MemoryDependenceAnalysis.cpp revision dd696052f0e4ecc973d105be19cf1b4b72f9a0c4
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/ADT/Statistic.h" 27#include "llvm/ADT/STLExtras.h" 28#include "llvm/Support/PredIteratorCache.h" 29#include "llvm/Support/Debug.h" 30using namespace llvm; 31 32STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 33STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 34STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 35 36STATISTIC(NumCacheNonLocalPtr, 37 "Number of fully cached non-local ptr responses"); 38STATISTIC(NumCacheDirtyNonLocalPtr, 39 "Number of cached, but dirty, non-local ptr responses"); 40STATISTIC(NumUncacheNonLocalPtr, 41 "Number of uncached non-local ptr responses"); 42STATISTIC(NumCacheCompleteNonLocalPtr, 43 "Number of block queries that were completely cached"); 44 45char MemoryDependenceAnalysis::ID = 0; 46 47// Register this pass... 48static RegisterPass<MemoryDependenceAnalysis> X("memdep", 49 "Memory Dependence Analysis", false, true); 50 51MemoryDependenceAnalysis::MemoryDependenceAnalysis() 52: FunctionPass(&ID), PredCache(0) { 53} 54MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 55} 56 57/// Clean up memory in between runs 58void MemoryDependenceAnalysis::releaseMemory() { 59 LocalDeps.clear(); 60 NonLocalDeps.clear(); 61 NonLocalPointerDeps.clear(); 62 ReverseLocalDeps.clear(); 63 ReverseNonLocalDeps.clear(); 64 ReverseNonLocalPtrDeps.clear(); 65 PredCache->clear(); 66} 67 68 69 70/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 71/// 72void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 73 AU.setPreservesAll(); 74 AU.addRequiredTransitive<AliasAnalysis>(); 75} 76 77bool MemoryDependenceAnalysis::runOnFunction(Function &) { 78 AA = &getAnalysis<AliasAnalysis>(); 79 if (PredCache == 0) 80 PredCache.reset(new PredIteratorCache()); 81 return false; 82} 83 84/// RemoveFromReverseMap - This is a helper function that removes Val from 85/// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry. 86template <typename KeyTy> 87static void RemoveFromReverseMap(DenseMap<Instruction*, 88 SmallPtrSet<KeyTy, 4> > &ReverseMap, 89 Instruction *Inst, KeyTy Val) { 90 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator 91 InstIt = ReverseMap.find(Inst); 92 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); 93 bool Found = InstIt->second.erase(Val); 94 assert(Found && "Invalid reverse map!"); Found=Found; 95 if (InstIt->second.empty()) 96 ReverseMap.erase(InstIt); 97} 98 99 100/// getCallSiteDependencyFrom - Private helper for finding the local 101/// dependencies of a call site. 102MemDepResult MemoryDependenceAnalysis:: 103getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 104 BasicBlock::iterator ScanIt, BasicBlock *BB) { 105 // Walk backwards through the block, looking for dependencies 106 while (ScanIt != BB->begin()) { 107 Instruction *Inst = --ScanIt; 108 109 // If this inst is a memory op, get the pointer it accessed 110 Value *Pointer = 0; 111 uint64_t PointerSize = 0; 112 if (StoreInst *S = dyn_cast<StoreInst>(Inst)) { 113 Pointer = S->getPointerOperand(); 114 PointerSize = AA->getTypeStoreSize(S->getOperand(0)->getType()); 115 } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 116 Pointer = V->getOperand(0); 117 PointerSize = AA->getTypeStoreSize(V->getType()); 118 } else if (isFreeCall(Inst)) { 119 Pointer = Inst->getOperand(1); 120 // calls to free() erase the entire structure 121 PointerSize = ~0ULL; 122 } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) { 123 // Debug intrinsics don't cause dependences. 124 if (isa<DbgInfoIntrinsic>(Inst)) continue; 125 CallSite InstCS = CallSite::get(Inst); 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 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 187 // If we pass an invariant-end marker, then we've just entered an 188 // invariant region and can start ignoring dependencies. 189 if (II->getIntrinsicID() == Intrinsic::invariant_end) { 190 uint64_t invariantSize = ~0ULL; 191 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getOperand(2))) 192 invariantSize = CI->getZExtValue(); 193 194 AliasAnalysis::AliasResult R = 195 AA->alias(II->getOperand(3), invariantSize, MemPtr, MemSize); 196 if (R == AliasAnalysis::MustAlias) { 197 invariantTag = II->getOperand(1); 198 continue; 199 } 200 201 // If we reach a lifetime begin or end marker, then the query ends here 202 // because the value is undefined. 203 } else if (II->getIntrinsicID() == Intrinsic::lifetime_start || 204 II->getIntrinsicID() == Intrinsic::lifetime_end) { 205 uint64_t invariantSize = ~0ULL; 206 if (ConstantInt *CI = dyn_cast<ConstantInt>(II->getOperand(1))) 207 invariantSize = CI->getZExtValue(); 208 209 AliasAnalysis::AliasResult R = 210 AA->alias(II->getOperand(2), invariantSize, MemPtr, MemSize); 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 // Debug intrinsics don't cause dependences. 221 if (isa<DbgInfoIntrinsic>(Inst)) 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) || extractMallocCall(Inst)) { 279 Value *AccessPtr = MemPtr->getUnderlyingObject(); 280 281 if (AccessPtr == Inst || 282 AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) 283 return MemDepResult::getDef(Inst); 284 continue; 285 } 286 287 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 288 switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) { 289 case AliasAnalysis::NoModRef: 290 // If the call has no effect on the queried pointer, just ignore it. 291 continue; 292 case AliasAnalysis::Mod: 293 // If we're in an invariant region, we can ignore calls that ONLY 294 // modify the pointer. 295 if (invariantTag) continue; 296 return MemDepResult::getClobber(Inst); 297 case AliasAnalysis::Ref: 298 // If the call is known to never store to the pointer, and if this is a 299 // load query, we can safely ignore it (scan past it). 300 if (isLoad) 301 continue; 302 default: 303 // Otherwise, there is a potential dependence. Return a clobber. 304 return MemDepResult::getClobber(Inst); 305 } 306 } 307 308 // No dependence found. If this is the entry block of the function, it is a 309 // clobber, otherwise it is non-local. 310 if (BB != &BB->getParent()->getEntryBlock()) 311 return MemDepResult::getNonLocal(); 312 return MemDepResult::getClobber(ScanIt); 313} 314 315/// getDependency - Return the instruction on which a memory operation 316/// depends. 317MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 318 Instruction *ScanPos = QueryInst; 319 320 // Check for a cached result 321 MemDepResult &LocalCache = LocalDeps[QueryInst]; 322 323 // If the cached entry is non-dirty, just return it. Note that this depends 324 // on MemDepResult's default constructing to 'dirty'. 325 if (!LocalCache.isDirty()) 326 return LocalCache; 327 328 // Otherwise, if we have a dirty entry, we know we can start the scan at that 329 // instruction, which may save us some work. 330 if (Instruction *Inst = LocalCache.getInst()) { 331 ScanPos = Inst; 332 333 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 334 } 335 336 BasicBlock *QueryParent = QueryInst->getParent(); 337 338 Value *MemPtr = 0; 339 uint64_t MemSize = 0; 340 341 // Do the scan. 342 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 343 // No dependence found. If this is the entry block of the function, it is a 344 // clobber, otherwise it is non-local. 345 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 346 LocalCache = MemDepResult::getNonLocal(); 347 else 348 LocalCache = MemDepResult::getClobber(QueryInst); 349 } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) { 350 // If this is a volatile store, don't mess around with it. Just return the 351 // previous instruction as a clobber. 352 if (SI->isVolatile()) 353 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 354 else { 355 MemPtr = SI->getPointerOperand(); 356 MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); 357 } 358 } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) { 359 // If this is a volatile load, don't mess around with it. Just return the 360 // previous instruction as a clobber. 361 if (LI->isVolatile()) 362 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 363 else { 364 MemPtr = LI->getPointerOperand(); 365 MemSize = AA->getTypeStoreSize(LI->getType()); 366 } 367 } else if (isFreeCall(QueryInst)) { 368 MemPtr = QueryInst->getOperand(1); 369 // calls to free() erase the entire structure, not just a field. 370 MemSize = ~0UL; 371 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 372 CallSite QueryCS = CallSite::get(QueryInst); 373 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 374 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 375 QueryParent); 376 } else { 377 // Non-memory instruction. 378 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 379 } 380 381 // If we need to do a pointer scan, make it happen. 382 if (MemPtr) 383 LocalCache = getPointerDependencyFrom(MemPtr, MemSize, 384 isa<LoadInst>(QueryInst), 385 ScanPos, QueryParent); 386 387 // Remember the result! 388 if (Instruction *I = LocalCache.getInst()) 389 ReverseLocalDeps[I].insert(QueryInst); 390 391 return LocalCache; 392} 393 394#ifndef NDEBUG 395/// AssertSorted - This method is used when -debug is specified to verify that 396/// cache arrays are properly kept sorted. 397static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 398 int Count = -1) { 399 if (Count == -1) Count = Cache.size(); 400 if (Count == 0) return; 401 402 for (unsigned i = 1; i != unsigned(Count); ++i) 403 assert(Cache[i-1] <= Cache[i] && "Cache isn't sorted!"); 404} 405#endif 406 407/// getNonLocalCallDependency - Perform a full dependency query for the 408/// specified call, returning the set of blocks that the value is 409/// potentially live across. The returned set of results will include a 410/// "NonLocal" result for all blocks where the value is live across. 411/// 412/// This method assumes the instruction returns a "NonLocal" dependency 413/// within its own block. 414/// 415/// This returns a reference to an internal data structure that may be 416/// invalidated on the next non-local query or when an instruction is 417/// removed. Clients must copy this data if they want it around longer than 418/// that. 419const MemoryDependenceAnalysis::NonLocalDepInfo & 420MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 421 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 422 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 423 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 424 NonLocalDepInfo &Cache = CacheP.first; 425 426 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 427 /// the cached case, this can happen due to instructions being deleted etc. In 428 /// the uncached case, this starts out as the set of predecessors we care 429 /// about. 430 SmallVector<BasicBlock*, 32> DirtyBlocks; 431 432 if (!Cache.empty()) { 433 // Okay, we have a cache entry. If we know it is not dirty, just return it 434 // with no computation. 435 if (!CacheP.second) { 436 NumCacheNonLocal++; 437 return Cache; 438 } 439 440 // If we already have a partially computed set of results, scan them to 441 // determine what is dirty, seeding our initial DirtyBlocks worklist. 442 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 443 I != E; ++I) 444 if (I->second.isDirty()) 445 DirtyBlocks.push_back(I->first); 446 447 // Sort the cache so that we can do fast binary search lookups below. 448 std::sort(Cache.begin(), Cache.end()); 449 450 ++NumCacheDirtyNonLocal; 451 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 452 // << Cache.size() << " cached: " << *QueryInst; 453 } else { 454 // Seed DirtyBlocks with each of the preds of QueryInst's block. 455 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 456 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 457 DirtyBlocks.push_back(*PI); 458 NumUncacheNonLocal++; 459 } 460 461 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 462 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 463 464 SmallPtrSet<BasicBlock*, 64> Visited; 465 466 unsigned NumSortedEntries = Cache.size(); 467 DEBUG(AssertSorted(Cache)); 468 469 // Iterate while we still have blocks to update. 470 while (!DirtyBlocks.empty()) { 471 BasicBlock *DirtyBB = DirtyBlocks.back(); 472 DirtyBlocks.pop_back(); 473 474 // Already processed this block? 475 if (!Visited.insert(DirtyBB)) 476 continue; 477 478 // Do a binary search to see if we already have an entry for this block in 479 // the cache set. If so, find it. 480 DEBUG(AssertSorted(Cache, NumSortedEntries)); 481 NonLocalDepInfo::iterator Entry = 482 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 483 std::make_pair(DirtyBB, MemDepResult())); 484 if (Entry != Cache.begin() && prior(Entry)->first == DirtyBB) 485 --Entry; 486 487 MemDepResult *ExistingResult = 0; 488 if (Entry != Cache.begin()+NumSortedEntries && 489 Entry->first == DirtyBB) { 490 // If we already have an entry, and if it isn't already dirty, the block 491 // is done. 492 if (!Entry->second.isDirty()) 493 continue; 494 495 // Otherwise, remember this slot so we can update the value. 