MemoryDependenceAnalysis.cpp revision a3351a0e5db0b5b2c53920f2f15d3e862fecfad3
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/Analysis/ValueTracking.h" 20#include "llvm/Instructions.h" 21#include "llvm/IntrinsicInst.h" 22#include "llvm/Function.h" 23#include "llvm/LLVMContext.h" 24#include "llvm/Analysis/AliasAnalysis.h" 25#include "llvm/Analysis/Dominators.h" 26#include "llvm/Analysis/InstructionSimplify.h" 27#include "llvm/Analysis/MemoryBuiltins.h" 28#include "llvm/Analysis/PHITransAddr.h" 29#include "llvm/Analysis/ValueTracking.h" 30#include "llvm/ADT/Statistic.h" 31#include "llvm/ADT/STLExtras.h" 32#include "llvm/Support/PredIteratorCache.h" 33#include "llvm/Support/Debug.h" 34#include "llvm/Target/TargetData.h" 35using namespace llvm; 36 37STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 38STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 39STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 40 41STATISTIC(NumCacheNonLocalPtr, 42 "Number of fully cached non-local ptr responses"); 43STATISTIC(NumCacheDirtyNonLocalPtr, 44 "Number of cached, but dirty, non-local ptr responses"); 45STATISTIC(NumUncacheNonLocalPtr, 46 "Number of uncached non-local ptr responses"); 47STATISTIC(NumCacheCompleteNonLocalPtr, 48 "Number of block queries that were completely cached"); 49 50char MemoryDependenceAnalysis::ID = 0; 51 52// Register this pass... 53INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep", 54 "Memory Dependence Analysis", false, true) 55INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 56INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep", 57 "Memory Dependence Analysis", false, true) 58 59MemoryDependenceAnalysis::MemoryDependenceAnalysis() 60: FunctionPass(ID), PredCache(0) { 61 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry()); 62} 63MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 64} 65 66/// Clean up memory in between runs 67void MemoryDependenceAnalysis::releaseMemory() { 68 LocalDeps.clear(); 69 NonLocalDeps.clear(); 70 NonLocalPointerDeps.clear(); 71 ReverseLocalDeps.clear(); 72 ReverseNonLocalDeps.clear(); 73 ReverseNonLocalPtrDeps.clear(); 74 PredCache->clear(); 75} 76 77 78 79/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 80/// 81void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 82 AU.setPreservesAll(); 83 AU.addRequiredTransitive<AliasAnalysis>(); 84} 85 86bool MemoryDependenceAnalysis::runOnFunction(Function &) { 87 AA = &getAnalysis<AliasAnalysis>(); 88 TD = getAnalysisIfAvailable<TargetData>(); 89 if (PredCache == 0) 90 PredCache.reset(new PredIteratorCache()); 91 return false; 92} 93 94/// RemoveFromReverseMap - This is a helper function that removes Val from 95/// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry. 96template <typename KeyTy> 97static void RemoveFromReverseMap(DenseMap<Instruction*, 98 SmallPtrSet<KeyTy, 4> > &ReverseMap, 99 Instruction *Inst, KeyTy Val) { 100 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator 101 InstIt = ReverseMap.find(Inst); 102 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); 103 bool Found = InstIt->second.erase(Val); 104 assert(Found && "Invalid reverse map!"); (void)Found; 105 if (InstIt->second.empty()) 106 ReverseMap.erase(InstIt); 107} 108 109/// GetLocation - If the given instruction references a specific memory 110/// location, fill in Loc with the details, otherwise set Loc.Ptr to null. 111/// Return a ModRefInfo value describing the general behavior of the 112/// instruction. 113static 114AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst, 115 AliasAnalysis::Location &Loc, 116 AliasAnalysis *AA) { 117 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 118 if (LI->isVolatile()) { 119 Loc = AliasAnalysis::Location(); 120 return AliasAnalysis::ModRef; 121 } 122 Loc = AA->getLocation(LI); 123 return AliasAnalysis::Ref; 124 } 125 126 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 127 if (SI->isVolatile()) { 128 Loc = AliasAnalysis::Location(); 129 return AliasAnalysis::ModRef; 130 } 131 Loc = AA->getLocation(SI); 132 return AliasAnalysis::Mod; 133 } 134 135 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 136 Loc = AA->getLocation(V); 137 return AliasAnalysis::ModRef; 138 } 139 140 if (const CallInst *CI = isFreeCall(Inst)) { 141 // calls to free() deallocate the entire structure 142 Loc = AliasAnalysis::Location(CI->getArgOperand(0)); 143 return AliasAnalysis::Mod; 144 } 145 146 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) 147 switch (II->getIntrinsicID()) { 148 case Intrinsic::lifetime_start: 149 case Intrinsic::lifetime_end: 150 case Intrinsic::invariant_start: 151 Loc = AliasAnalysis::Location(II->getArgOperand(1), 152 cast<ConstantInt>(II->getArgOperand(0)) 153 ->getZExtValue(), 154 II->getMetadata(LLVMContext::MD_tbaa)); 155 // These intrinsics don't really modify the memory, but returning Mod 156 // will allow them to be handled conservatively. 157 return AliasAnalysis::Mod; 158 case Intrinsic::invariant_end: 159 Loc = AliasAnalysis::Location(II->getArgOperand(2), 160 cast<ConstantInt>(II->getArgOperand(1)) 161 ->getZExtValue(), 162 II->getMetadata(LLVMContext::MD_tbaa)); 163 // These intrinsics don't really modify the memory, but returning Mod 164 // will allow them to be handled conservatively. 165 return AliasAnalysis::Mod; 166 default: 167 break; 168 } 169 170 // Otherwise, just do the coarse-grained thing that always works. 171 if (Inst->mayWriteToMemory()) 172 return AliasAnalysis::ModRef; 173 if (Inst->mayReadFromMemory()) 174 return AliasAnalysis::Ref; 175 return AliasAnalysis::NoModRef; 176} 177 178/// getCallSiteDependencyFrom - Private helper for finding the local 179/// dependencies of a call site. 180MemDepResult MemoryDependenceAnalysis:: 181getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 182 BasicBlock::iterator ScanIt, BasicBlock *BB) { 183 // Walk backwards through the block, looking for dependencies 184 while (ScanIt != BB->begin()) { 185 Instruction *Inst = --ScanIt; 186 187 // If this inst is a memory op, get the pointer it accessed 188 AliasAnalysis::Location Loc; 189 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA); 190 if (Loc.Ptr) { 191 // A simple instruction. 192 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef) 193 return MemDepResult::getClobber(Inst); 194 continue; 195 } 196 197 if (CallSite InstCS = cast<Value>(Inst)) { 198 // Debug intrinsics don't cause dependences. 199 if (isa<DbgInfoIntrinsic>(Inst)) continue; 200 // If these two calls do not interfere, look past it. 201 switch (AA->getModRefInfo(CS, InstCS)) { 202 case AliasAnalysis::NoModRef: 203 // If the two calls are the same, return InstCS as a Def, so that 204 // CS can be found redundant and eliminated. 205 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) && 206 CS.getInstruction()->isIdenticalToWhenDefined(Inst)) 207 return MemDepResult::getDef(Inst); 208 209 // Otherwise if the two calls don't interact (e.g. InstCS is readnone) 210 // keep scanning. 