI != DE; ++DI) 1150 if (Instruction *Inst = DI->second.getInst()) 1151 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1152 NonLocalDeps.erase(NLDI); 1153 } 1154 1155 // If we have a cached local dependence query for this instruction, remove it. 1156 // 1157 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1158 if (LocalDepEntry != LocalDeps.end()) { 1159 // Remove us from DepInst's reverse set now that the local dep info is gone. 1160 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1161 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1162 1163 // Remove this local dependency info. 1164 LocalDeps.erase(LocalDepEntry); 1165 } 1166 1167 // If we have any cached pointer dependencies on this instruction, remove 1168 // them. If the instruction has non-pointer type, then it can't be a pointer 1169 // base. 1170 1171 // Remove it from both the load info and the store info. The instruction 1172 // can't be in either of these maps if it is non-pointer. 1173 if (isa<PointerType>(RemInst->getType())) { 1174 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1175 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1176 } 1177 1178 // Loop over all of the things that depend on the instruction we're removing. 1179 // 1180 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1181 1182 // If we find RemInst as a clobber or Def in any of the maps for other values, 1183 // we need to replace its entry with a dirty version of the instruction after 1184 // it. If RemInst is a terminator, we use a null dirty value. 1185 // 1186 // Using a dirty version of the instruction after RemInst saves having to scan 1187 // the entire block to get to this point. 1188 MemDepResult NewDirtyVal; 1189 if (!RemInst->isTerminator()) 1190 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1191 1192 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1193 if (ReverseDepIt != ReverseLocalDeps.end()) { 1194 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1195 // RemInst can't be the terminator if it has local stuff depending on it. 1196 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1197 "Nothing can locally depend on a terminator"); 1198 1199 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1200 E = ReverseDeps.end(); I != E; ++I) { 1201 Instruction *InstDependingOnRemInst = *I; 1202 assert(InstDependingOnRemInst != RemInst && 1203 "Already removed our local dep info"); 1204 1205 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1206 1207 // Make sure to remember that new things depend on NewDepInst. 1208 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1209 "a local dep on this if it is a terminator!"); 1210 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1211 InstDependingOnRemInst)); 1212 } 1213 1214 ReverseLocalDeps.erase(ReverseDepIt); 1215 1216 // Add new reverse deps after scanning the set, to avoid invalidating the 1217 // 'ReverseDeps' reference. 1218 while (!ReverseDepsToAdd.empty()) { 1219 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1220 .insert(ReverseDepsToAdd.back().second); 1221 ReverseDepsToAdd.pop_back(); 1222 } 1223 } 1224 1225 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1226 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1227 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1228 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1229 I != E; ++I) { 1230 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1231 1232 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1233 // The information is now dirty! 1234 INLD.second = true; 1235 1236 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1237 DE = INLD.first.end(); DI != DE; ++DI) { 1238 if (DI->second.getInst() != RemInst) continue; 1239 1240 // Convert to a dirty entry for the subsequent instruction. 1241 DI->second = NewDirtyVal; 1242 1243 if (Instruction *NextI = NewDirtyVal.getInst()) 1244 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1245 } 1246 } 1247 1248 ReverseNonLocalDeps.erase(ReverseDepIt); 1249 1250 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1251 while (!ReverseDepsToAdd.empty()) { 1252 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1253 .insert(ReverseDepsToAdd.back().second); 1254 ReverseDepsToAdd.pop_back(); 1255 } 1256 } 1257 1258 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1259 // value in the NonLocalPointerDeps info. 1260 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1261 ReverseNonLocalPtrDeps.find(RemInst); 1262 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1263 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1264 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1265 1266 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1267 E = Set.end(); I != E; ++I) { 1268 ValueIsLoadPair P = *I; 1269 assert(P.getPointer() != RemInst && 1270 "Already removed NonLocalPointerDeps info for RemInst"); 1271 1272 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second; 1273 1274 // The cache is not valid for any specific block anymore. 1275 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair(); 1276 1277 // Update any entries for RemInst to use the instruction after it. 1278 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1279 DI != DE; ++DI) { 1280 if (DI->second.getInst() != RemInst) continue; 1281 1282 // Convert to a dirty entry for the subsequent instruction. 1283 DI->second = NewDirtyVal; 1284 1285 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1286 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1287 } 1288 1289 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1290 // subsequent value may invalidate the sortedness. 1291 std::sort(NLPDI.begin(), NLPDI.end()); 1292 } 1293 1294 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1295 1296 while (!ReversePtrDepsToAdd.empty()) { 1297 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1298 .insert(ReversePtrDepsToAdd.back().second); 1299 ReversePtrDepsToAdd.pop_back(); 1300 } 1301 } 1302 1303 1304 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1305 AA->deleteValue(RemInst); 1306 DEBUG(verifyRemoved(RemInst)); 1307} 1308/// verifyRemoved - Verify that the specified instruction does not occur 1309/// in our internal data structures. 1310void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1311 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1312 E = LocalDeps.end(); I != E; ++I) { 1313 assert(I->first != D && "Inst occurs in data structures"); 1314 assert(I->second.getInst() != D && 1315 "Inst occurs in data structures"); 1316 } 1317 1318 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1319 E = NonLocalPointerDeps.end(); I != E; ++I) { 1320 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1321 const NonLocalDepInfo &Val = I->second.second; 1322 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1323 II != E; ++II) 1324 assert(II->second.getInst() != D && "Inst occurs as NLPD value"); 1325 } 1326 1327 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1328 E = NonLocalDeps.end(); I != E; ++I) { 1329 assert(I->first != D && "Inst occurs in data structures"); 1330 const PerInstNLInfo &INLD = I->second; 1331 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1332 EE = INLD.first.end(); II != EE; ++II) 1333 assert(II->second.getInst() != D && "Inst occurs in data structures"); 1334 } 1335 1336 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1337 E = ReverseLocalDeps.end(); I != E; ++I) { 1338 assert(I->first != D && "Inst occurs in data structures"); 1339 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1340 EE = I->second.end(); II != EE; ++II) 1341 assert(*II != D && "Inst occurs in data structures"); 1342 } 1343 1344 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1345 E = ReverseNonLocalDeps.end(); 1346 I != E; ++I) { 1347 assert(I->first != D && "Inst occurs in data structures"); 1348 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1349 EE = I->second.end(); II != EE; ++II) 1350 assert(*II != D && "Inst occurs in data structures"); 1351 } 1352 1353 for (ReverseNonLocalPtrDepTy::const_iterator 1354 I = ReverseNonLocalPtrDeps.begin(), 1355 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1356 assert(I->first != D && "Inst occurs in rev NLPD map"); 1357 1358 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1359 E = I->second.end(); II != E; ++II) 1360 assert(*II != ValueIsLoadPair(D, false) && 1361 *II != ValueIsLoadPair(D, true) && 1362 "Inst occurs in ReverseNonLocalPtrDeps map"); 1363 } 1364 1365} 1366