MemoryDependenceAnalysis.cpp revision 1df9859c40492511b8aa4321eb76496005d3b75b
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... 49static RegisterPass<MemoryDependenceAnalysis> X("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 (isFreeCall(Inst)) { 120 Pointer = Inst->getOperand(1); 121 // calls to free() erase the entire structure 122 PointerSize = ~0ULL; 123 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) { 124 // Debug intrinsics don't cause dependences. 125 if (isa<DbgInfoIntrinsic>(Inst)) continue; 126 CallSite InstCS = CallSite::get(Inst); 127 // If these two calls do not interfere, look past it. 128 switch (AA->getModRefInfo(CS, InstCS)) { 129 case AliasAnalysis::NoModRef: 130 // If the two calls don't interact (e.g. InstCS is readnone) keep 131 // scanning. 132 continue; 133 case AliasAnalysis::Ref: 134 // If the two calls read the same memory locations and CS is a readonly 135 // function, then we have two cases: 1) the calls may not interfere with 136 // each other at all. 2) the calls may produce the same value. In case 137 // #1 we want to ignore the values, in case #2, we want to return Inst 138 // as a Def dependence. This allows us to CSE in cases like: 139 // X = strlen(P); 140 // memchr(...); 141 // Y = strlen(P); // Y = X 142 if (isReadOnlyCall) { 143 if (CS.getCalledFunction() != 0 && 144 CS.getCalledFunction() == InstCS.getCalledFunction()) 145 return MemDepResult::getDef(Inst); 146 // Ignore unrelated read/read call dependences. 147 continue; 148 } 149 // FALL THROUGH 150 default: 151 return MemDepResult::getClobber(Inst); 152 } 153 } else { 154 // Non-memory instruction. 155 continue; 156 } 157 158 if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef) 159 return MemDepResult::getClobber(Inst); 160 } 161 162 // No dependence found. If this is the entry block of the function, it is a 163 // clobber, otherwise it is non-local. 164 if (BB != &BB->getParent()->getEntryBlock()) 165 return MemDepResult::getNonLocal(); 166 return MemDepResult::getClobber(ScanIt); 167} 168 169/// getPointerDependencyFrom - Return the instruction on which a memory 170/// location depends. If isLoad is true, this routine ignore may-aliases with 171/// read-only operations. 172MemDepResult MemoryDependenceAnalysis:: 173getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, 174 BasicBlock::iterator ScanIt, BasicBlock *BB) { 175 176 Value *InvariantTag = 0; 177 178 // Walk backwards through the basic block, looking for dependencies. 179 while (ScanIt != BB->begin()) { 180 Instruction *Inst = --ScanIt; 181 182 // If we're in an invariant region, no dependencies can be found before 183 // we pass an invariant-begin marker. 184 if (InvariantTag == Inst) { 185 InvariantTag = 0; 186 continue; 187 } 188 189 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 190 // Debug intrinsics don't cause dependences. 191 if (isa<DbgInfoIntrinsic>(Inst)) continue; 192 193 // If we pass an invariant-end marker, then we've just entered an 194 // invariant region and can start ignoring dependencies. 195 if (II->getIntrinsicID() == Intrinsic::invariant_end) { 196 // FIXME: This only considers queries directly on the invariant-tagged 197 // pointer, not on query pointers that are indexed off of them. It'd 198 // be nice to handle that at some point. 199 AliasAnalysis::AliasResult R = 200 AA->alias(II->getOperand(3), ~0U, MemPtr, ~0U); 201 if (R == AliasAnalysis::MustAlias) { 202 InvariantTag = II->getOperand(1); 203 continue; 204 } 205 206 // If we reach a lifetime begin or end marker, then the query ends here 207 // because the value is undefined. 208 } else if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 209 // FIXME: This only considers queries directly on the invariant-tagged 210 // pointer, not on query pointers that are indexed off of them. It'd 211 // be nice to handle that at some point. 212 AliasAnalysis::AliasResult R = 213 AA->alias(II->getOperand(2), ~0U, MemPtr, ~0U); 214 if (R == AliasAnalysis::MustAlias) 215 return MemDepResult::getDef(II); 216 } 217 } 218 219 // If we're querying on a load and we're in an invariant region, we're done 220 // at this point. Nothing a load depends on can live in an invariant region. 221 if (isLoad && InvariantTag) continue; 222 223 // Values depend on loads if the pointers are must aliased. This means that 224 // a load depends on another must aliased load from the same value. 225 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 226 Value *Pointer = LI->getPointerOperand(); 227 uint64_t PointerSize = AA->getTypeStoreSize(LI->getType()); 228 229 // If we found a pointer, check if it could be the same as our pointer. 230 AliasAnalysis::AliasResult R = 231 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 232 if (R == AliasAnalysis::NoAlias) 233 continue; 234 235 // May-alias loads don't depend on each other without a dependence. 236 if (isLoad && R == AliasAnalysis::MayAlias) 237 continue; 238 // Stores depend on may and must aliased loads, loads depend on must-alias 239 // loads. 240 return MemDepResult::getDef(Inst); 241 } 242 243 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 244 // There can't be stores to the value we care about inside an 245 // invariant region. 246 if (InvariantTag) continue; 247 248 // If alias analysis can tell that this store is guaranteed to not modify 249 // the query pointer, ignore it. Use getModRefInfo to handle cases where 250 // the query pointer points to constant memory etc. 251 if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef) 252 continue; 253 254 // Ok, this store might clobber the query pointer. Check to see if it is 255 // a must alias: in this case, we want to return this as a def. 256 Value *Pointer = SI->getPointerOperand(); 257 uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); 258 259 // If we found a pointer, check if it could be the same as our pointer. 