211 break; 212 default: 213 return MemDepResult::getClobber(Inst); 214 } 215 } 216 } 217 218 // No dependence found. If this is the entry block of the function, it is a 219 // clobber, otherwise it is non-local. 220 if (BB != &BB->getParent()->getEntryBlock()) 221 return MemDepResult::getNonLocal(); 222 return MemDepResult::getClobber(ScanIt); 223} 224 225/// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that 226/// would fully overlap MemLoc if done as a wider legal integer load. 227/// 228/// MemLocBase, MemLocOffset are lazily computed here the first time the 229/// base/offs of memloc is needed. 230static bool 231isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc, 232 const Value *&MemLocBase, 233 int64_t &MemLocOffs, 234 const LoadInst *LI, 235 const TargetData *TD) { 236 // If we have no target data, we can't do this. 237 if (TD == 0) return false; 238 239 // If we haven't already computed the base/offset of MemLoc, do so now. 240 if (MemLocBase == 0) 241 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, *TD); 242 243 unsigned Size = MemoryDependenceAnalysis:: 244 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size, 245 LI, *TD); 246 return Size != 0; 247} 248 249/// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that 250/// looks at a memory location for a load (specified by MemLocBase, Offs, 251/// and Size) and compares it against a load. If the specified load could 252/// be safely widened to a larger integer load that is 1) still efficient, 253/// 2) safe for the target, and 3) would provide the specified memory 254/// location value, then this function returns the size in bytes of the 255/// load width to use. If not, this returns zero. 256unsigned MemoryDependenceAnalysis:: 257getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs, 258 unsigned MemLocSize, const LoadInst *LI, 259 const TargetData &TD) { 260 // We can only extend non-volatile integer loads. 261 if (!isa<IntegerType>(LI->getType()) || LI->isVolatile()) return 0; 262 263 // Get the base of this load. 264 int64_t LIOffs = 0; 265 const Value *LIBase = 266 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, TD); 267 268 // If the two pointers are not based on the same pointer, we can't tell that 269 // they are related. 270 if (LIBase != MemLocBase) return 0; 271 272 // Okay, the two values are based on the same pointer, but returned as 273 // no-alias. This happens when we have things like two byte loads at "P+1" 274 // and "P+3". Check to see if increasing the size of the "LI" load up to its 275 // alignment (or the largest native integer type) will allow us to load all 276 // the bits required by MemLoc. 277 278 // If MemLoc is before LI, then no widening of LI will help us out. 279 if (MemLocOffs < LIOffs) return 0; 280 281 // Get the alignment of the load in bytes. We assume that it is safe to load 282 // any legal integer up to this size without a problem. For example, if we're 283 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can 284 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it 285 // to i16. 286 unsigned LoadAlign = LI->getAlignment(); 287 288 int64_t MemLocEnd = MemLocOffs+MemLocSize; 289 290 // If no amount of rounding up will let MemLoc fit into LI, then bail out. 291 if (LIOffs+LoadAlign < MemLocEnd) return 0; 292 293 // This is the size of the load to try. Start with the next larger power of 294 // two. 295 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U; 296 NewLoadByteSize = NextPowerOf2(NewLoadByteSize); 297 298 while (1) { 299 // If this load size is bigger than our known alignment or would not fit 300 // into a native integer register, then we fail. 301 if (NewLoadByteSize > LoadAlign || 302 !TD.fitsInLegalInteger(NewLoadByteSize*8)) 303 return 0; 304 305 // If a load of this width would include all of MemLoc, then we succeed. 306 if (LIOffs+NewLoadByteSize >= MemLocEnd) 307 return NewLoadByteSize; 308 309 NewLoadByteSize <<= 1; 310 } 311 312 return 0; 313} 314 315/// getPointerDependencyFrom - Return the instruction on which a memory 316/// location depends. If isLoad is true, this routine ignores may-aliases with 317/// read-only operations. If isLoad is false, this routine ignores may-aliases 318/// with reads from read-only locations. 319MemDepResult MemoryDependenceAnalysis:: 320getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad, 321 BasicBlock::iterator ScanIt, BasicBlock *BB) { 322 323 const Value *MemLocBase = 0; 324 int64_t MemLocOffset = 0; 325 326 // Walk backwards through the basic block, looking for dependencies. 327 while (ScanIt != BB->begin()) { 328 Instruction *Inst = --ScanIt; 329 330 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 331 // Debug intrinsics don't (and can't) cause dependences. 332 if (isa<DbgInfoIntrinsic>(II)) continue; 333 334 // If we reach a lifetime begin or end marker, then the query ends here 335 // because the value is undefined. 336 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 337 // FIXME: This only considers queries directly on the invariant-tagged 338 // pointer, not on query pointers that are indexed off of them. It'd 339 // be nice to handle that at some point (the right approach is to use 340 // GetPointerBaseWithConstantOffset). 341 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)), 342 MemLoc)) 343 return MemDepResult::getDef(II); 344 continue; 345 } 346 } 347 348 // Values depend on loads if the pointers are must aliased. This means that 349 // a load depends on another must aliased load from the same value. 350 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 351 AliasAnalysis::Location LoadLoc = AA->getLocation(LI); 352 353 // If we found a pointer, check if it could be the same as our pointer. 354 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc); 355 356 if (isLoad) { 357 if (R == AliasAnalysis::NoAlias) { 358 // If this is an over-aligned integer load (for example, 359 // "load i8* %P, align 4") see if it would obviously overlap with the 360 // queried location if widened to a larger load (e.g. if the queried 361 // location is 1 byte at P+1). If so, return it as a load/load 362 // clobber result, allowing the client to decide to widen the load if 363 // it wants to. 364 if (const IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) 365 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() && 366 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase, 367 MemLocOffset, LI, TD)) 368 return MemDepResult::getClobber(Inst); 369 370 continue; 371 } 372 373 // Must aliased loads are defs of each other. 374 if (R == AliasAnalysis::MustAlias) 375 return MemDepResult::getDef(Inst); 376 377#if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads 378 // in terms of clobbering loads, but since it does this by looking 379 // at the clobbering load directly, it doesn't know about any 380 // phi translation that may have happened along the way. 381 382 // If we have a partial alias, then return this as a clobber for the 383 // client to handle. 