MemoryDependenceAnalysis.cpp revision 847a84efd23a2c7d90429b82f6e0f19d1f913d9a
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/Dominators.h" 24#include "llvm/Analysis/InstructionSimplify.h" 25#include "llvm/Analysis/MemoryBuiltins.h" 26#include "llvm/Analysis/PHITransAddr.h" 27#include "llvm/ADT/Statistic.h" 28#include "llvm/ADT/STLExtras.h" 29#include "llvm/Support/PredIteratorCache.h" 30#include "llvm/Support/Debug.h" 31using namespace llvm; 32 33STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 34STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 35STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 36 37STATISTIC(NumCacheNonLocalPtr, 38 "Number of fully cached non-local ptr responses"); 39STATISTIC(NumCacheDirtyNonLocalPtr, 40 "Number of cached, but dirty, non-local ptr responses"); 41STATISTIC(NumUncacheNonLocalPtr, 42 "Number of uncached non-local ptr responses"); 43STATISTIC(NumCacheCompleteNonLocalPtr, 44 "Number of block queries that were completely cached"); 45 46char MemoryDependenceAnalysis::ID = 0; 47 48// Register this pass... 49INITIALIZE_PASS(MemoryDependenceAnalysis, "memdep", 50 "Memory Dependence Analysis", false, true); 51 52MemoryDependenceAnalysis::MemoryDependenceAnalysis() 53: FunctionPass(&ID), PredCache(0) { 54} 55MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 56} 57 58/// Clean up memory in between runs 59void MemoryDependenceAnalysis::releaseMemory() { 60 LocalDeps.clear(); 61 NonLocalDeps.clear(); 62 NonLocalPointerDeps.clear(); 63 ReverseLocalDeps.clear(); 64 ReverseNonLocalDeps.clear(); 65 ReverseNonLocalPtrDeps.clear(); 66 PredCache->clear(); 67} 68 69 70 71/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 72/// 73void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 74 AU.setPreservesAll(); 75 AU.addRequiredTransitive<AliasAnalysis>(); 76} 77 78bool MemoryDependenceAnalysis::runOnFunction(Function &) { 79 AA = &getAnalysis<AliasAnalysis>(); 80 if (PredCache == 0) 81 PredCache.reset(new PredIteratorCache()); 82 return false; 83} 84 85/// RemoveFromReverseMap - This is a helper function that removes Val from 86/// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry. 87template <typename KeyTy> 88static void RemoveFromReverseMap(DenseMap<Instruction*, 89 SmallPtrSet<KeyTy, 4> > &ReverseMap, 90 Instruction *Inst, KeyTy Val) { 91 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator 92 InstIt = ReverseMap.find(Inst); 93 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); 94 bool Found = InstIt->second.erase(Val); 95 assert(Found && "Invalid reverse map!"); Found=Found; 96 if (InstIt->second.empty()) 97 ReverseMap.erase(InstIt); 98} 99 100 101/// getCallSiteDependencyFrom - Private helper for finding the local 102/// dependencies of a call site. 103MemDepResult MemoryDependenceAnalysis:: 104getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 105 BasicBlock::iterator ScanIt, BasicBlock *BB) { 106 // Walk backwards through the block, looking for dependencies 107 while (ScanIt != BB->begin()) { 108 Instruction *Inst = --ScanIt; 109 110 // If this inst is a memory op, get the pointer it accessed 111 Value *Pointer = 0; 112 uint64_t PointerSize = 0; 113 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) { 114 Pointer = S->getPointerOperand(); 115 PointerSize = AA->getTypeStoreSize(S->getOperand(0)->getType()); 116 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 117 Pointer = V->getOperand(0); 118 PointerSize = AA->getTypeStoreSize(V->getType()); 119 } else if (const CallInst *CI = isFreeCall(Inst)) { 120 Pointer = CI->getArgOperand(0); 121 // calls to free() erase the entire structure 122 PointerSize = ~0ULL; 123 } else if (CallSite InstCS = cast<Value>(Inst)) { 124 // Debug intrinsics don't cause dependences. 125 if (isa<DbgInfoIntrinsic>(Inst)) continue; 126 // If these two calls do not interfere, look past it. 127 switch (AA->getModRefInfo(CS, InstCS)) { 128 case AliasAnalysis::NoModRef: 129 // If the two calls don't interact (e.g. InstCS is readnone) keep 130 // scanning. 131 continue; 132 case AliasAnalysis::Ref: 133 // If the two calls read the same memory locations and CS is a readonly 134 // function, then we have two cases: 1) the calls may not interfere with 135 // each other at all. 2) the calls may produce the same value. In case 136 // #1 we want to ignore the values, in case #2, we want to return Inst 137 // as a Def dependence. This allows us to CSE in cases like: 138 // X = strlen(P); 139 // memchr(...); 140 // Y = strlen(P); // Y = X 141 if (isReadOnlyCall) { 142 if (CS.getCalledFunction() != 0 && 143 CS.getCalledFunction() == InstCS.getCalledFunction()) 144 return MemDepResult::getDef(Inst); 145 // Ignore unrelated read/read call dependences. 146 continue; 147 } 148 // FALL THROUGH 149 default: 150 return MemDepResult::getClobber(Inst); 151 } 152 } else { 153 // Non-memory instruction. 154 continue; 155 } 156 157 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef) 158 return MemDepResult::getClobber(Inst); 159 } 160 161 // No dependence found. If this is the entry block of the function, it is a 162 // clobber, otherwise it is non-local. 163 if (BB != &BB->getParent()->getEntryBlock()) 164 return MemDepResult::getNonLocal(); 165 return MemDepResult::getClobber(ScanIt); 166} 167 168/// getPointerDependencyFrom - Return the instruction on which a memory 169/// location depends. If isLoad is true, this routine ignore may-aliases with 170/// read-only operations. 171MemDepResult MemoryDependenceAnalysis:: 172getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, 173 BasicBlock::iterator ScanIt, BasicBlock *BB) { 174 175 Value *InvariantTag = 0; 176 177 // Walk backwards through the basic block, looking for dependencies. 178 while (ScanIt != BB->begin()) { 179 Instruction *Inst = --ScanIt; 180 181 // If we're in an invariant region, no dependencies can be found before 182 // we pass an invariant-begin marker. 183 if (InvariantTag == Inst) { 184 InvariantTag = 0; 185 continue; 186 } 187 188 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 189 // Debug intrinsics don't cause dependences. 