260 AliasAnalysis::AliasResult R = 261 AA->alias(Pointer, PointerSize, MemPtr, MemSize); 262 263 if (R == AliasAnalysis::NoAlias) 264 continue; 265 if (R == AliasAnalysis::MayAlias) 266 return MemDepResult::getClobber(Inst); 267 return MemDepResult::getDef(Inst); 268 } 269 270 // If this is an allocation, and if we know that the accessed pointer is to 271 // the allocation, return Def. This means that there is no dependence and 272 // the access can be optimized based on that. For example, a load could 273 // turn into undef. 274 // Note: Only determine this to be a malloc if Inst is the malloc call, not 275 // a subsequent bitcast of the malloc call result. There can be stores to 276 // the malloced memory between the malloc call and its bitcast uses, and we 277 // need to continue scanning until the malloc call. 278 if (isa<AllocaInst>(Inst) || 279 (isa<CallInst>(Inst) && extractMallocCall(Inst))) { 280 Value *AccessPtr = MemPtr->getUnderlyingObject(); 281 282 if (AccessPtr == Inst || 283 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 284 return MemDepResult::getDef(Inst); 285 continue; 286 } 287 288 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 289 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) { 290 case AliasAnalysis::NoModRef: 291 // If the call has no effect on the queried pointer, just ignore it. 292 continue; 293 case AliasAnalysis::Mod: 294 // If we're in an invariant region, we can ignore calls that ONLY 295 // modify the pointer. 296 if (InvariantTag) continue; 297 return MemDepResult::getClobber(Inst); 298 case AliasAnalysis::Ref: 299 // If the call is known to never store to the pointer, and if this is a 300 // load query, we can safely ignore it (scan past it). 301 if (isLoad) 302 continue; 303 default: 304 // Otherwise, there is a potential dependence. Return a clobber. 305 return MemDepResult::getClobber(Inst); 306 } 307 } 308 309 // No dependence found. If this is the entry block of the function, it is a 310 // clobber, otherwise it is non-local. 311 if (BB != &BB->getParent()->getEntryBlock()) 312 return MemDepResult::getNonLocal(); 313 return MemDepResult::getClobber(ScanIt); 314} 315 316/// getDependency - Return the instruction on which a memory operation 317/// depends. 318MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 319 Instruction *ScanPos = QueryInst; 320 321 // Check for a cached result 322 MemDepResult &LocalCache = LocalDeps[QueryInst]; 323 324 // If the cached entry is non-dirty, just return it. Note that this depends 325 // on MemDepResult's default constructing to 'dirty'. 326 if (!LocalCache.isDirty()) 327 return LocalCache; 328 329 // Otherwise, if we have a dirty entry, we know we can start the scan at that 330 // instruction, which may save us some work. 331 if (Instruction *Inst = LocalCache.getInst()) { 332 ScanPos = Inst; 333 334 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 335 } 336 337 BasicBlock *QueryParent = QueryInst->getParent(); 338 339 Value *MemPtr = 0; 340 uint64_t MemSize = 0; 341 342 // Do the scan. 343 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 344 // No dependence found. If this is the entry block of the function, it is a 345 // clobber, otherwise it is non-local. 346 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 347 LocalCache = MemDepResult::getNonLocal(); 348 else 349 LocalCache = MemDepResult::getClobber(QueryInst); 350 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) { 351 // If this is a volatile store, don't mess around with it. Just return the 352 // previous instruction as a clobber. 353 if (SI->isVolatile()) 354 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 355 else { 356 MemPtr = SI->getPointerOperand(); 357 MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); 358 } 359 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) { 360 // If this is a volatile load, don't mess around with it. Just return the 361 // previous instruction as a clobber. 362 if (LI->isVolatile()) 363 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 364 else { 365 MemPtr = LI->getPointerOperand(); 366 MemSize = AA->getTypeStoreSize(LI->getType()); 367 } 368 } else if (isFreeCall(QueryInst)) { 369 MemPtr = QueryInst->getOperand(1); 370 // calls to free() erase the entire structure, not just a field. 371 MemSize = ~0UL; 372 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 373 int IntrinsicID = 0; // Intrinsic IDs start at 1. 374 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst)) 375 IntrinsicID = II->getIntrinsicID(); 376 377 switch (IntrinsicID) { 378 case Intrinsic::lifetime_start: 379 case Intrinsic::lifetime_end: 380 case Intrinsic::invariant_start: 381 MemPtr = QueryInst->getOperand(2); 382 MemSize = cast<ConstantInt>(QueryInst->getOperand(1))->getZExtValue(); 383 break; 384 case Intrinsic::invariant_end: 385 MemPtr = QueryInst->getOperand(3); 386 MemSize = cast<ConstantInt>(QueryInst->getOperand(2))->getZExtValue(); 387 break; 388 default: 389 CallSite QueryCS = CallSite::get(QueryInst); 390 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 391 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 392 QueryParent); 393 break; 394 } 395 } else { 396 // Non-memory instruction. 397 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 398 } 399 400 // If we need to do a pointer scan, make it happen. 