384 if (R == AliasAnalysis::PartialAlias) 385 return MemDepResult::getClobber(Inst); 386#endif 387 388 // Random may-alias loads don't depend on each other without a 389 // dependence. 390 continue; 391 } 392 393 // Stores don't depend on other no-aliased accesses. 394 if (R == AliasAnalysis::NoAlias) 395 continue; 396 397 // Stores don't alias loads from read-only memory. 398 if (AA->pointsToConstantMemory(LoadLoc)) 399 continue; 400 401 // Stores depend on may/must aliased loads. 402 return MemDepResult::getDef(Inst); 403 } 404 405 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 406 // If alias analysis can tell that this store is guaranteed to not modify 407 // the query pointer, ignore it. Use getModRefInfo to handle cases where 408 // the query pointer points to constant memory etc. 409 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef) 410 continue; 411 412 // Ok, this store might clobber the query pointer. Check to see if it is 413 // a must alias: in this case, we want to return this as a def. 414 AliasAnalysis::Location StoreLoc = AA->getLocation(SI); 415 416 // If we found a pointer, check if it could be the same as our pointer. 417 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc); 418 419 if (R == AliasAnalysis::NoAlias) 420 continue; 421 if (R == AliasAnalysis::MustAlias) 422 return MemDepResult::getDef(Inst); 423 return MemDepResult::getClobber(Inst); 424 } 425 426 // If this is an allocation, and if we know that the accessed pointer is to 427 // the allocation, return Def. This means that there is no dependence and 428 // the access can be optimized based on that. For example, a load could 429 // turn into undef. 430 // Note: Only determine this to be a malloc if Inst is the malloc call, not 431 // a subsequent bitcast of the malloc call result. There can be stores to 432 // the malloced memory between the malloc call and its bitcast uses, and we 433 // need to continue scanning until the malloc call. 434 if (isa<AllocaInst>(Inst) || 435 (isa<CallInst>(Inst) && extractMallocCall(Inst))) { 436 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD); 437 438 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr)) 439 return MemDepResult::getDef(Inst); 440 continue; 441 } 442 443 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 444 switch (AA->getModRefInfo(Inst, MemLoc)) { 445 case AliasAnalysis::NoModRef: 446 // If the call has no effect on the queried pointer, just ignore it. 447 continue; 448 case AliasAnalysis::Mod: 449 return MemDepResult::getClobber(Inst); 450 case AliasAnalysis::Ref: 451 // If the call is known to never store to the pointer, and if this is a 452 // load query, we can safely ignore it (scan past it). 453 if (isLoad) 454 continue; 455 default: 456 // Otherwise, there is a potential dependence. Return a clobber. 457 return MemDepResult::getClobber(Inst); 458 } 459 } 460 461 // No dependence found. If this is the entry block of the function, it is a 462 // clobber, otherwise it is non-local. 463 if (BB != &BB->getParent()->getEntryBlock()) 464 return MemDepResult::getNonLocal(); 465 return MemDepResult::getClobber(ScanIt); 466} 467 468/// getDependency - Return the instruction on which a memory operation 469/// depends. 470MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 471 Instruction *ScanPos = QueryInst; 472 473 // Check for a cached result 474 MemDepResult &LocalCache = LocalDeps[QueryInst]; 475 476 // If the cached entry is non-dirty, just return it. Note that this depends 477 // on MemDepResult's default constructing to 'dirty'. 478 if (!LocalCache.isDirty()) 479 return LocalCache; 480 481 // Otherwise, if we have a dirty entry, we know we can start the scan at that 482 // instruction, which may save us some work. 483 if (Instruction *Inst = LocalCache.getInst()) { 484 ScanPos = Inst; 485 486 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 487 } 488 489 BasicBlock *QueryParent = QueryInst->getParent(); 490 491 // Do the scan. 492 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 493 // No dependence found. If this is the entry block of the function, it is a 494 // clobber, otherwise it is non-local. 495 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 496 LocalCache = MemDepResult::getNonLocal(); 497 else 498 LocalCache = MemDepResult::getClobber(QueryInst); 499 } else { 500 AliasAnalysis::Location MemLoc; 501 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA); 502 if (MemLoc.Ptr) { 503 // If we can do a pointer scan, make it happen. 504 bool isLoad = !(MR & AliasAnalysis::Mod); 505 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst)) 506 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start; 507 508 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos, 509 QueryParent); 510 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 511 CallSite QueryCS(QueryInst); 512 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 513 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 514 QueryParent); 515 } else 516 // Non-memory instruction. 517 LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); 518 } 519 520 // Remember the result! 521 if (Instruction *I = LocalCache.getInst()) 522 ReverseLocalDeps[I].insert(QueryInst); 523 524 return LocalCache; 525} 526 527#ifndef NDEBUG 528/// AssertSorted - This method is used when -debug is specified to verify that 529/// cache arrays are properly kept sorted. 530static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 531 int Count = -1) { 532 if (Count == -1) Count = Cache.size(); 533 if (Count == 0) return; 534 535 for (unsigned i = 1; i != unsigned(Count); ++i) 536 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 537} 538#endif 539 540/// getNonLocalCallDependency - Perform a full dependency query for the 541/// specified call, returning the set of blocks that the value is 542/// potentially live across. The returned set of results will include a 543/// "NonLocal" result for all blocks where the value is live across. 544/// 545/// This method assumes the instruction returns a "NonLocal" dependency 546/// within its own block. 547/// 548/// This returns a reference to an internal data structure that may be 549/// invalidated on the next non-local query or when an instruction is 550/// removed. Clients must copy this data if they want it around longer than 551/// that. 552const MemoryDependenceAnalysis::NonLocalDepInfo & 553MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 554 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 555 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 556 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 557 NonLocalDepInfo &Cache = CacheP.first; 558 559 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 560 /// the cached case, this can happen due to instructions being deleted etc. In 561 /// the uncached case, this starts out as the set of predecessors we care 562 /// about. 563 SmallVector<BasicBlock*, 32> DirtyBlocks; 564 565 if (!Cache.empty()) { 566 // Okay, we have a cache entry. If we know it is not dirty, just return it 567 // with no computation. 