496 ExistingResult = &Entry->second; 497 } 498 499 // If the dirty entry has a pointer, start scanning from it so we don't have 500 // to rescan the entire block. 501 BasicBlock::iterator ScanPos = DirtyBB->end(); 502 if (ExistingResult) { 503 if (Instruction *Inst = ExistingResult->getInst()) { 504 ScanPos = Inst; 505 // We're removing QueryInst's use of Inst. 506 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 507 QueryCS.getInstruction()); 508 } 509 } 510 511 // Find out if this block has a local dependency for QueryInst. 512 MemDepResult Dep; 513 514 if (ScanPos != DirtyBB->begin()) { 515 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 516 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 517 // No dependence found. If this is the entry block of the function, it is 518 // a clobber, otherwise it is non-local. 519 Dep = MemDepResult::getNonLocal(); 520 } else { 521 Dep = MemDepResult::getClobber(ScanPos); 522 } 523 524 // If we had a dirty entry for the block, update it. Otherwise, just add 525 // a new entry. 526 if (ExistingResult) 527 *ExistingResult = Dep; 528 else 529 Cache.push_back(std::make_pair(DirtyBB, Dep)); 530 531 // If the block has a dependency (i.e. it isn't completely transparent to 532 // the value), remember the association! 533 if (!Dep.isNonLocal()) { 534 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 535 // update this when we remove instructions. 536 if (Instruction *Inst = Dep.getInst()) 537 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 538 } else { 539 540 // If the block *is* completely transparent to the load, we need to check 541 // the predecessors of this block. Add them to our worklist. 542 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 543 DirtyBlocks.push_back(*PI); 544 } 545 } 546 547 return Cache; 548} 549 550/// getNonLocalPointerDependency - Perform a full dependency query for an 551/// access to the specified (non-volatile) memory location, returning the 552/// set of instructions that either define or clobber the value. 553/// 554/// This method assumes the pointer has a "NonLocal" dependency within its 555/// own block. 556/// 557void MemoryDependenceAnalysis:: 558getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB, 559 SmallVectorImpl<NonLocalDepEntry> &Result) { 560 assert(isa<PointerType>(Pointer->getType()) && 561 "Can't get pointer deps of a non-pointer!"); 562 Result.clear(); 563 564 // We know that the pointer value is live into FromBB find the def/clobbers 565 // from presecessors. 566 const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType(); 567 uint64_t PointeeSize = AA->getTypeStoreSize(EltTy); 568 569 // This is the set of blocks we've inspected, and the pointer we consider in 570 // each block. Because of critical edges, we currently bail out if querying 571 // a block with multiple different pointers. This can happen during PHI 572 // translation. 573 DenseMap<BasicBlock*, Value*> Visited; 574 if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB, 575 Result, Visited, true)) 576 return; 577 Result.clear(); 578 Result.push_back(std::make_pair(FromBB, 579 MemDepResult::getClobber(FromBB->begin()))); 580} 581 582/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 583/// Pointer/PointeeSize using either cached information in Cache or by doing a 584/// lookup (which may use dirty cache info if available). If we do a lookup, 585/// add the result to the cache. 586MemDepResult MemoryDependenceAnalysis:: 587GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, 588 bool isLoad, BasicBlock *BB, 589 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 590 591 // Do a binary search to see if we already have an entry for this block in 592 // the cache set. If so, find it. 593 NonLocalDepInfo::iterator Entry = 594 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 595 std::make_pair(BB, MemDepResult())); 596 if (Entry != Cache->begin() && prior(Entry)->first == BB) 597 --Entry; 598 599 MemDepResult *ExistingResult = 0; 600 if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB) 601 ExistingResult = &Entry->second; 602 603 // If we have a cached entry, and it is non-dirty, use it as the value for 604 // this dependency. 605 if (ExistingResult && !ExistingResult->isDirty()) { 606 ++NumCacheNonLocalPtr; 607 return *ExistingResult; 608 } 609 610 // Otherwise, we have to scan for the value. If we have a dirty cache 611 // entry, start scanning from its position, otherwise we scan from the end 612 // of the block. 613 BasicBlock::iterator ScanPos = BB->end(); 614 if (ExistingResult && ExistingResult->getInst()) { 615 assert(ExistingResult->getInst()->getParent() == BB && 616 "Instruction invalidated?"); 617 ++NumCacheDirtyNonLocalPtr; 618 ScanPos = ExistingResult->getInst(); 619 620 // Eliminating the dirty entry from 'Cache', so update the reverse info. 621 ValueIsLoadPair CacheKey(Pointer, isLoad); 622 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 623 } else { 624 ++NumUncacheNonLocalPtr; 625 } 626 627 // Scan the block for the dependency. 628 MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, 629 ScanPos, BB); 630 631 // If we had a dirty entry for the block, update it. Otherwise, just add 632 // a new entry. 633 if (ExistingResult) 634 *ExistingResult = Dep; 635 else 636 Cache->push_back(std::make_pair(BB, Dep)); 637 638 // If the block has a dependency (i.e. it isn't completely transparent to 639 // the value), remember the reverse association because we just added it 640 // to Cache! 641 if (Dep.isNonLocal()) 642 return Dep; 643 644 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 645 // update MemDep when we remove instructions. 646 Instruction *Inst = Dep.getInst(); 647 assert(Inst && "Didn't depend on anything?"); 648 ValueIsLoadPair CacheKey(Pointer, isLoad); 649 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 650 return Dep; 651} 652 653/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 654/// number of elements in the array that are already properly ordered. This is 655/// optimized for the case when only a few entries are added. 656static void 657SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 658 unsigned NumSortedEntries) { 659 switch (Cache.size() - NumSortedEntries) { 660 case 0: 661 // done, no new entries. 662 break; 663 case 2: { 664 // Two new entries, insert the last one into place. 