190 if (isa<DbgInfoIntrinsic>(Inst)) continue; 191 192 // If we pass an invariant-end marker, then we've just entered an 193 // invariant region and can start ignoring dependencies. 194 if (II->getIntrinsicID() == Intrinsic::invariant_end) { 195 // FIXME: This only considers queries directly on the invariant-tagged 196 // pointer, not on query pointers that are indexed off of them. It'd 197 // be nice to handle that at some point. 198 AliasAnalysis::AliasResult R = AA->alias(II->getArgOperand(2), MemPtr); 199 if (R == AliasAnalysis::MustAlias) { 200 InvariantTag = II->getArgOperand(0); 201 continue; 202 } 203 204 // If we reach a lifetime begin or end marker, then the query ends here 205 // because the value is undefined. 206 } else if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 207 // FIXME: This only considers queries directly on the invariant-tagged 208 // pointer, not on query pointers that are indexed off of them. It'd 209 // be nice to handle that at some point. 210 AliasAnalysis::AliasResult R = AA->alias(II->getArgOperand(1), MemPtr); 211 if (R == AliasAnalysis::MustAlias) 212 return MemDepResult::getDef(II); 213 } 214 } 215 216 // If we're querying on a load and we're in an invariant region, we're done 217 // at this point. Nothing a load depends on can live in an invariant region. 218 if (isLoad && InvariantTag) continue; 219 220 // Values depend on loads if the pointers are must aliased. This means that 221 // a load depends on another must aliased load from the same value. 222 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 223 Value *Pointer = LI->getPointerOperand(); 224 uint64_t PointerSize = AA->getTypeStoreSize(LI->getType()); 225 226 // If we found a pointer, check if it could be the same as our pointer. 227 AliasAnalysis::AliasResult R = 228 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 229 if (R == AliasAnalysis::NoAlias) 230 continue; 231 232 // May-alias loads don't depend on each other without a dependence. 233 if (isLoad && R == AliasAnalysis::MayAlias) 234 continue; 235 // Stores depend on may and must aliased loads, loads depend on must-alias 236 // loads. 237 return MemDepResult::getDef(Inst); 238 } 239 240 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 241 // There can't be stores to the value we care about inside an 242 // invariant region. 243 if (InvariantTag) continue; 244 245 // If alias analysis can tell that this store is guaranteed to not modify 246 // the query pointer, ignore it. Use getModRefInfo to handle cases where 247 // the query pointer points to constant memory etc. 248 if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef) 249 continue; 250 251 // Ok, this store might clobber the query pointer. Check to see if it is 252 // a must alias: in this case, we want to return this as a def. 253 Value *Pointer = SI->getPointerOperand(); 254 uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); 255 256 // If we found a pointer, check if it could be the same as our pointer. 257 AliasAnalysis::AliasResult R = 258 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 259 260 if (R == AliasAnalysis::NoAlias) 261 continue; 262 if (R == AliasAnalysis::MayAlias) 263 return MemDepResult::getClobber(Inst); 264 return MemDepResult::getDef(Inst); 265 } 266 267 // If this is an allocation, and if we know that the accessed pointer is to 268 // the allocation, return Def. This means that there is no dependence and 269 // the access can be optimized based on that. For example, a load could 270 // turn into undef. 271 // Note: Only determine this to be a malloc if Inst is the malloc call, not 272 // a subsequent bitcast of the malloc call result. There can be stores to 273 // the malloced memory between the malloc call and its bitcast uses, and we 274 // need to continue scanning until the malloc call. 275 if (isa<AllocaInst>(Inst) || 276 (isa<CallInst>(Inst) && extractMallocCall(Inst))) { 277 Value *AccessPtr = MemPtr->getUnderlyingObject(); 278 279 if (AccessPtr == Inst || 280 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 281 return MemDepResult::getDef(Inst); 282 continue; 283 } 284 285 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 286 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) { 287 case AliasAnalysis::NoModRef: 288 // If the call has no effect on the queried pointer, just ignore it. 289 continue; 290 case AliasAnalysis::Mod: 291 // If we're in an invariant region, we can ignore calls that ONLY 292 // modify the pointer. 293 if (InvariantTag) continue; 294 return MemDepResult::getClobber(Inst); 295 case AliasAnalysis::Ref: 296 // If the call is known to never store to the pointer, and if this is a 297 // load query, we can safely ignore it (scan past it). 298 if (isLoad) 299 continue; 300 default: 301 // Otherwise, there is a potential dependence. Return a clobber. 302 return MemDepResult::getClobber(Inst); 303 } 304 } 305 306 // No dependence found. If this is the entry block of the function, it is a 307 // clobber, otherwise it is non-local. 308 if (BB != &BB->getParent()->getEntryBlock()) 309 return MemDepResult::getNonLocal(); 310 return MemDepResult::getClobber(ScanIt); 311} 312 313/// getDependency - Return the instruction on which a memory operation 314/// depends. 315MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 316 Instruction *ScanPos = QueryInst; 317 318 // Check for a cached result 319 MemDepResult &LocalCache = LocalDeps[QueryInst]; 320 321 // If the cached entry is non-dirty, just return it. Note that this depends 322 // on MemDepResult's default constructing to 'dirty'. 323 if (!LocalCache.isDirty()) 324 return LocalCache; 325 326 // Otherwise, if we have a dirty entry, we know we can start the scan at that 327 // instruction, which may save us some work. 328 if (Instruction *Inst = LocalCache.getInst()) { 329 ScanPos = Inst; 330 331 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 332 } 333 334 BasicBlock *QueryParent = QueryInst->getParent(); 335 336 Value *MemPtr = 0; 337 uint64_t MemSize = 0; 338 339 // Do the scan. 340 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 341 // No dependence found. If this is the entry block of the function, it is a 342 // clobber, otherwise it is non-local. 