568 if (!CacheP.second) { 569 ++NumCacheNonLocal; 570 return Cache; 571 } 572 573 // If we already have a partially computed set of results, scan them to 574 // determine what is dirty, seeding our initial DirtyBlocks worklist. 575 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 576 I != E; ++I) 577 if (I->getResult().isDirty()) 578 DirtyBlocks.push_back(I->getBB()); 579 580 // Sort the cache so that we can do fast binary search lookups below. 581 std::sort(Cache.begin(), Cache.end()); 582 583 ++NumCacheDirtyNonLocal; 584 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 585 // << Cache.size() << " cached: " << *QueryInst; 586 } else { 587 // Seed DirtyBlocks with each of the preds of QueryInst's block. 588 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 589 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 590 DirtyBlocks.push_back(*PI); 591 ++NumUncacheNonLocal; 592 } 593 594 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 595 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 596 597 SmallPtrSet<BasicBlock*, 64> Visited; 598 599 unsigned NumSortedEntries = Cache.size(); 600 DEBUG(AssertSorted(Cache)); 601 602 // Iterate while we still have blocks to update. 603 while (!DirtyBlocks.empty()) { 604 BasicBlock *DirtyBB = DirtyBlocks.back(); 605 DirtyBlocks.pop_back(); 606 607 // Already processed this block? 608 if (!Visited.insert(DirtyBB)) 609 continue; 610 611 // Do a binary search to see if we already have an entry for this block in 612 // the cache set. If so, find it. 613 DEBUG(AssertSorted(Cache, NumSortedEntries)); 614 NonLocalDepInfo::iterator Entry = 615 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 616 NonLocalDepEntry(DirtyBB)); 617 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB) 618 --Entry; 619 620 NonLocalDepEntry *ExistingResult = 0; 621 if (Entry != Cache.begin()+NumSortedEntries && 622 Entry->getBB() == DirtyBB) { 623 // If we already have an entry, and if it isn't already dirty, the block 624 // is done. 625 if (!Entry->getResult().isDirty()) 626 continue; 627 628 // Otherwise, remember this slot so we can update the value. 629 ExistingResult = &*Entry; 630 } 631 632 // If the dirty entry has a pointer, start scanning from it so we don't have 633 // to rescan the entire block. 634 BasicBlock::iterator ScanPos = DirtyBB->end(); 635 if (ExistingResult) { 636 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 637 ScanPos = Inst; 638 // We're removing QueryInst's use of Inst. 639 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 640 QueryCS.getInstruction()); 641 } 642 } 643 644 // Find out if this block has a local dependency for QueryInst. 645 MemDepResult Dep; 646 647 if (ScanPos != DirtyBB->begin()) { 648 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 649 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 650 // No dependence found. If this is the entry block of the function, it is 651 // a clobber, otherwise it is non-local. 652 Dep = MemDepResult::getNonLocal(); 653 } else { 654 Dep = MemDepResult::getClobber(ScanPos); 655 } 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(DirtyBB, Dep)); 663 664 // If the block has a dependency (i.e. it isn't completely transparent to 665 // the value), remember the association! 666 if (!Dep.isNonLocal()) { 667 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 668 // update this when we remove instructions. 669 if (Instruction *Inst = Dep.getInst()) 670 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 671 } else { 672 673 // If the block *is* completely transparent to the load, we need to check 674 // the predecessors of this block. Add them to our worklist. 675 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 676 DirtyBlocks.push_back(*PI); 677 } 678 } 679 680 return Cache; 681} 682 683/// getNonLocalPointerDependency - Perform a full dependency query for an 684/// access to the specified (non-volatile) memory location, returning the 685/// set of instructions that either define or clobber the value. 686/// 687/// This method assumes the pointer has a "NonLocal" dependency within its 688/// own block. 689/// 690void MemoryDependenceAnalysis:: 691getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad, 692 BasicBlock *FromBB, 693 SmallVectorImpl<NonLocalDepResult> &Result) { 694 assert(Loc.Ptr->getType()->isPointerTy() && 695 "Can't get pointer deps of a non-pointer!"); 696 Result.clear(); 697 698 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD); 699 700 // This is the set of blocks we've inspected, and the pointer we consider in 701 // each block. Because of critical edges, we currently bail out if querying 702 // a block with multiple different pointers. This can happen during PHI 703 // translation. 704 DenseMap<BasicBlock*, Value*> Visited; 705 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB, 706 Result, Visited, true)) 707 return; 708 Result.clear(); 709 Result.push_back(NonLocalDepResult(FromBB, 710 MemDepResult::getClobber(FromBB->begin()), 711 const_cast<Value *>(Loc.Ptr))); 712} 713 714/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 715/// Pointer/PointeeSize using either cached information in Cache or by doing a 716/// lookup (which may use dirty cache info if available). If we do a lookup, 717/// add the result to the cache. 718MemDepResult MemoryDependenceAnalysis:: 719GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc, 720 bool isLoad, BasicBlock *BB, 721 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 722 723 // Do a binary search to see if we already have an entry for this block in 724 // the cache set. If so, find it. 725 NonLocalDepInfo::iterator Entry = 726 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 727 NonLocalDepEntry(BB)); 728 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 729 --Entry; 730 731 NonLocalDepEntry *ExistingResult = 0; 732 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 733 ExistingResult = &*Entry; 734 735 // If we have a cached entry, and it is non-dirty, use it as the value for 736 // this dependency. 737 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 738 ++NumCacheNonLocalPtr; 739 return ExistingResult->getResult(); 740 } 741 742 // Otherwise, we have to scan for the value. If we have a dirty cache 743 // entry, start scanning from its position, otherwise we scan from the end 744 // of the block. 745 BasicBlock::iterator ScanPos = BB->end(); 746 if (ExistingResult && ExistingResult->getResult().getInst()) { 747 assert(ExistingResult->getResult().getInst()->getParent() == BB && 748 "Instruction invalidated?"); 749 ++NumCacheDirtyNonLocalPtr; 750 ScanPos = ExistingResult->getResult().getInst(); 751 752 // Eliminating the dirty entry from 'Cache', so update the reverse info. 753 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 754 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 755 } else { 756 ++NumUncacheNonLocalPtr; 757 } 758 759 // Scan the block for the dependency. 