665 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back(); 666 Cache.pop_back(); 667 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 668 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 669 Cache.insert(Entry, Val); 670 // FALL THROUGH. 671 } 672 case 1: 673 // One new entry, Just insert the new value at the appropriate position. 674 if (Cache.size() != 1) { 675 MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back(); 676 Cache.pop_back(); 677 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 678 std::upper_bound(Cache.begin(), Cache.end(), Val); 679 Cache.insert(Entry, Val); 680 } 681 break; 682 default: 683 // Added many values, do a full scale sort. 684 std::sort(Cache.begin(), Cache.end()); 685 break; 686 } 687} 688 689/// isPHITranslatable - Return true if the specified computation is derived from 690/// a PHI node in the current block and if it is simple enough for us to handle. 691static bool isPHITranslatable(Instruction *Inst) { 692 if (isa<PHINode>(Inst)) 693 return true; 694 695 // We can handle bitcast of a PHI, but the PHI needs to be in the same block 696 // as the bitcast. 697 if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) { 698 Instruction *OpI = dyn_cast<Instruction>(BC->getOperand(0)); 699 if (OpI == 0 || OpI->getParent() != Inst->getParent()) 700 return true; 701 return isPHITranslatable(OpI); 702 } 703 704 // We can translate a GEP if all of its operands defined in this block are phi 705 // translatable. 706 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) { 707 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) { 708 Instruction *OpI = dyn_cast<Instruction>(GEP->getOperand(i)); 709 if (OpI == 0 || OpI->getParent() != Inst->getParent()) 710 continue; 711 712 if (!isPHITranslatable(OpI)) 713 return false; 714 } 715 return true; 716 } 717 718 if (Inst->getOpcode() == Instruction::Add && 719 isa<ConstantInt>(Inst->getOperand(1))) { 720 Instruction *OpI = dyn_cast<Instruction>(Inst->getOperand(0)); 721 if (OpI == 0 || OpI->getParent() != Inst->getParent()) 722 return true; 723 return isPHITranslatable(OpI); 724 } 725 726 // cerr << "MEMDEP: Could not PHI translate: " << *Pointer; 727 // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst)) 728 // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0); 729 730 return false; 731} 732 733/// GetPHITranslatedValue - Given a computation that satisfied the 734/// isPHITranslatable predicate, see if we can translate the computation into 735/// the specified predecessor block. If so, return that value. 736Value *MemoryDependenceAnalysis:: 737GetPHITranslatedValue(Value *InVal, BasicBlock *CurBB, BasicBlock *Pred, 738 const TargetData *TD) const { 739 // If the input value is not an instruction, or if it is not defined in CurBB, 740 // then we don't need to phi translate it. 741 Instruction *Inst = dyn_cast<Instruction>(InVal); 742 if (Inst == 0 || Inst->getParent() != CurBB) 743 return InVal; 744 745 if (PHINode *PN = dyn_cast<PHINode>(Inst)) 746 return PN->getIncomingValueForBlock(Pred); 747 748 // Handle bitcast of PHI. 749 if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) { 750 // PHI translate the input operand. 751 Value *PHIIn = GetPHITranslatedValue(BC->getOperand(0), CurBB, Pred, TD); 752 if (PHIIn == 0) return 0; 753 754 // Constants are trivial to phi translate. 755 if (Constant *C = dyn_cast<Constant>(PHIIn)) 756 return ConstantExpr::getBitCast(C, BC->getType()); 757 758 // Otherwise we have to see if a bitcasted version of the incoming pointer 759 // is available. If so, we can use it, otherwise we have to fail. 760 for (Value::use_iterator UI = PHIIn->use_begin(), E = PHIIn->use_end(); 761 UI != E; ++UI) { 762 if (BitCastInst *BCI = dyn_cast<BitCastInst>(*UI)) 763 if (BCI->getType() == BC->getType()) 764 return BCI; 765 } 766 return 0; 767 } 768 769 // Handle getelementptr with at least one PHI translatable operand. 770 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) { 771 SmallVector<Value*, 8> GEPOps; 772 BasicBlock *CurBB = GEP->getParent(); 773 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) { 774 Value *GEPOp = GEP->getOperand(i); 775 // No PHI translation is needed of operands whose values are live in to 776 // the predecessor block. 777 if (!isa<Instruction>(GEPOp) || 778 cast<Instruction>(GEPOp)->getParent() != CurBB) { 779 GEPOps.push_back(GEPOp); 780 continue; 781 } 782 783 // If the operand is a phi node, do phi translation. 784 Value *InOp = GetPHITranslatedValue(GEPOp, CurBB, Pred, TD); 785 if (InOp == 0) return 0; 786 787 GEPOps.push_back(InOp); 788 } 789 790 // Simplify the GEP to handle 'gep x, 0' -> x etc. 791 if (Value *V = SimplifyGEPInst(&GEPOps[0], GEPOps.size(), TD)) 792 return V; 793 794 // Scan to see if we have this GEP available. 795 Value *APHIOp = GEPOps[0]; 796 for (Value::use_iterator UI = APHIOp->use_begin(), E = APHIOp->use_end(); 797 UI != E; ++UI) { 798 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(*UI)) 799 if (GEPI->getType() == GEP->getType() && 800 GEPI->getNumOperands() == GEPOps.size() && 801 GEPI->getParent()->getParent() == CurBB->getParent()) { 802 bool Mismatch = false; 803 for (unsigned i = 0, e = GEPOps.size(); i != e; ++i) 804 if (GEPI->getOperand(i) != GEPOps[i]) { 805 Mismatch = true; 806 break; 807 } 808 if (!Mismatch) 809 return GEPI; 810 } 811 } 812 return 0; 813 } 814 815 // Handle add with a constant RHS. 816 if (Inst->getOpcode() == Instruction::Add && 817 isa<ConstantInt>(Inst->getOperand(1))) { 818 // PHI translate the LHS. 819 Value *LHS; 820 Constant *RHS = cast<ConstantInt>(Inst->getOperand(1)); 821 Instruction *OpI = dyn_cast<Instruction>(Inst->getOperand(0)); 822 bool isNSW = cast<BinaryOperator>(Inst)->hasNoSignedWrap(); 823 bool isNUW = cast<BinaryOperator>(Inst)->hasNoUnsignedWrap(); 824 825 if (OpI == 0 || OpI->getParent() != Inst->getParent()) 826 LHS = Inst->getOperand(0); 827 else { 828 LHS = GetPHITranslatedValue(Inst->getOperand(0), CurBB, Pred, TD); 829 if (LHS == 0) 830 return 0; 831 } 832 833 // If the PHI translated LHS is an add of a constant, fold the immediates. 834 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(LHS)) 835 if (BOp->getOpcode() == Instruction::Add) 836 if (ConstantInt *CI = dyn_cast<ConstantInt>(BOp->getOperand(1))) { 837 LHS = BOp->getOperand(0); 838 RHS = ConstantExpr::getAdd(RHS, CI); 839 isNSW = isNUW = false; 840 } 841 842 // See if the add simplifies away. 843 if (Value *Res = SimplifyAddInst(LHS, RHS, isNSW, isNUW, TD)) 844 return Res; 845 846 // Otherwise, see if we have this add available somewhere. 