343 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 344 LocalCache = MemDepResult::getNonLocal(); 345 else 346 LocalCache = MemDepResult::getClobber(QueryInst); 347 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) { 348 // If this is a volatile store, don't mess around with it. Just return the 349 // previous instruction as a clobber. 350 if (SI->isVolatile()) 351 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 352 else { 353 MemPtr = SI->getPointerOperand(); 354 MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); 355 } 356 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) { 357 // If this is a volatile load, don't mess around with it. Just return the 358 // previous instruction as a clobber. 359 if (LI->isVolatile()) 360 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 361 else { 362 MemPtr = LI->getPointerOperand(); 363 MemSize = AA->getTypeStoreSize(LI->getType()); 364 } 365 } else if (const CallInst *CI = isFreeCall(QueryInst)) { 366 MemPtr = CI->getArgOperand(0); 367 // calls to free() erase the entire structure, not just a field. 368 MemSize = ~0UL; 369 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 370 int IntrinsicID = 0; // Intrinsic IDs start at 1. 371 IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst); 372 if (II) 373 IntrinsicID = II->getIntrinsicID(); 374 375 switch (IntrinsicID) { 376 case Intrinsic::lifetime_start: 377 case Intrinsic::lifetime_end: 378 case Intrinsic::invariant_start: 379 MemPtr = II->getArgOperand(1); 380 MemSize = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue(); 381 break; 382 case Intrinsic::invariant_end: 383 MemPtr = II->getArgOperand(2); 384 MemSize = cast<ConstantInt>(II->getArgOperand(1))->getZExtValue(); 385 break; 386 default: 387 CallSite QueryCS(QueryInst); 388 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 389 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 390 QueryParent); 391 break; 392 } 393 } else { 394 // Non-memory instruction. 395 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 396 } 397 398 // If we need to do a pointer scan, make it happen. 399 if (MemPtr) { 400 bool isLoad = !QueryInst->mayWriteToMemory(); 401 if (IntrinsicInst *II = dyn_cast<MemoryUseIntrinsic>(QueryInst)) { 402 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end; 403 } 404 LocalCache = getPointerDependencyFrom(MemPtr, MemSize, isLoad, ScanPos, 405 QueryParent); 406 } 407 408 // Remember the result! 409 if (Instruction *I = LocalCache.getInst()) 410 ReverseLocalDeps[I].insert(QueryInst); 411 412 return LocalCache; 413} 414 415#ifndef NDEBUG 416/// AssertSorted - This method is used when -debug is specified to verify that 417/// cache arrays are properly kept sorted. 418static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 419 int Count = -1) { 420 if (Count == -1) Count = Cache.size(); 421 if (Count == 0) return; 422 423 for (unsigned i = 1; i != unsigned(Count); ++i) 424 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 425} 426#endif 427 428/// getNonLocalCallDependency - Perform a full dependency query for the 429/// specified call, returning the set of blocks that the value is 430/// potentially live across. The returned set of results will include a 431/// "NonLocal" result for all blocks where the value is live across. 432/// 433/// This method assumes the instruction returns a "NonLocal" dependency 434/// within its own block. 435/// 436/// This returns a reference to an internal data structure that may be 437/// invalidated on the next non-local query or when an instruction is 438/// removed. Clients must copy this data if they want it around longer than 439/// that. 440const MemoryDependenceAnalysis::NonLocalDepInfo & 441MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 442 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 443 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 444 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 445 NonLocalDepInfo &Cache = CacheP.first; 446 447 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 448 /// the cached case, this can happen due to instructions being deleted etc. In 449 /// the uncached case, this starts out as the set of predecessors we care 450 /// about. 451 SmallVector<BasicBlock*, 32> DirtyBlocks; 452 453 if (!Cache.empty()) { 454 // Okay, we have a cache entry. If we know it is not dirty, just return it 455 // with no computation. 456 if (!CacheP.second) { 457 ++NumCacheNonLocal; 458 return Cache; 459 } 460 461 // If we already have a partially computed set of results, scan them to 462 // determine what is dirty, seeding our initial DirtyBlocks worklist. 463 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 464 I != E; ++I) 465 if (I->getResult().isDirty()) 466 DirtyBlocks.push_back(I->getBB()); 467 468 // Sort the cache so that we can do fast binary search lookups below. 469 std::sort(Cache.begin(), Cache.end()); 470 471 ++NumCacheDirtyNonLocal; 472 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 473 // << Cache.size() << " cached: " << *QueryInst; 474 } else { 475 // Seed DirtyBlocks with each of the preds of QueryInst's block. 476 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 477 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 478 DirtyBlocks.push_back(*PI); 479 ++NumUncacheNonLocal; 480 } 481 482 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 483 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 484 485 SmallPtrSet<BasicBlock*, 64> Visited; 486 487 unsigned NumSortedEntries = Cache.size(); 488 DEBUG(AssertSorted(Cache)); 489 490 // Iterate while we still have blocks to update. 491 while (!DirtyBlocks.empty()) { 492 BasicBlock *DirtyBB = DirtyBlocks.back(); 493 DirtyBlocks.pop_back(); 494 495 // Already processed this block? 