760 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB); 761 762 // If we had a dirty entry for the block, update it. Otherwise, just add 763 // a new entry. 764 if (ExistingResult) 765 ExistingResult->setResult(Dep); 766 else 767 Cache->push_back(NonLocalDepEntry(BB, Dep)); 768 769 // If the block has a dependency (i.e. it isn't completely transparent to 770 // the value), remember the reverse association because we just added it 771 // to Cache! 772 if (Dep.isNonLocal()) 773 return Dep; 774 775 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 776 // update MemDep when we remove instructions. 777 Instruction *Inst = Dep.getInst(); 778 assert(Inst && "Didn't depend on anything?"); 779 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 780 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 781 return Dep; 782} 783 784/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 785/// number of elements in the array that are already properly ordered. This is 786/// optimized for the case when only a few entries are added. 787static void 788SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 789 unsigned NumSortedEntries) { 790 switch (Cache.size() - NumSortedEntries) { 791 case 0: 792 // done, no new entries. 793 break; 794 case 2: { 795 // Two new entries, insert the last one into place. 796 NonLocalDepEntry Val = Cache.back(); 797 Cache.pop_back(); 798 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 799 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 800 Cache.insert(Entry, Val); 801 // FALL THROUGH. 802 } 803 case 1: 804 // One new entry, Just insert the new value at the appropriate position. 805 if (Cache.size() != 1) { 806 NonLocalDepEntry Val = Cache.back(); 807 Cache.pop_back(); 808 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 809 std::upper_bound(Cache.begin(), Cache.end(), Val); 810 Cache.insert(Entry, Val); 811 } 812 break; 813 default: 814 // Added many values, do a full scale sort. 815 std::sort(Cache.begin(), Cache.end()); 816 break; 817 } 818} 819 820/// getNonLocalPointerDepFromBB - Perform a dependency query based on 821/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 822/// results to the results vector and keep track of which blocks are visited in 823/// 'Visited'. 824/// 825/// This has special behavior for the first block queries (when SkipFirstBlock 826/// is true). In this special case, it ignores the contents of the specified 827/// block and starts returning dependence info for its predecessors. 828/// 829/// This function returns false on success, or true to indicate that it could 830/// not compute dependence information for some reason. This should be treated 831/// as a clobber dependence on the first instruction in the predecessor block. 832bool MemoryDependenceAnalysis:: 833getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, 834 const AliasAnalysis::Location &Loc, 835 bool isLoad, BasicBlock *StartBB, 836 SmallVectorImpl<NonLocalDepResult> &Result, 837 DenseMap<BasicBlock*, Value*> &Visited, 838 bool SkipFirstBlock) { 839 840 // Look up the cached info for Pointer. 841 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 842 843 // Set up a temporary NLPI value. If the map doesn't yet have an entry for 844 // CacheKey, this value will be inserted as the associated value. Otherwise, 845 // it'll be ignored, and we'll have to check to see if the cached size and 846 // tbaa tag are consistent with the current query. 847 NonLocalPointerInfo InitialNLPI; 848 InitialNLPI.Size = Loc.Size; 849 InitialNLPI.TBAATag = Loc.TBAATag; 850 851 // Get the NLPI for CacheKey, inserting one into the map if it doesn't 852 // already have one. 853 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair = 854 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI)); 855 NonLocalPointerInfo *CacheInfo = &Pair.first->second; 856 857 // If we already have a cache entry for this CacheKey, we may need to do some 858 // work to reconcile the cache entry and the current query. 859 if (!Pair.second) { 860 if (CacheInfo->Size < Loc.Size) { 861 // The query's Size is greater than the cached one. Throw out the 862 // cached data and procede with the query at the greater size. 863 CacheInfo->Pair = BBSkipFirstBlockPair(); 864 CacheInfo->Size = Loc.Size; 865 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 866 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 867 if (Instruction *Inst = DI->getResult().getInst()) 868 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 869 CacheInfo->NonLocalDeps.clear(); 870 } else if (CacheInfo->Size > Loc.Size) { 871 // This query's Size is less than the cached one. Conservatively restart 872 // the query using the greater size. 873 return getNonLocalPointerDepFromBB(Pointer, 874 Loc.getWithNewSize(CacheInfo->Size), 875 isLoad, StartBB, Result, Visited, 876 SkipFirstBlock); 877 } 878 879 // If the query's TBAATag is inconsistent with the cached one, 880 // conservatively throw out the cached data and restart the query with 881 // no tag if needed. 882 if (CacheInfo->TBAATag != Loc.TBAATag) { 883 if (CacheInfo->TBAATag) { 884 CacheInfo->Pair = BBSkipFirstBlockPair(); 885 CacheInfo->TBAATag = 0; 886 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 887 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 888 if (Instruction *Inst = DI->getResult().getInst()) 889 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 890 CacheInfo->NonLocalDeps.clear(); 891 } 892 if (Loc.TBAATag) 893 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(), 894 isLoad, StartBB, Result, Visited, 895 SkipFirstBlock); 896 } 897 } 898 899 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps; 900 901 // If we have valid cached information for exactly the block we are 902 // investigating, just return it with no recomputation. 903 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 904 // We have a fully cached result for this query then we can just return the 905 // cached results and populate the visited set. However, we have to verify 906 // that we don't already have conflicting results for these blocks. Check 907 // to ensure that if a block in the results set is in the visited set that 908 // it was for the same pointer query. 