847 for (Value::use_iterator UI = LHS->use_begin(), E = LHS->use_end(); 848 UI != E; ++UI) { 849 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(*UI)) 850 if (BO->getOperand(0) == LHS && BO->getOperand(1) == RHS && 851 BO->getParent()->getParent() == CurBB->getParent()) 852 return BO; 853 } 854 855 return 0; 856 } 857 858 return 0; 859} 860 861/// GetAvailablePHITranslatePointer - Return the value computed by 862/// PHITranslatePointer if it dominates PredBB, otherwise return null. 863Value *MemoryDependenceAnalysis:: 864GetAvailablePHITranslatedValue(Value *V, 865 BasicBlock *CurBB, BasicBlock *PredBB, 866 const TargetData *TD, 867 const DominatorTree &DT) const { 868 // See if PHI translation succeeds. 869 V = GetPHITranslatedValue(V, CurBB, PredBB, TD); 870 if (V == 0) return 0; 871 872 // Make sure the value is live in the predecessor. 873 if (Instruction *Inst = dyn_cast_or_null<Instruction>(V)) 874 if (!DT.dominates(Inst->getParent(), PredBB)) 875 return 0; 876 return V; 877} 878 879 880/// InsertPHITranslatedPointer - Insert a computation of the PHI translated 881/// version of 'V' for the edge PredBB->CurBB into the end of the PredBB 882/// block. All newly created instructions are added to the NewInsts list. 883/// 884Value *MemoryDependenceAnalysis:: 885InsertPHITranslatedPointer(Value *InVal, BasicBlock *CurBB, 886 BasicBlock *PredBB, const TargetData *TD, 887 const DominatorTree &DT, 888 SmallVectorImpl<Instruction*> &NewInsts) const { 889 // See if we have a version of this value already available and dominating 890 // PredBB. If so, there is no need to insert a new copy. 891 if (Value *Res = GetAvailablePHITranslatedValue(InVal, CurBB, PredBB, TD, DT)) 892 return Res; 893 894 // If we don't have an available version of this value, it must be an 895 // instruction. 896 Instruction *Inst = cast<Instruction>(InVal); 897 898 // Handle bitcast of PHI translatable value. 899 if (BitCastInst *BC = dyn_cast<BitCastInst>(Inst)) { 900 Value *OpVal = InsertPHITranslatedPointer(BC->getOperand(0), 901 CurBB, PredBB, TD, DT, NewInsts); 902 if (OpVal == 0) return 0; 903 904 // Otherwise insert a bitcast at the end of PredBB. 905 BitCastInst *New = new BitCastInst(OpVal, InVal->getType(), 906 InVal->getName()+".phi.trans.insert", 907 PredBB->getTerminator()); 908 NewInsts.push_back(New); 909 return New; 910 } 911 912 // Handle getelementptr with at least one PHI operand. 913 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Inst)) { 914 SmallVector<Value*, 8> GEPOps; 915 BasicBlock *CurBB = GEP->getParent(); 916 for (unsigned i = 0, e = GEP->getNumOperands(); i != e; ++i) { 917 Value *OpVal = InsertPHITranslatedPointer(GEP->getOperand(i), 918 CurBB, PredBB, TD, DT, NewInsts); 919 if (OpVal == 0) return 0; 920 GEPOps.push_back(OpVal); 921 } 922 923 GetElementPtrInst *Result = 924 GetElementPtrInst::Create(GEPOps[0], GEPOps.begin()+1, GEPOps.end(), 925 InVal->getName()+".phi.trans.insert", 926 PredBB->getTerminator()); 927 Result->setIsInBounds(GEP->isInBounds()); 928 NewInsts.push_back(Result); 929 return Result; 930 } 931 932#if 0 933 // FIXME: This code works, but it is unclear that we actually want to insert 934 // a big chain of computation in order to make a value available in a block. 935 // This needs to be evaluated carefully to consider its cost trade offs. 936 937 // Handle add with a constant RHS. 938 if (Inst->getOpcode() == Instruction::Add && 939 isa<ConstantInt>(Inst->getOperand(1))) { 940 // PHI translate the LHS. 941 Value *OpVal = InsertPHITranslatedPointer(Inst->getOperand(0), 942 CurBB, PredBB, TD, DT, NewInsts); 943 if (OpVal == 0) return 0; 944 945 BinaryOperator *Res = BinaryOperator::CreateAdd(OpVal, Inst->getOperand(1), 946 InVal->getName()+".phi.trans.insert", 947 PredBB->getTerminator()); 948 Res->setHasNoSignedWrap(cast<BinaryOperator>(Inst)->hasNoSignedWrap()); 949 Res->setHasNoUnsignedWrap(cast<BinaryOperator>(Inst)->hasNoUnsignedWrap()); 950 NewInsts.push_back(Res); 951 return Res; 952 } 953#endif 954 955 return 0; 956} 957 958/// getNonLocalPointerDepFromBB - Perform a dependency query based on 959/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 960/// results to the results vector and keep track of which blocks are visited in 961/// 'Visited'. 962/// 963/// This has special behavior for the first block queries (when SkipFirstBlock 964/// is true). In this special case, it ignores the contents of the specified 965/// block and starts returning dependence info for its predecessors. 966/// 967/// This function returns false on success, or true to indicate that it could 968/// not compute dependence information for some reason. This should be treated 969/// as a clobber dependence on the first instruction in the predecessor block. 970bool MemoryDependenceAnalysis:: 971getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize, 972 bool isLoad, BasicBlock *StartBB, 973 SmallVectorImpl<NonLocalDepEntry> &Result, 974 DenseMap<BasicBlock*, Value*> &Visited, 975 bool SkipFirstBlock) { 976 977 // Look up the cached info for Pointer. 978 ValueIsLoadPair CacheKey(Pointer, isLoad); 979 980 std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo = 981 &NonLocalPointerDeps[CacheKey]; 982 NonLocalDepInfo *Cache = &CacheInfo->second; 983 984 // If we have valid cached information for exactly the block we are 985 // investigating, just return it with no recomputation. 986 if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 987 // We have a fully cached result for this query then we can just return the 988 // cached results and populate the visited set. However, we have to verify 989 // that we don't already have conflicting results for these blocks. Check 990 // to ensure that if a block in the results set is in the visited set that 991 // it was for the same pointer query. 