496 if (!Visited.insert(DirtyBB)) 497 continue; 498 499 // Do a binary search to see if we already have an entry for this block in 500 // the cache set. If so, find it. 501 DEBUG(AssertSorted(Cache, NumSortedEntries)); 502 NonLocalDepInfo::iterator Entry = 503 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 504 NonLocalDepEntry(DirtyBB)); 505 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB) 506 --Entry; 507 508 NonLocalDepEntry *ExistingResult = 0; 509 if (Entry != Cache.begin()+NumSortedEntries && 510 Entry->getBB() == DirtyBB) { 511 // If we already have an entry, and if it isn't already dirty, the block 512 // is done. 513 if (!Entry->getResult().isDirty()) 514 continue; 515 516 // Otherwise, remember this slot so we can update the value. 517 ExistingResult = &*Entry; 518 } 519 520 // If the dirty entry has a pointer, start scanning from it so we don't have 521 // to rescan the entire block. 522 BasicBlock::iterator ScanPos = DirtyBB->end(); 523 if (ExistingResult) { 524 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 525 ScanPos = Inst; 526 // We're removing QueryInst's use of Inst. 527 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 528 QueryCS.getInstruction()); 529 } 530 } 531 532 // Find out if this block has a local dependency for QueryInst. 533 MemDepResult Dep; 534 535 if (ScanPos != DirtyBB->begin()) { 536 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 537 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 538 // No dependence found. If this is the entry block of the function, it is 539 // a clobber, otherwise it is non-local. 540 Dep = MemDepResult::getNonLocal(); 541 } else { 542 Dep = MemDepResult::getClobber(ScanPos); 543 } 544 545 // If we had a dirty entry for the block, update it. Otherwise, just add 546 // a new entry. 547 if (ExistingResult) 548 ExistingResult->setResult(Dep); 549 else 550 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 551 552 // If the block has a dependency (i.e. it isn't completely transparent to 553 // the value), remember the association! 554 if (!Dep.isNonLocal()) { 555 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 556 // update this when we remove instructions. 557 if (Instruction *Inst = Dep.getInst()) 558 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 559 } else { 560 561 // If the block *is* completely transparent to the load, we need to check 562 // the predecessors of this block. Add them to our worklist. 563 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 564 DirtyBlocks.push_back(*PI); 565 } 566 } 567 568 return Cache; 569} 570 571/// getNonLocalPointerDependency - Perform a full dependency query for an 572/// access to the specified (non-volatile) memory location, returning the 573/// set of instructions that either define or clobber the value. 574/// 575/// This method assumes the pointer has a "NonLocal" dependency within its 576/// own block. 577/// 578void MemoryDependenceAnalysis:: 579getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB, 580 SmallVectorImpl<NonLocalDepResult> &Result) { 581 assert(Pointer->getType()->isPointerTy() && 582 "Can't get pointer deps of a non-pointer!"); 583 Result.clear(); 584 585 // We know that the pointer value is live into FromBB find the def/clobbers 586 // from presecessors. 587 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType(); 588 uint64_t PointeeSize = AA->getTypeStoreSize(EltTy); 589 590 PHITransAddr Address(Pointer, TD); 591 592 // This is the set of blocks we've inspected, and the pointer we consider in 593 // each block. Because of critical edges, we currently bail out if querying 594 // a block with multiple different pointers. This can happen during PHI 595 // translation. 596 DenseMap<BasicBlock*, Value*> Visited; 597 if (!getNonLocalPointerDepFromBB(Address, PointeeSize, isLoad, FromBB, 598 Result, Visited, true)) 599 return; 600 Result.clear(); 601 Result.push_back(NonLocalDepResult(FromBB, 602 MemDepResult::getClobber(FromBB->begin()), 603 Pointer)); 604} 605 606/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 607/// Pointer/PointeeSize using either cached information in Cache or by doing a 608/// lookup (which may use dirty cache info if available). If we do a lookup, 609/// add the result to the cache. 610MemDepResult MemoryDependenceAnalysis:: 611GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, 612 bool isLoad, BasicBlock *BB, 613 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 614 615 // Do a binary search to see if we already have an entry for this block in 616 // the cache set. If so, find it. 617 NonLocalDepInfo::iterator Entry = 618 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 619 NonLocalDepEntry(BB)); 620 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 621 --Entry; 622 623 NonLocalDepEntry *ExistingResult = 0; 624 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 625 ExistingResult = &*Entry; 626 627 // If we have a cached entry, and it is non-dirty, use it as the value for 628 // this dependency. 629 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 630 ++NumCacheNonLocalPtr; 631 return ExistingResult->getResult(); 632 } 633 634 // Otherwise, we have to scan for the value. If we have a dirty cache 635 // entry, start scanning from its position, otherwise we scan from the end 636 // of the block. 637 BasicBlock::iterator ScanPos = BB->end(); 638 if (ExistingResult && ExistingResult->getResult().getInst()) { 639 assert(ExistingResult->getResult().getInst()->getParent() == BB && 640 "Instruction invalidated?"); 641 ++NumCacheDirtyNonLocalPtr; 642 ScanPos = ExistingResult->getResult().getInst(); 643 644 // Eliminating the dirty entry from 'Cache', so update the reverse info. 645 ValueIsLoadPair CacheKey(Pointer, isLoad); 646 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 647 } else { 648 ++NumUncacheNonLocalPtr; 649 } 650 651 // Scan the block for the dependency. 