909 if (!Visited.empty()) { 910 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 911 I != E; ++I) { 912 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 913 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 914 continue; 915 916 // We have a pointer mismatch in a block. Just return clobber, saying 917 // that something was clobbered in this result. We could also do a 918 // non-fully cached query, but there is little point in doing this. 919 return true; 920 } 921 } 922 923 Value *Addr = Pointer.getAddr(); 924 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 925 I != E; ++I) { 926 Visited.insert(std::make_pair(I->getBB(), Addr)); 927 if (!I->getResult().isNonLocal()) 928 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 929 } 930 ++NumCacheCompleteNonLocalPtr; 931 return false; 932 } 933 934 // Otherwise, either this is a new block, a block with an invalid cache 935 // pointer or one that we're about to invalidate by putting more info into it 936 // than its valid cache info. If empty, the result will be valid cache info, 937 // otherwise it isn't. 938 if (Cache->empty()) 939 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 940 else 941 CacheInfo->Pair = BBSkipFirstBlockPair(); 942 943 SmallVector<BasicBlock*, 32> Worklist; 944 Worklist.push_back(StartBB); 945 946 // PredList used inside loop. 947 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList; 948 949 // Keep track of the entries that we know are sorted. Previously cached 950 // entries will all be sorted. The entries we add we only sort on demand (we 951 // don't insert every element into its sorted position). We know that we 952 // won't get any reuse from currently inserted values, because we don't 953 // revisit blocks after we insert info for them. 954 unsigned NumSortedEntries = Cache->size(); 955 DEBUG(AssertSorted(*Cache)); 956 957 while (!Worklist.empty()) { 958 BasicBlock *BB = Worklist.pop_back_val(); 959 960 // Skip the first block if we have it. 961 if (!SkipFirstBlock) { 962 // Analyze the dependency of *Pointer in FromBB. See if we already have 963 // been here. 964 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 965 966 // Get the dependency info for Pointer in BB. If we have cached 967 // information, we will use it, otherwise we compute it. 968 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 969 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache, 970 NumSortedEntries); 971 972 // If we got a Def or Clobber, add this to the list of results. 973 if (!Dep.isNonLocal()) { 974 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 975 continue; 976 } 977 } 978 979 // If 'Pointer' is an instruction defined in this block, then we need to do 980 // phi translation to change it into a value live in the predecessor block. 981 // If not, we just add the predecessors to the worklist and scan them with 982 // the same Pointer. 983 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 984 SkipFirstBlock = false; 985 SmallVector<BasicBlock*, 16> NewBlocks; 986 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 987 // Verify that we haven't looked at this block yet. 988 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 989 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 990 if (InsertRes.second) { 991 // First time we've looked at *PI. 992 NewBlocks.push_back(*PI); 993 continue; 994 } 995 996 // If we have seen this block before, but it was with a different 997 // pointer then we have a phi translation failure and we have to treat 998 // this as a clobber. 999 if (InsertRes.first->second != Pointer.getAddr()) { 1000 // Make sure to clean up the Visited map before continuing on to 1001 // PredTranslationFailure. 1002 for (unsigned i = 0; i < NewBlocks.size(); i++) 1003 Visited.erase(NewBlocks[i]); 1004 goto PredTranslationFailure; 1005 } 1006 } 1007 Worklist.append(NewBlocks.begin(), NewBlocks.end()); 1008 continue; 1009 } 1010 1011 // We do need to do phi translation, if we know ahead of time we can't phi 1012 // translate this value, don't even try. 1013 if (!Pointer.IsPotentiallyPHITranslatable()) 1014 goto PredTranslationFailure; 1015 1016 // We may have added values to the cache list before this PHI translation. 1017 // If so, we haven't done anything to ensure that the cache remains sorted. 1018 // Sort it now (if needed) so that recursive invocations of 1019 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 1020 // value will only see properly sorted cache arrays. 1021 if (Cache && NumSortedEntries != Cache->size()) { 1022 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1023 NumSortedEntries = Cache->size(); 1024 } 1025 Cache = 0; 1026 1027 PredList.clear(); 1028 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1029 BasicBlock *Pred = *PI; 1030 PredList.push_back(std::make_pair(Pred, Pointer)); 1031 1032 // Get the PHI translated pointer in this predecessor. This can fail if 1033 // not translatable, in which case the getAddr() returns null. 1034 PHITransAddr &PredPointer = PredList.back().second; 1035 PredPointer.PHITranslateValue(BB, Pred, 0); 1036 1037 Value *PredPtrVal = PredPointer.getAddr(); 1038 1039 // Check to see if we have already visited this pred block with another 1040 // pointer. If so, we can't do this lookup. This failure can occur 1041 // with PHI translation when a critical edge exists and the PHI node in 1042 // the successor translates to a pointer value different than the 1043 // pointer the block was first analyzed with. 1044 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1045 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 1046 1047 if (!InsertRes.second) { 1048 // We found the pred; take it off the list of preds to visit. 1049 PredList.pop_back(); 1050 1051 // If the predecessor was visited with PredPtr, then we already did 1052 // the analysis and can ignore it. 1053 if (InsertRes.first->second == PredPtrVal) 1054 continue; 1055 1056 // Otherwise, the block was previously analyzed with a different 1057 // pointer. We can't represent the result of this case, so we just 1058 // treat this as a phi translation failure. 1059 1060 // Make sure to clean up the Visited map before continuing on to 1061 // PredTranslationFailure. 1062 for (unsigned i = 0; i < PredList.size(); i++) 1063 Visited.erase(PredList[i].first); 1064 1065 goto PredTranslationFailure; 1066 } 1067 } 1068 1069 // Actually process results here; this need to be a separate loop to avoid 1070 // calling getNonLocalPointerDepFromBB for blocks we don't want to return 1071 // any results for. (getNonLocalPointerDepFromBB will modify our 1072 // datastructures in ways the code after the PredTranslationFailure label 1073 // doesn't expect.) 1074 for (unsigned i = 0; i < PredList.size(); i++) { 1075 BasicBlock *Pred = PredList[i].first; 1076 PHITransAddr &PredPointer = PredList[i].second; 1077 Value *PredPtrVal = PredPointer.getAddr(); 1078 1079 bool CanTranslate = true; 1080 // If PHI translation was unable to find an available pointer in this 1081 // predecessor, then we have to assume that the pointer is clobbered in 1082 // that predecessor. We can still do PRE of the load, which would insert 1083 // a computation of the pointer in this predecessor. 1084 if (PredPtrVal == 0) 1085 CanTranslate = false; 1086 1087 // FIXME: it is entirely possible that PHI translating will end up with 1088 // the same value. Consider PHI translating something like: 1089 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 1090 // to recurse here, pedantically speaking. 