992 if (!Visited.empty()) { 993 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 994 I != E; ++I) { 995 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->first); 996 if (VI == Visited.end() || VI->second == Pointer) continue; 997 998 // We have a pointer mismatch in a block. Just return clobber, saying 999 // that something was clobbered in this result. We could also do a 1000 // non-fully cached query, but there is little point in doing this. 1001 return true; 1002 } 1003 } 1004 1005 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 1006 I != E; ++I) { 1007 Visited.insert(std::make_pair(I->first, Pointer)); 1008 if (!I->second.isNonLocal()) 1009 Result.push_back(*I); 1010 } 1011 ++NumCacheCompleteNonLocalPtr; 1012 return false; 1013 } 1014 1015 // Otherwise, either this is a new block, a block with an invalid cache 1016 // pointer or one that we're about to invalidate by putting more info into it 1017 // than its valid cache info. If empty, the result will be valid cache info, 1018 // otherwise it isn't. 1019 if (Cache->empty()) 1020 CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 1021 else 1022 CacheInfo->first = BBSkipFirstBlockPair(); 1023 1024 SmallVector<BasicBlock*, 32> Worklist; 1025 Worklist.push_back(StartBB); 1026 1027 // Keep track of the entries that we know are sorted. Previously cached 1028 // entries will all be sorted. The entries we add we only sort on demand (we 1029 // don't insert every element into its sorted position). We know that we 1030 // won't get any reuse from currently inserted values, because we don't 1031 // revisit blocks after we insert info for them. 1032 unsigned NumSortedEntries = Cache->size(); 1033 DEBUG(AssertSorted(*Cache)); 1034 1035 while (!Worklist.empty()) { 1036 BasicBlock *BB = Worklist.pop_back_val(); 1037 1038 // Skip the first block if we have it. 1039 if (!SkipFirstBlock) { 1040 // Analyze the dependency of *Pointer in FromBB. See if we already have 1041 // been here. 1042 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 1043 1044 // Get the dependency info for Pointer in BB. If we have cached 1045 // information, we will use it, otherwise we compute it. 1046 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 1047 MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad, 1048 BB, Cache, NumSortedEntries); 1049 1050 // If we got a Def or Clobber, add this to the list of results. 1051 if (!Dep.isNonLocal()) { 1052 Result.push_back(NonLocalDepEntry(BB, Dep)); 1053 continue; 1054 } 1055 } 1056 1057 // If 'Pointer' is an instruction defined in this block, then we need to do 1058 // phi translation to change it into a value live in the predecessor block. 1059 // If phi translation fails, then we can't continue dependence analysis. 1060 Instruction *PtrInst = dyn_cast<Instruction>(Pointer); 1061 bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB; 1062 1063 // If no PHI translation is needed, just add all the predecessors of this 1064 // block to scan them as well. 1065 if (!NeedsPHITranslation) { 1066 SkipFirstBlock = false; 1067 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1068 // Verify that we haven't looked at this block yet. 1069 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1070 InsertRes = Visited.insert(std::make_pair(*PI, Pointer)); 1071 if (InsertRes.second) { 1072 // First time we've looked at *PI. 1073 Worklist.push_back(*PI); 1074 continue; 1075 } 1076 1077 // If we have seen this block before, but it was with a different 1078 // pointer then we have a phi translation failure and we have to treat 1079 // this as a clobber. 1080 if (InsertRes.first->second != Pointer) 1081 goto PredTranslationFailure; 1082 } 1083 continue; 1084 } 1085 1086 // If we do need to do phi translation, then there are a bunch of different 1087 // cases, because we have to find a Value* live in the predecessor block. We 1088 // know that PtrInst is defined in this block at least. 1089 1090 // We may have added values to the cache list before this PHI translation. 1091 // If so, we haven't done anything to ensure that the cache remains sorted. 1092 // Sort it now (if needed) so that recursive invocations of 1093 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 1094 // value will only see properly sorted cache arrays. 1095 if (Cache && NumSortedEntries != Cache->size()) { 1096 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1097 NumSortedEntries = Cache->size(); 1098 } 1099 1100 // If this is a computation derived from a PHI node, use the suitably 1101 // translated incoming values for each pred as the phi translated version. 1102 if (!isPHITranslatable(PtrInst)) 1103 goto PredTranslationFailure; 1104 1105 Cache = 0; 1106 1107 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1108 BasicBlock *Pred = *PI; 1109 // Get the PHI translated pointer in this predecessor. This can fail and 1110 // return null if not translatable. 1111 Value *PredPtr = GetPHITranslatedValue(PtrInst, BB, Pred, TD); 1112 1113 // Check to see if we have already visited this pred block with another 1114 // pointer. If so, we can't do this lookup. This failure can occur 1115 // with PHI translation when a critical edge exists and the PHI node in 1116 // the successor translates to a pointer value different than the 1117 // pointer the block was first analyzed with. 1118 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1119 InsertRes = Visited.insert(std::make_pair(Pred, PredPtr)); 1120 1121 if (!InsertRes.second) { 1122 // If the predecessor was visited with PredPtr, then we already did 1123 // the analysis and can ignore it. 1124 if (InsertRes.first->second == PredPtr) 1125 continue; 1126 1127 // Otherwise, the block was previously analyzed with a different 1128 // pointer. We can't represent the result of this case, so we just 1129 // treat this as a phi translation failure. 1130 goto PredTranslationFailure; 1131 } 1132 1133 // If PHI translation was unable to find an available pointer in this 1134 // predecessor, then we have to assume that the pointer is clobbered in 1135 // that predecessor. We can still do PRE of the load, which would insert 1136 // a computation of the pointer in this predecessor. 1137 if (PredPtr == 0) { 1138 Result.push_back(NonLocalDepEntry(Pred, 1139 MemDepResult::getClobber(Pred->getTerminator()))); 1140 continue; 1141 } 1142 1143 // FIXME: it is entirely possible that PHI translating will end up with 1144 // the same value. Consider PHI translating something like: 1145 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 1146 // to recurse here, pedantically speaking. 1147 1148 // If we have a problem phi translating, fall through to the code below 1149 // to handle the failure condition. 