652 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, 653 ScanPos, BB); 654 655 // If we had a dirty entry for the block, update it. Otherwise, just add 656 // a new entry. 657 if (ExistingResult) 658 ExistingResult->setResult(Dep); 659 else 660 Cache->push_back(NonLocalDepEntry(BB, Dep)); 661 662 // If the block has a dependency (i.e. it isn't completely transparent to 663 // the value), remember the reverse association because we just added it 664 // to Cache! 665 if (Dep.isNonLocal()) 666 return Dep; 667 668 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 669 // update MemDep when we remove instructions. 670 Instruction *Inst = Dep.getInst(); 671 assert(Inst && "Didn't depend on anything?"); 672 ValueIsLoadPair CacheKey(Pointer, isLoad); 673 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 674 return Dep; 675} 676 677/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 678/// number of elements in the array that are already properly ordered. This is 679/// optimized for the case when only a few entries are added. 680static void 681SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 682 unsigned NumSortedEntries) { 683 switch (Cache.size() - NumSortedEntries) { 684 case 0: 685 // done, no new entries. 686 break; 687 case 2: { 688 // Two new entries, insert the last one into place. 689 NonLocalDepEntry Val = Cache.back(); 690 Cache.pop_back(); 691 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 692 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 693 Cache.insert(Entry, Val); 694 // FALL THROUGH. 695 } 696 case 1: 697 // One new entry, Just insert the new value at the appropriate position. 698 if (Cache.size() != 1) { 699 NonLocalDepEntry Val = Cache.back(); 700 Cache.pop_back(); 701 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 702 std::upper_bound(Cache.begin(), Cache.end(), Val); 703 Cache.insert(Entry, Val); 704 } 705 break; 706 default: 707 // Added many values, do a full scale sort. 708 std::sort(Cache.begin(), Cache.end()); 709 break; 710 } 711} 712 713/// getNonLocalPointerDepFromBB - Perform a dependency query based on 714/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 715/// results to the results vector and keep track of which blocks are visited in 716/// 'Visited'. 717/// 718/// This has special behavior for the first block queries (when SkipFirstBlock 719/// is true). In this special case, it ignores the contents of the specified 720/// block and starts returning dependence info for its predecessors. 721/// 722/// This function returns false on success, or true to indicate that it could 723/// not compute dependence information for some reason. This should be treated 724/// as a clobber dependence on the first instruction in the predecessor block. 725bool MemoryDependenceAnalysis:: 726getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, 727 bool isLoad, BasicBlock *StartBB, 728 SmallVectorImpl<NonLocalDepResult> &Result, 729 DenseMap<BasicBlock*, Value*> &Visited, 730 bool SkipFirstBlock) { 731 732 // Look up the cached info for Pointer. 733 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 734 735 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo = 736 &NonLocalPointerDeps[CacheKey]; 737 NonLocalDepInfo *Cache = &CacheInfo->second; 738 739 // If we have valid cached information for exactly the block we are 740 // investigating, just return it with no recomputation. 741 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 742 // We have a fully cached result for this query then we can just return the 743 // cached results and populate the visited set. However, we have to verify 744 // that we don't already have conflicting results for these blocks. Check 745 // to ensure that if a block in the results set is in the visited set that 746 // it was for the same pointer query. 747 if (!Visited.empty()) { 748 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 749 I != E; ++I) { 750 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 751 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 752 continue; 753 754 // We have a pointer mismatch in a block. Just return clobber, saying 755 // that something was clobbered in this result. We could also do a 756 // non-fully cached query, but there is little point in doing this. 757 return true; 758 } 759 } 760 761 Value *Addr = Pointer.getAddr(); 762 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 763 I != E; ++I) { 764 Visited.insert(std::make_pair(I->getBB(), Addr)); 765 if (!I->getResult().isNonLocal()) 766 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 767 } 768 ++NumCacheCompleteNonLocalPtr; 769 return false; 770 } 771 772 // Otherwise, either this is a new block, a block with an invalid cache 773 // pointer or one that we're about to invalidate by putting more info into it 774 // than its valid cache info. If empty, the result will be valid cache info, 775 // otherwise it isn't. 776 if (Cache->empty()) 777 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 778 else 779 CacheInfo->first = BBSkipFirstBlockPair(); 780 781 SmallVector<BasicBlock*, 32> Worklist; 782 Worklist.push_back(StartBB); 783 784 // Keep track of the entries that we know are sorted. Previously cached 785 // entries will all be sorted. The entries we add we only sort on demand (we 786 // don't insert every element into its sorted position). We know that we 787 // won't get any reuse from currently inserted values, because we don't 788 // revisit blocks after we insert info for them. 789 unsigned NumSortedEntries = Cache->size(); 790 DEBUG(AssertSorted(*Cache)); 791 792 while (!Worklist.empty()) { 793 BasicBlock *BB = Worklist.pop_back_val(); 794 795 // Skip the first block if we have it. 796 if (!SkipFirstBlock) { 797 // Analyze the dependency of *Pointer in FromBB. See if we already have 798 // been here. 799 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 800 801 // Get the dependency info for Pointer in BB. If we have cached 802 // information, we will use it, otherwise we compute it. 803 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 804 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer.getAddr(), PointeeSize, 805 isLoad, BB, Cache, 806 NumSortedEntries); 807 808 // If we got a Def or Clobber, add this to the list of results. 