401 if (MemPtr) { 402 bool isLoad = !QueryInst->mayWriteToMemory(); 403 if (IntrinsicInst *II = dyn_cast<MemoryUseIntrinsic>(QueryInst)) { 404 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end; 405 } 406 LocalCache = getPointerDependencyFrom(MemPtr, MemSize, isLoad, ScanPos, 407 QueryParent); 408 } 409 410 // Remember the result! 411 if (Instruction *I = LocalCache.getInst()) 412 ReverseLocalDeps[I].insert(QueryInst); 413 414 return LocalCache; 415} 416 417#ifndef NDEBUG 418/// AssertSorted - This method is used when -debug is specified to verify that 419/// cache arrays are properly kept sorted. 420static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 421 int Count = -1) { 422 if (Count == -1) Count = Cache.size(); 423 if (Count == 0) return; 424 425 for (unsigned i = 1; i != unsigned(Count); ++i) 426 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 427} 428#endif 429 430/// getNonLocalCallDependency - Perform a full dependency query for the 431/// specified call, returning the set of blocks that the value is 432/// potentially live across. The returned set of results will include a 433/// "NonLocal" result for all blocks where the value is live across. 434/// 435/// This method assumes the instruction returns a "NonLocal" dependency 436/// within its own block. 437/// 438/// This returns a reference to an internal data structure that may be 439/// invalidated on the next non-local query or when an instruction is 440/// removed. Clients must copy this data if they want it around longer than 441/// that. 442const MemoryDependenceAnalysis::NonLocalDepInfo & 443MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 444 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 445 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 446 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 447 NonLocalDepInfo &Cache = CacheP.first; 448 449 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 450 /// the cached case, this can happen due to instructions being deleted etc. In 451 /// the uncached case, this starts out as the set of predecessors we care 452 /// about. 453 SmallVector<BasicBlock*, 32> DirtyBlocks; 454 455 if (!Cache.empty()) { 456 // Okay, we have a cache entry. If we know it is not dirty, just return it 457 // with no computation. 458 if (!CacheP.second) { 459 NumCacheNonLocal++; 460 return Cache; 461 } 462 463 // If we already have a partially computed set of results, scan them to 464 // determine what is dirty, seeding our initial DirtyBlocks worklist. 465 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 466 I != E; ++I) 467 if (I->getResult().isDirty()) 468 DirtyBlocks.push_back(I->getBB()); 469 470 // Sort the cache so that we can do fast binary search lookups below. 471 std::sort(Cache.begin(), Cache.end()); 472 473 ++NumCacheDirtyNonLocal; 474 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 475 // << Cache.size() << " cached: " << *QueryInst; 476 } else { 477 // Seed DirtyBlocks with each of the preds of QueryInst's block. 478 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 479 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 480 DirtyBlocks.push_back(*PI); 481 NumUncacheNonLocal++; 482 } 483 484 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 485 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 486 487 SmallPtrSet<BasicBlock*, 64> Visited; 488 489 unsigned NumSortedEntries = Cache.size(); 490 DEBUG(AssertSorted(Cache)); 491 492 // Iterate while we still have blocks to update. 493 while (!DirtyBlocks.empty()) { 494 BasicBlock *DirtyBB = DirtyBlocks.back(); 495 DirtyBlocks.pop_back(); 496 497 // Already processed this block? 498 if (!Visited.insert(DirtyBB)) 499 continue; 500 501 // Do a binary search to see if we already have an entry for this block in 502 // the cache set. If so, find it. 503 DEBUG(AssertSorted(Cache, NumSortedEntries)); 504 NonLocalDepInfo::iterator Entry = 505 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 506 NonLocalDepEntry(DirtyBB)); 507 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB) 508 --Entry; 509 510 NonLocalDepEntry *ExistingResult = 0; 511 if (Entry != Cache.begin()+NumSortedEntries && 512 Entry->getBB() == DirtyBB) { 513 // If we already have an entry, and if it isn't already dirty, the block 514 // is done. 515 if (!Entry->getResult().isDirty()) 516 continue; 517 518 // Otherwise, remember this slot so we can update the value. 519 ExistingResult = &*Entry; 520 } 521 522 // If the dirty entry has a pointer, start scanning from it so we don't have 523 // to rescan the entire block. 524 BasicBlock::iterator ScanPos = DirtyBB->end(); 525 if (ExistingResult) { 526 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 527 ScanPos = Inst; 528 // We're removing QueryInst's use of Inst. 529 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 530 QueryCS.getInstruction()); 531 } 532 } 533 534 // Find out if this block has a local dependency for QueryInst. 535 MemDepResult Dep; 536 537 if (ScanPos != DirtyBB->begin()) { 538 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 539 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 540 // No dependence found. If this is the entry block of the function, it is 541 // a clobber, otherwise it is non-local. 542 Dep = MemDepResult::getNonLocal(); 543 } else { 544 Dep = MemDepResult::getClobber(ScanPos); 545 } 546 547 // If we had a dirty entry for the block, update it. Otherwise, just add 548 // a new entry. 