1091 1092 // If getNonLocalPointerDepFromBB fails here, that means the cached 1093 // result conflicted with the Visited list; we have to conservatively 1094 // assume a clobber, but this also does not block PRE of the load. 1095 if (!CanTranslate || 1096 getNonLocalPointerDepFromBB(PredPointer, 1097 Loc.getWithNewPtr(PredPtrVal), 1098 isLoad, Pred, 1099 Result, Visited)) { 1100 // Add the entry to the Result list. 1101 NonLocalDepResult Entry(Pred, 1102 MemDepResult::getClobber(Pred->getTerminator()), 1103 PredPtrVal); 1104 Result.push_back(Entry); 1105 1106 // Since we had a phi translation failure, the cache for CacheKey won't 1107 // include all of the entries that we need to immediately satisfy future 1108 // queries. Mark this in NonLocalPointerDeps by setting the 1109 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 1110 // cached value to do more work but not miss the phi trans failure. 1111 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey]; 1112 NLPI.Pair = BBSkipFirstBlockPair(); 1113 continue; 1114 } 1115 } 1116 1117 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 1118 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1119 Cache = &CacheInfo->NonLocalDeps; 1120 NumSortedEntries = Cache->size(); 1121 1122 // Since we did phi translation, the "Cache" set won't contain all of the 1123 // results for the query. This is ok (we can still use it to accelerate 1124 // specific block queries) but we can't do the fastpath "return all 1125 // results from the set" Clear out the indicator for this. 1126 CacheInfo->Pair = BBSkipFirstBlockPair(); 1127 SkipFirstBlock = false; 1128 continue; 1129 1130 PredTranslationFailure: 1131 // The following code is "failure"; we can't produce a sane translation 1132 // for the given block. It assumes that we haven't modified any of 1133 // our datastructures while processing the current block. 1134 1135 if (Cache == 0) { 1136 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1137 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1138 Cache = &CacheInfo->NonLocalDeps; 1139 NumSortedEntries = Cache->size(); 1140 } 1141 1142 // Since we failed phi translation, the "Cache" set won't contain all of the 1143 // results for the query. This is ok (we can still use it to accelerate 1144 // specific block queries) but we can't do the fastpath "return all 1145 // results from the set". Clear out the indicator for this. 1146 CacheInfo->Pair = BBSkipFirstBlockPair(); 1147 1148 // If *nothing* works, mark the pointer as being clobbered by the first 1149 // instruction in this block. 1150 // 1151 // If this is the magic first block, return this as a clobber of the whole 1152 // incoming value. Since we can't phi translate to one of the predecessors, 1153 // we have to bail out. 1154 if (SkipFirstBlock) 1155 return true; 1156 1157 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 1158 assert(I != Cache->rend() && "Didn't find current block??"); 1159 if (I->getBB() != BB) 1160 continue; 1161 1162 assert(I->getResult().isNonLocal() && 1163 "Should only be here with transparent block"); 1164 I->setResult(MemDepResult::getClobber(BB->getTerminator())); 1165 ReverseNonLocalPtrDeps[BB->getTerminator()].insert(CacheKey); 1166 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 1167 Pointer.getAddr())); 1168 break; 1169 } 1170 } 1171 1172 // Okay, we're done now. If we added new values to the cache, re-sort it. 1173 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1174 DEBUG(AssertSorted(*Cache)); 1175 return false; 1176} 1177 1178/// RemoveCachedNonLocalPointerDependencies - If P exists in 1179/// CachedNonLocalPointerInfo, remove it. 1180void MemoryDependenceAnalysis:: 1181RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 1182 CachedNonLocalPointerInfo::iterator It = 1183 NonLocalPointerDeps.find(P); 1184 if (It == NonLocalPointerDeps.end()) return; 1185 1186 // Remove all of the entries in the BB->val map. This involves removing 1187 // instructions from the reverse map. 1188 NonLocalDepInfo &PInfo = It->second.NonLocalDeps; 1189 1190 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 1191 Instruction *Target = PInfo[i].getResult().getInst(); 1192 if (Target == 0) continue; // Ignore non-local dep results. 1193 assert(Target->getParent() == PInfo[i].getBB()); 1194 1195 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1196 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1197 } 1198 1199 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1200 NonLocalPointerDeps.erase(It); 1201} 1202 1203 1204/// invalidateCachedPointerInfo - This method is used to invalidate cached 1205/// information about the specified pointer, because it may be too 1206/// conservative in memdep. This is an optional call that can be used when 1207/// the client detects an equivalence between the pointer and some other 1208/// value and replaces the other value with ptr. This can make Ptr available 1209/// in more places that cached info does not necessarily keep. 1210void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1211 // If Ptr isn't really a pointer, just ignore it. 1212 if (!Ptr->getType()->isPointerTy()) return; 1213 // Flush store info for the pointer. 1214 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1215 // Flush load info for the pointer. 1216 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1217} 1218 1219/// invalidateCachedPredecessors - Clear the PredIteratorCache info. 1220/// This needs to be done when the CFG changes, e.g., due to splitting 1221/// critical edges. 1222void MemoryDependenceAnalysis::invalidateCachedPredecessors() { 1223 PredCache->clear(); 1224} 1225 1226/// removeInstruction - Remove an instruction from the dependence analysis, 1227/// updating the dependence of instructions that previously depended on it. 1228/// This method attempts to keep the cache coherent using the reverse map. 1229void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1230 // Walk through the Non-local dependencies, removing this one as the value 1231 // for any cached queries. 1232 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1233 if (NLDI != NonLocalDeps.end()) { 1234 NonLocalDepInfo &BlockMap = NLDI->second.first; 1235 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1236 DI != DE; ++DI) 1237 if (Instruction *Inst = DI->getResult().getInst()) 1238 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1239 NonLocalDeps.erase(NLDI); 1240 } 1241 1242 // If we have a cached local dependence query for this instruction, remove it. 1243 // 1244 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1245 if (LocalDepEntry != LocalDeps.end()) { 1246 // Remove us from DepInst's reverse set now that the local dep info is gone. 1247 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1248 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1249 1250 // Remove this local dependency info. 1251 LocalDeps.erase(LocalDepEntry); 1252 } 1253 1254 // If we have any cached pointer dependencies on this instruction, remove 1255 // them. If the instruction has non-pointer type, then it can't be a pointer 1256 // base. 1257 1258 // Remove it from both the load info and the store info. The instruction 1259 // can't be in either of these maps if it is non-pointer. 