1150 if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred, 1151 Result, Visited)) 1152 goto PredTranslationFailure; 1153 } 1154 1155 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 1156 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1157 Cache = &CacheInfo->second; 1158 NumSortedEntries = Cache->size(); 1159 1160 // Since we did phi translation, the "Cache" set won't contain all of the 1161 // results for the query. This is ok (we can still use it to accelerate 1162 // specific block queries) but we can't do the fastpath "return all 1163 // results from the set" Clear out the indicator for this. 1164 CacheInfo->first = BBSkipFirstBlockPair(); 1165 SkipFirstBlock = false; 1166 continue; 1167 1168 PredTranslationFailure: 1169 1170 if (Cache == 0) { 1171 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1172 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1173 Cache = &CacheInfo->second; 1174 NumSortedEntries = Cache->size(); 1175 } 1176 1177 // Since we did phi translation, the "Cache" set won't contain all of the 1178 // results for the query. This is ok (we can still use it to accelerate 1179 // specific block queries) but we can't do the fastpath "return all 1180 // results from the set" Clear out the indicator for this. 1181 CacheInfo->first = BBSkipFirstBlockPair(); 1182 1183 // If *nothing* works, mark the pointer as being clobbered by the first 1184 // instruction in this block. 1185 // 1186 // If this is the magic first block, return this as a clobber of the whole 1187 // incoming value. Since we can't phi translate to one of the predecessors, 1188 // we have to bail out. 1189 if (SkipFirstBlock) 1190 return true; 1191 1192 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 1193 assert(I != Cache->rend() && "Didn't find current block??"); 1194 if (I->first != BB) 1195 continue; 1196 1197 assert(I->second.isNonLocal() && 1198 "Should only be here with transparent block"); 1199 I->second = MemDepResult::getClobber(BB->begin()); 1200 ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey); 1201 Result.push_back(*I); 1202 break; 1203 } 1204 } 1205 1206 // Okay, we're done now. If we added new values to the cache, re-sort it. 1207 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1208 DEBUG(AssertSorted(*Cache)); 1209 return false; 1210} 1211 1212/// RemoveCachedNonLocalPointerDependencies - If P exists in 1213/// CachedNonLocalPointerInfo, remove it. 1214void MemoryDependenceAnalysis:: 1215RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 1216 CachedNonLocalPointerInfo::iterator It = 1217 NonLocalPointerDeps.find(P); 1218 if (It == NonLocalPointerDeps.end()) return; 1219 1220 // Remove all of the entries in the BB->val map. This involves removing 1221 // instructions from the reverse map. 1222 NonLocalDepInfo &PInfo = It->second.second; 1223 1224 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 1225 Instruction *Target = PInfo[i].second.getInst(); 1226 if (Target == 0) continue; // Ignore non-local dep results. 1227 assert(Target->getParent() == PInfo[i].first); 1228 1229 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1230 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1231 } 1232 1233 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1234 NonLocalPointerDeps.erase(It); 1235} 1236 1237 1238/// invalidateCachedPointerInfo - This method is used to invalidate cached 1239/// information about the specified pointer, because it may be too 1240/// conservative in memdep. This is an optional call that can be used when 1241/// the client detects an equivalence between the pointer and some other 1242/// value and replaces the other value with ptr. This can make Ptr available 1243/// in more places that cached info does not necessarily keep. 1244void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1245 // If Ptr isn't really a pointer, just ignore it. 1246 if (!isa<PointerType>(Ptr->getType())) return; 1247 // Flush store info for the pointer. 1248 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1249 // Flush load info for the pointer. 1250 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1251} 1252 1253/// removeInstruction - Remove an instruction from the dependence analysis, 1254/// updating the dependence of instructions that previously depended on it. 1255/// This method attempts to keep the cache coherent using the reverse map. 1256void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1257 // Walk through the Non-local dependencies, removing this one as the value 1258 // for any cached queries. 1259 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1260 if (NLDI != NonLocalDeps.end()) { 1261 NonLocalDepInfo &BlockMap = NLDI->second.first; 1262 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1263 DI != DE; ++DI) 1264 if (Instruction *Inst = DI->second.getInst()) 1265 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1266 NonLocalDeps.erase(NLDI); 1267 } 1268 1269 // If we have a cached local dependence query for this instruction, remove it. 1270 // 1271 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1272 if (LocalDepEntry != LocalDeps.end()) { 1273 // Remove us from DepInst's reverse set now that the local dep info is gone. 1274 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1275 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1276 1277 // Remove this local dependency info. 1278 LocalDeps.erase(LocalDepEntry); 1279 } 1280 1281 // If we have any cached pointer dependencies on this instruction, remove 1282 // them. If the instruction has non-pointer type, then it can't be a pointer 1283 // base. 1284 1285 // Remove it from both the load info and the store info. The instruction 1286 // can't be in either of these maps if it is non-pointer. 1287 if (isa<PointerType>(RemInst->getType())) { 1288 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1289 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1290 } 1291 1292 // Loop over all of the things that depend on the instruction we're removing. 1293 // 1294 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1295 1296 // If we find RemInst as a clobber or Def in any of the maps for other values, 1297 // we need to replace its entry with a dirty version of the instruction after 1298 // it. If RemInst is a terminator, we use a null dirty value. 1299 // 1300 // Using a dirty version of the instruction after RemInst saves having to scan 1301 // the entire block to get to this point. 