809 if (!Dep.isNonLocal()) { 810 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 811 continue; 812 } 813 } 814 815 // If 'Pointer' is an instruction defined in this block, then we need to do 816 // phi translation to change it into a value live in the predecessor block. 817 // If not, we just add the predecessors to the worklist and scan them with 818 // the same Pointer. 819 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 820 SkipFirstBlock = false; 821 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 822 // Verify that we haven't looked at this block yet. 823 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 824 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 825 if (InsertRes.second) { 826 // First time we've looked at *PI. 827 Worklist.push_back(*PI); 828 continue; 829 } 830 831 // If we have seen this block before, but it was with a different 832 // pointer then we have a phi translation failure and we have to treat 833 // this as a clobber. 834 if (InsertRes.first->second != Pointer.getAddr()) 835 goto PredTranslationFailure; 836 } 837 continue; 838 } 839 840 // We do need to do phi translation, if we know ahead of time we can't phi 841 // translate this value, don't even try. 842 if (!Pointer.IsPotentiallyPHITranslatable()) 843 goto PredTranslationFailure; 844 845 // We may have added values to the cache list before this PHI translation. 846 // If so, we haven't done anything to ensure that the cache remains sorted. 847 // Sort it now (if needed) so that recursive invocations of 848 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 849 // value will only see properly sorted cache arrays. 850 if (Cache && NumSortedEntries != Cache->size()) { 851 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 852 NumSortedEntries = Cache->size(); 853 } 854 Cache = 0; 855 856 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 857 BasicBlock *Pred = *PI; 858 859 // Get the PHI translated pointer in this predecessor. This can fail if 860 // not translatable, in which case the getAddr() returns null. 861 PHITransAddr PredPointer(Pointer); 862 PredPointer.PHITranslateValue(BB, Pred, 0); 863 864 Value *PredPtrVal = PredPointer.getAddr(); 865 866 // Check to see if we have already visited this pred block with another 867 // pointer. If so, we can't do this lookup. This failure can occur 868 // with PHI translation when a critical edge exists and the PHI node in 869 // the successor translates to a pointer value different than the 870 // pointer the block was first analyzed with. 871 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 872 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 873 874 if (!InsertRes.second) { 875 // If the predecessor was visited with PredPtr, then we already did 876 // the analysis and can ignore it. 877 if (InsertRes.first->second == PredPtrVal) 878 continue; 879 880 // Otherwise, the block was previously analyzed with a different 881 // pointer. We can't represent the result of this case, so we just 882 // treat this as a phi translation failure. 883 goto PredTranslationFailure; 884 } 885 886 // If PHI translation was unable to find an available pointer in this 887 // predecessor, then we have to assume that the pointer is clobbered in 888 // that predecessor. We can still do PRE of the load, which would insert 889 // a computation of the pointer in this predecessor. 890 if (PredPtrVal == 0) { 891 // Add the entry to the Result list. 892 NonLocalDepResult Entry(Pred, 893 MemDepResult::getClobber(Pred->getTerminator()), 894 PredPtrVal); 895 Result.push_back(Entry); 896 897 // Since we had a phi translation failure, the cache for CacheKey won't 898 // include all of the entries that we need to immediately satisfy future 899 // queries. Mark this in NonLocalPointerDeps by setting the 900 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 901 // cached value to do more work but not miss the phi trans failure. 902 NonLocalPointerDeps[CacheKey].first = BBSkipFirstBlockPair(); 903 continue; 904 } 905 906 // FIXME: it is entirely possible that PHI translating will end up with 907 // the same value. Consider PHI translating something like: 908 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 909 // to recurse here, pedantically speaking. 910 911 // If we have a problem phi translating, fall through to the code below 912 // to handle the failure condition. 913 if (getNonLocalPointerDepFromBB(PredPointer, PointeeSize, isLoad, Pred, 914 Result, Visited)) 915 goto PredTranslationFailure; 916 } 917 918 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 919 CacheInfo = &NonLocalPointerDeps[CacheKey]; 920 Cache = &CacheInfo->second; 921 NumSortedEntries = Cache->size(); 922 923 // Since we did phi translation, the "Cache" set won't contain all of the 924 // results for the query. This is ok (we can still use it to accelerate 925 // specific block queries) but we can't do the fastpath "return all 926 // results from the set" Clear out the indicator for this. 927 CacheInfo->first = BBSkipFirstBlockPair(); 928 SkipFirstBlock = false; 929 continue; 930 931 PredTranslationFailure: 932 933 if (Cache == 0) { 934 // Refresh the CacheInfo/Cache pointer if it got invalidated. 935 CacheInfo = &NonLocalPointerDeps[CacheKey]; 936 Cache = &CacheInfo->second; 937 NumSortedEntries = Cache->size(); 938 } 939 940 // Since we failed phi translation, the "Cache" set won't contain all of the 941 // results for the query. This is ok (we can still use it to accelerate 942 // specific block queries) but we can't do the fastpath "return all 943 // results from the set". Clear out the indicator for this. 944 CacheInfo->first = BBSkipFirstBlockPair(); 945 946 // If *nothing* works, mark the pointer as being clobbered by the first 947 // instruction in this block. 