549 if (ExistingResult) 550 ExistingResult->setResult(Dep); 551 else 552 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 553 554 // If the block has a dependency (i.e. it isn't completely transparent to 555 // the value), remember the association! 556 if (!Dep.isNonLocal()) { 557 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 558 // update this when we remove instructions. 559 if (Instruction *Inst = Dep.getInst()) 560 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 561 } else { 562 563 // If the block *is* completely transparent to the load, we need to check 564 // the predecessors of this block. Add them to our worklist. 565 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 566 DirtyBlocks.push_back(*PI); 567 } 568 } 569 570 return Cache; 571} 572 573/// getNonLocalPointerDependency - Perform a full dependency query for an 574/// access to the specified (non-volatile) memory location, returning the 575/// set of instructions that either define or clobber the value. 576/// 577/// This method assumes the pointer has a "NonLocal" dependency within its 578/// own block. 579/// 580void MemoryDependenceAnalysis:: 581getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB, 582 SmallVectorImpl<NonLocalDepResult> &Result) { 583 assert(Pointer->getType()->isPointerTy() && 584 "Can't get pointer deps of a non-pointer!"); 585 Result.clear(); 586 587 // We know that the pointer value is live into FromBB find the def/clobbers 588 // from presecessors. 589 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType(); 590 uint64_t PointeeSize = AA->getTypeStoreSize(EltTy); 591 592 PHITransAddr Address(Pointer, TD); 593 594 // This is the set of blocks we've inspected, and the pointer we consider in 595 // each block. Because of critical edges, we currently bail out if querying 596 // a block with multiple different pointers. This can happen during PHI 597 // translation. 598 DenseMap<BasicBlock*, Value*> Visited; 599 if (!getNonLocalPointerDepFromBB(Address, PointeeSize, isLoad, FromBB, 600 Result, Visited, true)) 601 return; 602 Result.clear(); 603 Result.push_back(NonLocalDepResult(FromBB, 604 MemDepResult::getClobber(FromBB->begin()), 605 Pointer)); 606} 607 608/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 609/// Pointer/PointeeSize using either cached information in Cache or by doing a 610/// lookup (which may use dirty cache info if available). If we do a lookup, 611/// add the result to the cache. 612MemDepResult MemoryDependenceAnalysis:: 613GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, 614 bool isLoad, BasicBlock *BB, 615 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 616 617 // Do a binary search to see if we already have an entry for this block in 618 // the cache set. If so, find it. 619 NonLocalDepInfo::iterator Entry = 620 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 621 NonLocalDepEntry(BB)); 622 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 623 --Entry; 624 625 NonLocalDepEntry *ExistingResult = 0; 626 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 627 ExistingResult = &*Entry; 628 629 // If we have a cached entry, and it is non-dirty, use it as the value for 630 // this dependency. 631 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 632 ++NumCacheNonLocalPtr; 633 return ExistingResult->getResult(); 634 } 635 636 // Otherwise, we have to scan for the value. If we have a dirty cache 637 // entry, start scanning from its position, otherwise we scan from the end 638 // of the block. 639 BasicBlock::iterator ScanPos = BB->end(); 640 if (ExistingResult && ExistingResult->getResult().getInst()) { 641 assert(ExistingResult->getResult().getInst()->getParent() == BB && 642 "Instruction invalidated?"); 643 ++NumCacheDirtyNonLocalPtr; 644 ScanPos = ExistingResult->getResult().getInst(); 645 646 // Eliminating the dirty entry from 'Cache', so update the reverse info. 647 ValueIsLoadPair CacheKey(Pointer, isLoad); 648 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 649 } else { 650 ++NumUncacheNonLocalPtr; 651 } 652 653 // Scan the block for the dependency. 654 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, 655 ScanPos, BB); 656 657 // If we had a dirty entry for the block, update it. Otherwise, just add 658 // a new entry. 659 if (ExistingResult) 660 ExistingResult->setResult(Dep); 661 else 662 Cache->push_back(NonLocalDepEntry(BB, Dep)); 663 664 // If the block has a dependency (i.e. it isn't completely transparent to 665 // the value), remember the reverse association because we just added it 666 // to Cache! 667 if (Dep.isNonLocal()) 668 return Dep; 669 670 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 671 // update MemDep when we remove instructions. 672 Instruction *Inst = Dep.getInst(); 673 assert(Inst && "Didn't depend on anything?"); 674 ValueIsLoadPair CacheKey(Pointer, isLoad); 675 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 676 return Dep; 677} 678 679/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 680/// number of elements in the array that are already properly ordered. This is 681/// optimized for the case when only a few entries are added. 682static void 683SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 684 unsigned NumSortedEntries) { 685 switch (Cache.size() - NumSortedEntries) { 686 case 0: 687 // done, no new entries. 688 break; 689 case 2: { 690 // Two new entries, insert the last one into place. 691 NonLocalDepEntry Val = Cache.back(); 692 Cache.pop_back(); 693 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 694 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 695 Cache.insert(Entry, Val); 696 // FALL THROUGH. 