1302 MemDepResult NewDirtyVal; 1303 if (!RemInst->isTerminator()) 1304 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1305 1306 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1307 if (ReverseDepIt != ReverseLocalDeps.end()) { 1308 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1309 // RemInst can't be the terminator if it has local stuff depending on it. 1310 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1311 "Nothing can locally depend on a terminator"); 1312 1313 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1314 E = ReverseDeps.end(); I != E; ++I) { 1315 Instruction *InstDependingOnRemInst = *I; 1316 assert(InstDependingOnRemInst != RemInst && 1317 "Already removed our local dep info"); 1318 1319 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1320 1321 // Make sure to remember that new things depend on NewDepInst. 1322 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1323 "a local dep on this if it is a terminator!"); 1324 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1325 InstDependingOnRemInst)); 1326 } 1327 1328 ReverseLocalDeps.erase(ReverseDepIt); 1329 1330 // Add new reverse deps after scanning the set, to avoid invalidating the 1331 // 'ReverseDeps' reference. 1332 while (!ReverseDepsToAdd.empty()) { 1333 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1334 .insert(ReverseDepsToAdd.back().second); 1335 ReverseDepsToAdd.pop_back(); 1336 } 1337 } 1338 1339 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1340 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1341 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1342 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1343 I != E; ++I) { 1344 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1345 1346 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1347 // The information is now dirty! 1348 INLD.second = true; 1349 1350 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1351 DE = INLD.first.end(); DI != DE; ++DI) { 1352 if (DI->second.getInst() != RemInst) continue; 1353 1354 // Convert to a dirty entry for the subsequent instruction. 1355 DI->second = NewDirtyVal; 1356 1357 if (Instruction *NextI = NewDirtyVal.getInst()) 1358 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1359 } 1360 } 1361 1362 ReverseNonLocalDeps.erase(ReverseDepIt); 1363 1364 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1365 while (!ReverseDepsToAdd.empty()) { 1366 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1367 .insert(ReverseDepsToAdd.back().second); 1368 ReverseDepsToAdd.pop_back(); 1369 } 1370 } 1371 1372 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1373 // value in the NonLocalPointerDeps info. 1374 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1375 ReverseNonLocalPtrDeps.find(RemInst); 1376 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1377 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1378 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1379 1380 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1381 E = Set.end(); I != E; ++I) { 1382 ValueIsLoadPair P = *I; 1383 assert(P.getPointer() != RemInst && 1384 "Already removed NonLocalPointerDeps info for RemInst"); 1385 1386 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second; 1387 1388 // The cache is not valid for any specific block anymore. 1389 NonLocalPointerDeps[P].first = BBSkipFirstBlockPair(); 1390 1391 // Update any entries for RemInst to use the instruction after it. 1392 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1393 DI != DE; ++DI) { 1394 if (DI->second.getInst() != RemInst) continue; 1395 1396 // Convert to a dirty entry for the subsequent instruction. 1397 DI->second = NewDirtyVal; 1398 1399 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1400 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1401 } 1402 1403 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1404 // subsequent value may invalidate the sortedness. 1405 std::sort(NLPDI.begin(), NLPDI.end()); 1406 } 1407 1408 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1409 1410 while (!ReversePtrDepsToAdd.empty()) { 1411 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1412 .insert(ReversePtrDepsToAdd.back().second); 1413 ReversePtrDepsToAdd.pop_back(); 1414 } 1415 } 1416 1417 1418 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1419 AA->deleteValue(RemInst); 1420 DEBUG(verifyRemoved(RemInst)); 1421} 1422/// verifyRemoved - Verify that the specified instruction does not occur 1423/// in our internal data structures. 1424void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1425 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1426 E = LocalDeps.end(); I != E; ++I) { 1427 assert(I->first != D && "Inst occurs in data structures"); 1428 assert(I->second.getInst() != D && 1429 "Inst occurs in data structures"); 1430 } 1431 1432 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1433 E = NonLocalPointerDeps.end(); I != E; ++I) { 1434 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1435 const NonLocalDepInfo &Val = I->second.second; 1436 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1437 II != E; ++II) 1438 assert(II->second.getInst() != D && "Inst occurs as NLPD value"); 1439 } 1440 1441 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1442 E = NonLocalDeps.end(); I != E; ++I) { 1443 assert(I->first != D && "Inst occurs in data structures"); 1444 const PerInstNLInfo &INLD = I->second; 1445 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1446 EE = INLD.first.end(); II != EE; ++II) 1447 assert(II->second.getInst() != D && "Inst occurs in data structures"); 1448 } 1449 1450 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1451 E = ReverseLocalDeps.end(); I != E; ++I) { 1452 assert(I->first != D && "Inst occurs in data structures"); 1453 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1454 EE = I->second.end(); II != EE; ++II) 1455 assert(*II != D && "Inst occurs in data structures"); 1456 } 1457 1458 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1459 E = ReverseNonLocalDeps.end(); 1460 I != E; ++I) { 1461 assert(I->first != D && "Inst occurs in data structures"); 1462 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1463 EE = I->second.end(); II != EE; ++II) 1464 assert(*II != D && "Inst occurs in data structures"); 1465 } 1466 1467 for (ReverseNonLocalPtrDepTy::const_iterator 1468 I = ReverseNonLocalPtrDeps.begin(), 1469 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1470 assert(I->first != D && "Inst occurs in rev NLPD map"); 1471 1472 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1473 E = I->second.end(); II != E; ++II) 1474 assert(*II != ValueIsLoadPair(D, false) && 1475 *II != ValueIsLoadPair(D, true) && 1476 "Inst occurs in ReverseNonLocalPtrDeps map"); 1477 } 1478 1479} 1480