948 // 949 // If this is the magic first block, return this as a clobber of the whole 950 // incoming value. Since we can't phi translate to one of the predecessors, 951 // we have to bail out. 952 if (SkipFirstBlock) 953 return true; 954 955 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 956 assert(I != Cache->rend() && "Didn't find current block??"); 957 if (I->getBB() != BB) 958 continue; 959 960 assert(I->getResult().isNonLocal() && 961 "Should only be here with transparent block"); 962 I->setResult(MemDepResult::getClobber(BB->begin())); 963 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey); 964 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 965 Pointer.getAddr())); 966 break; 967 } 968 } 969 970 // Okay, we're done now. If we added new values to the cache, re-sort it. 971 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 972 DEBUG(AssertSorted(*Cache)); 973 return false; 974} 975 976/// RemoveCachedNonLocalPointerDependencies - If P exists in 977/// CachedNonLocalPointerInfo, remove it. 978void MemoryDependenceAnalysis:: 979RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 980 CachedNonLocalPointerInfo::iterator It = 981 NonLocalPointerDeps.find(P); 982 if (It == NonLocalPointerDeps.end()) return; 983 984 // Remove all of the entries in the BB->val map. This involves removing 985 // instructions from the reverse map. 986 NonLocalDepInfo &PInfo = It->second.second; 987 988 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 989 Instruction *Target = PInfo[i].getResult().getInst(); 990 if (Target == 0) continue; // Ignore non-local dep results. 991 assert(Target->getParent() == PInfo[i].getBB()); 992 993 // Eliminating the dirty entry from 'Cache', so update the reverse info. 994 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 995 } 996 997 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 998 NonLocalPointerDeps.erase(It); 999} 1000 1001 1002/// invalidateCachedPointerInfo - This method is used to invalidate cached 1003/// information about the specified pointer, because it may be too 1004/// conservative in memdep. This is an optional call that can be used when 1005/// the client detects an equivalence between the pointer and some other 1006/// value and replaces the other value with ptr. This can make Ptr available 1007/// in more places that cached info does not necessarily keep. 1008void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1009 // If Ptr isn't really a pointer, just ignore it. 1010 if (!Ptr->getType()->isPointerTy()) return; 1011 // Flush store info for the pointer. 1012 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1013 // Flush load info for the pointer. 1014 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1015} 1016 1017/// invalidateCachedPredecessors - Clear the PredIteratorCache info. 1018/// This needs to be done when the CFG changes, e.g., due to splitting 1019/// critical edges. 1020void MemoryDependenceAnalysis::invalidateCachedPredecessors() { 1021 PredCache->clear(); 1022} 1023 1024/// removeInstruction - Remove an instruction from the dependence analysis, 1025/// updating the dependence of instructions that previously depended on it. 1026/// This method attempts to keep the cache coherent using the reverse map. 1027void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1028 // Walk through the Non-local dependencies, removing this one as the value 1029 // for any cached queries. 1030 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1031 if (NLDI != NonLocalDeps.end()) { 1032 NonLocalDepInfo &BlockMap = NLDI->second.first; 1033 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1034 DI != DE; ++DI) 1035 if (Instruction *Inst = DI->getResult().getInst()) 1036 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1037 NonLocalDeps.erase(NLDI); 1038 } 1039 1040 // If we have a cached local dependence query for this instruction, remove it. 1041 // 1042 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1043 if (LocalDepEntry != LocalDeps.end()) { 1044 // Remove us from DepInst's reverse set now that the local dep info is gone. 1045 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1046 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1047 1048 // Remove this local dependency info. 1049 LocalDeps.erase(LocalDepEntry); 1050 } 1051 1052 // If we have any cached pointer dependencies on this instruction, remove 1053 // them. If the instruction has non-pointer type, then it can't be a pointer 1054 // base. 1055 1056 // Remove it from both the load info and the store info. The instruction 1057 // can't be in either of these maps if it is non-pointer. 1058 if (RemInst->getType()->isPointerTy()) { 1059 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1060 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1061 } 1062 1063 // Loop over all of the things that depend on the instruction we're removing. 1064 // 1065 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1066 1067 // If we find RemInst as a clobber or Def in any of the maps for other values, 1068 // we need to replace its entry with a dirty version of the instruction after 1069 // it. If RemInst is a terminator, we use a null dirty value. 1070 // 1071 // Using a dirty version of the instruction after RemInst saves having to scan 1072 // the entire block to get to this point. 1073 MemDepResult NewDirtyVal; 1074 if (!RemInst->isTerminator()) 1075 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1076 1077 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1078 if (ReverseDepIt != ReverseLocalDeps.end()) { 1079 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1080 // RemInst can't be the terminator if it has local stuff depending on it. 1081 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1082 "Nothing can locally depend on a terminator"); 1083 1084 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1085 E = ReverseDeps.end(); I != E; ++I) { 1086 Instruction *InstDependingOnRemInst = *I; 1087 assert(InstDependingOnRemInst != RemInst && 1088 "Already removed our local dep info"); 1089 1090 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1091 1092 // Make sure to remember that new things depend on NewDepInst. 