697 } 698 case 1: 699 // One new entry, Just insert the new value at the appropriate position. 700 if (Cache.size() != 1) { 701 NonLocalDepEntry Val = Cache.back(); 702 Cache.pop_back(); 703 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 704 std::upper_bound(Cache.begin(), Cache.end(), Val); 705 Cache.insert(Entry, Val); 706 } 707 break; 708 default: 709 // Added many values, do a full scale sort. 710 std::sort(Cache.begin(), Cache.end()); 711 break; 712 } 713} 714 715/// getNonLocalPointerDepFromBB - Perform a dependency query based on 716/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 717/// results to the results vector and keep track of which blocks are visited in 718/// 'Visited'. 719/// 720/// This has special behavior for the first block queries (when SkipFirstBlock 721/// is true). In this special case, it ignores the contents of the specified 722/// block and starts returning dependence info for its predecessors. 723/// 724/// This function returns false on success, or true to indicate that it could 725/// not compute dependence information for some reason. This should be treated 726/// as a clobber dependence on the first instruction in the predecessor block. 727bool MemoryDependenceAnalysis:: 728getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, 729 bool isLoad, BasicBlock *StartBB, 730 SmallVectorImpl<NonLocalDepResult> &Result, 731 DenseMap<BasicBlock*, Value*> &Visited, 732 bool SkipFirstBlock) { 733 734 // Look up the cached info for Pointer. 735 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 736 737 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo = 738 &NonLocalPointerDeps[CacheKey]; 739 NonLocalDepInfo *Cache = &CacheInfo->second; 740 741 // If we have valid cached information for exactly the block we are 742 // investigating, just return it with no recomputation. 743 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 744 // We have a fully cached result for this query then we can just return the 745 // cached results and populate the visited set. However, we have to verify 746 // that we don't already have conflicting results for these blocks. Check 747 // to ensure that if a block in the results set is in the visited set that 748 // it was for the same pointer query. 749 if (!Visited.empty()) { 750 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 751 I != E; ++I) { 752 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 753 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 754 continue; 755 756 // We have a pointer mismatch in a block. Just return clobber, saying 757 // that something was clobbered in this result. We could also do a 758 // non-fully cached query, but there is little point in doing this. 759 return true; 760 } 761 } 762 763 Value *Addr = Pointer.getAddr(); 764 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 765 I != E; ++I) { 766 Visited.insert(std::make_pair(I->getBB(), Addr)); 767 if (!I->getResult().isNonLocal()) 768 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 769 } 770 ++NumCacheCompleteNonLocalPtr; 771 return false; 772 } 773 774 // Otherwise, either this is a new block, a block with an invalid cache 775 // pointer or one that we're about to invalidate by putting more info into it 776 // than its valid cache info. If empty, the result will be valid cache info, 777 // otherwise it isn't. 778 if (Cache->empty()) 779 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 780 else 781 CacheInfo->first = BBSkipFirstBlockPair(); 782 783 SmallVector<BasicBlock*, 32> Worklist; 784 Worklist.push_back(StartBB); 785 786 // Keep track of the entries that we know are sorted. Previously cached 787 // entries will all be sorted. The entries we add we only sort on demand (we 788 // don't insert every element into its sorted position). We know that we 789 // won't get any reuse from currently inserted values, because we don't 790 // revisit blocks after we insert info for them. 791 unsigned NumSortedEntries = Cache->size(); 792 DEBUG(AssertSorted(*Cache)); 793 794 while (!Worklist.empty()) { 795 BasicBlock *BB = Worklist.pop_back_val(); 796 797 // Skip the first block if we have it. 798 if (!SkipFirstBlock) { 799 // Analyze the dependency of *Pointer in FromBB. See if we already have 800 // been here. 801 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 802 803 // Get the dependency info for Pointer in BB. If we have cached 804 // information, we will use it, otherwise we compute it. 805 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 806 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer.getAddr(), PointeeSize, 807 isLoad, BB, Cache, 808 NumSortedEntries); 809 810 // If we got a Def or Clobber, add this to the list of results. 811 if (!Dep.isNonLocal()) { 812 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 813 continue; 814 } 815 } 816 817 // If 'Pointer' is an instruction defined in this block, then we need to do 818 // phi translation to change it into a value live in the predecessor block. 819 // If not, we just add the predecessors to the worklist and scan them with 820 // the same Pointer. 821 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 822 SkipFirstBlock = false; 823 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 824 // Verify that we haven't looked at this block yet. 825 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 826 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 827 if (InsertRes.second) { 828 // First time we've looked at *PI. 829 Worklist.push_back(*PI); 830 continue; 831 } 832 833 // If we have seen this block before, but it was with a different 834 // pointer then we have a phi translation failure and we have to treat 835 // this as a clobber. 