1093 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1094 "a local dep on this if it is a terminator!"); 1095 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1096 InstDependingOnRemInst)); 1097 } 1098 1099 ReverseLocalDeps.erase(ReverseDepIt); 1100 1101 // Add new reverse deps after scanning the set, to avoid invalidating the 1102 // 'ReverseDeps' reference. 1103 while (!ReverseDepsToAdd.empty()) { 1104 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1105 .insert(ReverseDepsToAdd.back().second); 1106 ReverseDepsToAdd.pop_back(); 1107 } 1108 } 1109 1110 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1111 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1112 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1113 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1114 I != E; ++I) { 1115 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1116 1117 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1118 // The information is now dirty! 1119 INLD.second = true; 1120 1121 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1122 DE = INLD.first.end(); DI != DE; ++DI) { 1123 if (DI->getResult().getInst() != RemInst) continue; 1124 1125 // Convert to a dirty entry for the subsequent instruction. 1126 DI->setResult(NewDirtyVal); 1127 1128 if (Instruction *NextI = NewDirtyVal.getInst()) 1129 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1130 } 1131 } 1132 1133 ReverseNonLocalDeps.erase(ReverseDepIt); 1134 1135 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1136 while (!ReverseDepsToAdd.empty()) { 1137 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1138 .insert(ReverseDepsToAdd.back().second); 1139 ReverseDepsToAdd.pop_back(); 1140 } 1141 } 1142 1143 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1144 // value in the NonLocalPointerDeps info. 1145 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1146 ReverseNonLocalPtrDeps.find(RemInst); 1147 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1148 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1149 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1150 1151 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1152 E = Set.end(); I != E; ++I) { 1153 ValueIsLoadPair P = *I; 1154 assert(P.getPointer() != RemInst && 1155 "Already removed NonLocalPointerDeps info for RemInst"); 1156 1157 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second; 1158 1159 // The cache is not valid for any specific block anymore. 1160 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair(); 1161 1162 // Update any entries for RemInst to use the instruction after it. 1163 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1164 DI != DE; ++DI) { 1165 if (DI->getResult().getInst() != RemInst) continue; 1166 1167 // Convert to a dirty entry for the subsequent instruction. 1168 DI->setResult(NewDirtyVal); 1169 1170 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1171 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1172 } 1173 1174 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1175 // subsequent value may invalidate the sortedness. 1176 std::sort(NLPDI.begin(), NLPDI.end()); 1177 } 1178 1179 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1180 1181 while (!ReversePtrDepsToAdd.empty()) { 1182 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1183 .insert(ReversePtrDepsToAdd.back().second); 1184 ReversePtrDepsToAdd.pop_back(); 1185 } 1186 } 1187 1188 1189 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1190 AA->deleteValue(RemInst); 1191 DEBUG(verifyRemoved(RemInst)); 1192} 1193/// verifyRemoved - Verify that the specified instruction does not occur 1194/// in our internal data structures. 1195void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1196 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1197 E = LocalDeps.end(); I != E; ++I) { 1198 assert(I->first != D && "Inst occurs in data structures"); 1199 assert(I->second.getInst() != D && 1200 "Inst occurs in data structures"); 1201 } 1202 1203 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1204 E = NonLocalPointerDeps.end(); I != E; ++I) { 1205 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1206 const NonLocalDepInfo &Val = I->second.second; 1207 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1208 II != E; ++II) 1209 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1210 } 1211 1212 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1213 E = NonLocalDeps.end(); I != E; ++I) { 1214 assert(I->first != D && "Inst occurs in data structures"); 1215 const PerInstNLInfo &INLD = I->second; 1216 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1217 EE = INLD.first.end(); II != EE; ++II) 1218 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1219 } 1220 1221 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1222 E = ReverseLocalDeps.end(); I != E; ++I) { 1223 assert(I->first != D && "Inst occurs in data structures"); 1224 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1225 EE = I->second.end(); II != EE; ++II) 1226 assert(*II != D && "Inst occurs in data structures"); 1227 } 1228 1229 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1230 E = ReverseNonLocalDeps.end(); 1231 I != E; ++I) { 1232 assert(I->first != D && "Inst occurs in data structures"); 1233 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1234 EE = I->second.end(); II != EE; ++II) 1235 assert(*II != D && "Inst occurs in data structures"); 1236 } 1237 1238 for (ReverseNonLocalPtrDepTy::const_iterator 1239 I = ReverseNonLocalPtrDeps.begin(), 1240 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1241 assert(I->first != D && "Inst occurs in rev NLPD map"); 1242 1243 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1244 E = I->second.end(); II != E; ++II) 1245 assert(*II != ValueIsLoadPair(D, false) && 1246 *II != ValueIsLoadPair(D, true) && 1247 "Inst occurs in ReverseNonLocalPtrDeps map"); 1248 } 1249 1250} 1251