836 if (InsertRes.first->second != Pointer.getAddr()) 837 goto PredTranslationFailure; 838 } 839 continue; 840 } 841 842 // We do need to do phi translation, if we know ahead of time we can't phi 843 // translate this value, don't even try. 844 if (!Pointer.IsPotentiallyPHITranslatable()) 845 goto PredTranslationFailure; 846 847 // We may have added values to the cache list before this PHI translation. 848 // If so, we haven't done anything to ensure that the cache remains sorted. 849 // Sort it now (if needed) so that recursive invocations of 850 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 851 // value will only see properly sorted cache arrays. 852 if (Cache && NumSortedEntries != Cache->size()) { 853 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 854 NumSortedEntries = Cache->size(); 855 } 856 Cache = 0; 857 858 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 859 BasicBlock *Pred = *PI; 860 861 // Get the PHI translated pointer in this predecessor. This can fail if 862 // not translatable, in which case the getAddr() returns null. 863 PHITransAddr PredPointer(Pointer); 864 PredPointer.PHITranslateValue(BB, Pred); 865 866 Value *PredPtrVal = PredPointer.getAddr(); 867 868 // Check to see if we have already visited this pred block with another 869 // pointer. If so, we can't do this lookup. This failure can occur 870 // with PHI translation when a critical edge exists and the PHI node in 871 // the successor translates to a pointer value different than the 872 // pointer the block was first analyzed with. 873 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 874 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 875 876 if (!InsertRes.second) { 877 // If the predecessor was visited with PredPtr, then we already did 878 // the analysis and can ignore it. 879 if (InsertRes.first->second == PredPtrVal) 880 continue; 881 882 // Otherwise, the block was previously analyzed with a different 883 // pointer. We can't represent the result of this case, so we just 884 // treat this as a phi translation failure. 885 goto PredTranslationFailure; 886 } 887 888 // If PHI translation was unable to find an available pointer in this 889 // predecessor, then we have to assume that the pointer is clobbered in 890 // that predecessor. We can still do PRE of the load, which would insert 891 // a computation of the pointer in this predecessor. 892 if (PredPtrVal == 0) { 893 // Add the entry to the Result list. 894 NonLocalDepResult Entry(Pred, 895 MemDepResult::getClobber(Pred->getTerminator()), 896 PredPtrVal); 897 Result.push_back(Entry); 898 899 // Since we had a phi translation failure, the cache for CacheKey won't 900 // include all of the entries that we need to immediately satisfy future 901 // queries. Mark this in NonLocalPointerDeps by setting the 902 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 903 // cached value to do more work but not miss the phi trans failure. 904 NonLocalPointerDeps[CacheKey].first = BBSkipFirstBlockPair(); 905 continue; 906 } 907 908 // FIXME: it is entirely possible that PHI translating will end up with 909 // the same value. Consider PHI translating something like: 910 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 911 // to recurse here, pedantically speaking. 912 913 // If we have a problem phi translating, fall through to the code below 914 // to handle the failure condition. 915 if (getNonLocalPointerDepFromBB(PredPointer, PointeeSize, isLoad, Pred, 916 Result, Visited)) 917 goto PredTranslationFailure; 918 } 919 920 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 921 CacheInfo = &NonLocalPointerDeps[CacheKey]; 922 Cache = &CacheInfo->second; 923 NumSortedEntries = Cache->size(); 924 925 // Since we did phi translation, the "Cache" set won't contain all of the 926 // results for the query. This is ok (we can still use it to accelerate 927 // specific block queries) but we can't do the fastpath "return all 928 // results from the set" Clear out the indicator for this. 929 CacheInfo->first = BBSkipFirstBlockPair(); 930 SkipFirstBlock = false; 931 continue; 932 933 PredTranslationFailure: 934 935 if (Cache == 0) { 936 // Refresh the CacheInfo/Cache pointer if it got invalidated. 937 CacheInfo = &NonLocalPointerDeps[CacheKey]; 938 Cache = &CacheInfo->second; 939 NumSortedEntries = Cache->size(); 940 } 941 942 // Since we failed phi translation, the "Cache" set won't contain all of the 943 // results for the query. This is ok (we can still use it to accelerate 944 // specific block queries) but we can't do the fastpath "return all 945 // results from the set". Clear out the indicator for this. 946 CacheInfo->first = BBSkipFirstBlockPair(); 947 948 // If *nothing* works, mark the pointer as being clobbered by the first 949 // instruction in this block. 950 // 951 // If this is the magic first block, return this as a clobber of the whole 952 // incoming value. Since we can't phi translate to one of the predecessors, 953 // we have to bail out. 954 if (SkipFirstBlock) 955 return true; 956 957 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 958 assert(I != Cache->rend() && "Didn't find current block??"); 959 if (I->getBB() != BB) 960 continue; 961 962 assert(I->getResult().isNonLocal() && 963 "Should only be here with transparent block"); 964 I->setResult(MemDepResult::getClobber(BB->begin())); 965 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey); 966 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 967 Pointer.getAddr())); 968 break; 969 } 970 } 971 972 // Okay, we're done now. If we added new values to the cache, re-sort it. 973 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 974 DEBUG(AssertSorted(*Cache)); 975 return false; 976} 977 978/// RemoveCachedNonLocalPointerDependencies - If P exists in 979/// CachedNonLocalPointerInfo, remove it. 980void MemoryDependenceAnalysis:: 981RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 982 CachedNonLocalPointerInfo::iterator It = 983 NonLocalPointerDeps.find(P); 984 if (It == NonLocalPointerDeps.end()) return; 985 986 // Remove all of the entries in the BB->val map. This involves removing 987 // instructions from the reverse map. 988 NonLocalDepInfo &PInfo = It->second.second; 989 990 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 991 Instruction *Target = PInfo[i].getResult().getInst(); 992 if (Target == 0) continue; // Ignore non-local dep results. 993 assert(Target->getParent() == PInfo[i].getBB()); 994 995 // Eliminating the dirty entry from 'Cache', so update the reverse info. 996 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 997 } 998 999 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1000 NonLocalPointerDeps.erase(It); 1001} 1002 1003 1004/// invalidateCachedPointerInfo - This method is used to invalidate cached 1005/// information about the specified pointer, because it may be too 1006/// conservative in memdep. This is an optional call that can be used when 1007/// the client detects an equivalence between the pointer and some other 1008/// value and replaces the other value with ptr. This can make Ptr available 1009/// in more places that cached info does not necessarily keep. 1010void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1011 // If Ptr isn't really a pointer, just ignore it. 1012 if (!Ptr->getType()->isPointerTy()) return; 1013 // Flush store info for the pointer. 1014 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1015 // Flush load info for the pointer. 1016 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1017} 1018 1019/// removeInstruction - Remove an instruction from the dependence analysis, 1020/// updating the dependence of instructions that previously depended on it. 1021/// This method attempts to keep the cache coherent using the reverse map. 1022void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1023 // Walk through the Non-local dependencies, removing this one as the value 1024 // for any cached queries. 1025 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1026 if (NLDI != NonLocalDeps.end()) { 1027 NonLocalDepInfo &BlockMap = NLDI->second.first; 1028 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1029 DI != DE; ++DI) 1030 if (Instruction *Inst = DI->getResult().getInst()) 1031 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1032 NonLocalDeps.erase(NLDI); 1033 } 1034 1035 // If we have a cached local dependence query for this instruction, remove it. 1036 // 1037 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1038 if (LocalDepEntry != LocalDeps.end()) { 1039 // Remove us from DepInst's reverse set now that the local dep info is gone. 1040 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1041 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1042 1043 // Remove this local dependency info. 1044 LocalDeps.erase(LocalDepEntry); 1045 } 1046 1047 // If we have any cached pointer dependencies on this instruction, remove 1048 // them. If the instruction has non-pointer type, then it can't be a pointer 1049 // base. 1050 1051 // Remove it from both the load info and the store info. The instruction 1052 // can't be in either of these maps if it is non-pointer. 1053 if (RemInst->getType()->isPointerTy()) { 1054 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1055 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1056 } 1057 1058 // Loop over all of the things that depend on the instruction we're removing. 1059 // 1060 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1061 1062 // If we find RemInst as a clobber or Def in any of the maps for other values, 1063 // we need to replace its entry with a dirty version of the instruction after 1064 // it. If RemInst is a terminator, we use a null dirty value. 1065 // 1066 // Using a dirty version of the instruction after RemInst saves having to scan 1067 // the entire block to get to this point. 1068 MemDepResult NewDirtyVal; 1069 if (!RemInst->isTerminator()) 1070 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1071 1072 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1073 if (ReverseDepIt != ReverseLocalDeps.end()) { 1074 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1075 // RemInst can't be the terminator if it has local stuff depending on it. 1076 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1077 "Nothing can locally depend on a terminator"); 1078 1079 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1080 E = ReverseDeps.end(); I != E; ++I) { 1081 Instruction *InstDependingOnRemInst = *I; 1082 assert(InstDependingOnRemInst != RemInst && 1083 "Already removed our local dep info"); 1084 1085 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1086 1087 // Make sure to remember that new things depend on NewDepInst. 1088 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1089 "a local dep on this if it is a terminator!"); 1090 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1091 InstDependingOnRemInst)); 1092 } 1093 1094 ReverseLocalDeps.erase(ReverseDepIt); 1095 1096 // Add new reverse deps after scanning the set, to avoid invalidating the 1097 // 'ReverseDeps' reference. 1098 while (!ReverseDepsToAdd.empty()) { 1099 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1100 .insert(ReverseDepsToAdd.back().second); 1101 ReverseDepsToAdd.pop_back(); 1102 } 1103 } 1104 1105 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1106 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1107 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1108 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1109 I != E; ++I) { 1110 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1111 1112 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1113 // The information is now dirty! 1114 INLD.second = true; 1115 1116 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1117 DE = INLD.first.end(); DI != DE; ++DI) { 1118 if (DI->getResult().getInst() != RemInst) continue; 1119 1120 // Convert to a dirty entry for the subsequent instruction. 1121 DI->setResult(NewDirtyVal); 1122 1123 if (Instruction *NextI = NewDirtyVal.getInst()) 1124 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1125 } 1126 } 1127 1128 ReverseNonLocalDeps.erase(ReverseDepIt); 1129 1130 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1131 while (!ReverseDepsToAdd.empty()) { 1132 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1133 .insert(ReverseDepsToAdd.back().second); 1134 ReverseDepsToAdd.pop_back(); 1135 } 1136 } 1137 1138 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1139 // value in the NonLocalPointerDeps info. 1140 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1141 ReverseNonLocalPtrDeps.find(RemInst); 1142 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1143 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1144 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1145 1146 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1147 E = Set.end(); I != E; ++I) { 1148 ValueIsLoadPair P = *I; 1149 assert(P.getPointer() != RemInst && 1150 "Already removed NonLocalPointerDeps info for RemInst"); 1151 1152 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second; 1153 1154 // The cache is not valid for any specific block anymore. 1155 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair(); 1156 1157 // Update any entries for RemInst to use the instruction after it. 1158 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1159 DI != DE; ++DI) { 1160 if (DI->getResult().getInst() != RemInst) continue; 1161 1162 // Convert to a dirty entry for the subsequent instruction. 1163 DI->setResult(NewDirtyVal); 1164 1165 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1166 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1167 } 1168 1169 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1170 // subsequent value may invalidate the sortedness. 1171 std::sort(NLPDI.begin(), NLPDI.end()); 1172 } 1173 1174 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1175 1176 while (!ReversePtrDepsToAdd.empty()) { 1177 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1178 .insert(ReversePtrDepsToAdd.back().second); 1179 ReversePtrDepsToAdd.pop_back(); 1180 } 1181 } 1182 1183 1184 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1185 AA->deleteValue(RemInst); 1186 DEBUG(verifyRemoved(RemInst)); 1187} 1188/// verifyRemoved - Verify that the specified instruction does not occur 1189/// in our internal data structures. 1190void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1191 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1192 E = LocalDeps.end(); I != E; ++I) { 1193 assert(I->first != D && "Inst occurs in data structures"); 1194 assert(I->second.getInst() != D && 1195 "Inst occurs in data structures"); 1196 } 1197 1198 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1199 E = NonLocalPointerDeps.end(); I != E; ++I) { 1200 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1201 const NonLocalDepInfo &Val = I->second.second; 1202 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1203 II != E; ++II) 1204 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1205 } 1206 1207 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1208 E = NonLocalDeps.end(); I != E; ++I) { 1209 assert(I->first != D && "Inst occurs in data structures"); 1210 const PerInstNLInfo &INLD = I->second; 1211 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1212 EE = INLD.first.end(); II != EE; ++II) 1213 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1214 } 1215 1216 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1217 E = ReverseLocalDeps.end(); I != E; ++I) { 1218 assert(I->first != D && "Inst occurs in data structures"); 1219 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1220 EE = I->second.end(); II != EE; ++II) 1221 assert(*II != D && "Inst occurs in data structures"); 1222 } 1223 1224 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1225 E = ReverseNonLocalDeps.end(); 1226 I != E; ++I) { 1227 assert(I->first != D && "Inst occurs in data structures"); 1228 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1229 EE = I->second.end(); II != EE; ++II) 1230 assert(*II != D && "Inst occurs in data structures"); 1231 } 1232 1233 for (ReverseNonLocalPtrDepTy::const_iterator 1234 I = ReverseNonLocalPtrDeps.begin(), 1235 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1236 assert(I->first != D && "Inst occurs in rev NLPD map"); 1237 1238 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1239 E = I->second.end(); II != E; ++II) 1240 assert(*II != ValueIsLoadPair(D, false) && 1241 *II != ValueIsLoadPair(D, true) && 1242 "Inst occurs in ReverseNonLocalPtrDeps map"); 1243 } 1244 1245} 1246