1260 if (RemInst->getType()->isPointerTy()) { 1261 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1262 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1263 } 1264 1265 // Loop over all of the things that depend on the instruction we're removing. 1266 // 1267 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1268 1269 // If we find RemInst as a clobber or Def in any of the maps for other values, 1270 // we need to replace its entry with a dirty version of the instruction after 1271 // it. If RemInst is a terminator, we use a null dirty value. 1272 // 1273 // Using a dirty version of the instruction after RemInst saves having to scan 1274 // the entire block to get to this point. 1275 MemDepResult NewDirtyVal; 1276 if (!RemInst->isTerminator()) 1277 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1278 1279 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1280 if (ReverseDepIt != ReverseLocalDeps.end()) { 1281 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1282 // RemInst can't be the terminator if it has local stuff depending on it. 1283 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1284 "Nothing can locally depend on a terminator"); 1285 1286 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1287 E = ReverseDeps.end(); I != E; ++I) { 1288 Instruction *InstDependingOnRemInst = *I; 1289 assert(InstDependingOnRemInst != RemInst && 1290 "Already removed our local dep info"); 1291 1292 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1293 1294 // Make sure to remember that new things depend on NewDepInst. 1295 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1296 "a local dep on this if it is a terminator!"); 1297 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1298 InstDependingOnRemInst)); 1299 } 1300 1301 ReverseLocalDeps.erase(ReverseDepIt); 1302 1303 // Add new reverse deps after scanning the set, to avoid invalidating the 1304 // 'ReverseDeps' reference. 1305 while (!ReverseDepsToAdd.empty()) { 1306 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1307 .insert(ReverseDepsToAdd.back().second); 1308 ReverseDepsToAdd.pop_back(); 1309 } 1310 } 1311 1312 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1313 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1314 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1315 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1316 I != E; ++I) { 1317 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1318 1319 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1320 // The information is now dirty! 1321 INLD.second = true; 1322 1323 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1324 DE = INLD.first.end(); DI != DE; ++DI) { 1325 if (DI->getResult().getInst() != RemInst) continue; 1326 1327 // Convert to a dirty entry for the subsequent instruction. 1328 DI->setResult(NewDirtyVal); 1329 1330 if (Instruction *NextI = NewDirtyVal.getInst()) 1331 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1332 } 1333 } 1334 1335 ReverseNonLocalDeps.erase(ReverseDepIt); 1336 1337 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1338 while (!ReverseDepsToAdd.empty()) { 1339 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1340 .insert(ReverseDepsToAdd.back().second); 1341 ReverseDepsToAdd.pop_back(); 1342 } 1343 } 1344 1345 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1346 // value in the NonLocalPointerDeps info. 1347 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1348 ReverseNonLocalPtrDeps.find(RemInst); 1349 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1350 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1351 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1352 1353 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1354 E = Set.end(); I != E; ++I) { 1355 ValueIsLoadPair P = *I; 1356 assert(P.getPointer() != RemInst && 1357 "Already removed NonLocalPointerDeps info for RemInst"); 1358 1359 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps; 1360 1361 // The cache is not valid for any specific block anymore. 1362 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair(); 1363 1364 // Update any entries for RemInst to use the instruction after it. 1365 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1366 DI != DE; ++DI) { 1367 if (DI->getResult().getInst() != RemInst) continue; 1368 1369 // Convert to a dirty entry for the subsequent instruction. 1370 DI->setResult(NewDirtyVal); 1371 1372 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1373 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1374 } 1375 1376 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1377 // subsequent value may invalidate the sortedness. 1378 std::sort(NLPDI.begin(), NLPDI.end()); 1379 } 1380 1381 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1382 1383 while (!ReversePtrDepsToAdd.empty()) { 1384 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1385 .insert(ReversePtrDepsToAdd.back().second); 1386 ReversePtrDepsToAdd.pop_back(); 1387 } 1388 } 1389 1390 1391 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1392 AA->deleteValue(RemInst); 1393 DEBUG(verifyRemoved(RemInst)); 1394} 1395/// verifyRemoved - Verify that the specified instruction does not occur 1396/// in our internal data structures. 1397void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1398 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1399 E = LocalDeps.end(); I != E; ++I) { 1400 assert(I->first != D && "Inst occurs in data structures"); 1401 assert(I->second.getInst() != D && 1402 "Inst occurs in data structures"); 1403 } 1404 1405 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1406 E = NonLocalPointerDeps.end(); I != E; ++I) { 1407 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1408 const NonLocalDepInfo &Val = I->second.NonLocalDeps; 1409 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1410 II != E; ++II) 1411 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1412 } 1413 1414 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1415 E = NonLocalDeps.end(); I != E; ++I) { 1416 assert(I->first != D && "Inst occurs in data structures"); 1417 const PerInstNLInfo &INLD = I->second; 1418 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1419 EE = INLD.first.end(); II != EE; ++II) 1420 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1421 } 1422 1423 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1424 E = ReverseLocalDeps.end(); I != E; ++I) { 1425 assert(I->first != D && "Inst occurs in data structures"); 1426 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1427 EE = I->second.end(); II != EE; ++II) 1428 assert(*II != D && "Inst occurs in data structures"); 1429 } 1430 1431 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1432 E = ReverseNonLocalDeps.end(); 1433 I != E; ++I) { 1434 assert(I->first != D && "Inst occurs in data structures"); 1435 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1436 EE = I->second.end(); II != EE; ++II) 1437 assert(*II != D && "Inst occurs in data structures"); 1438 } 1439 1440 for (ReverseNonLocalPtrDepTy::const_iterator 1441 I = ReverseNonLocalPtrDeps.begin(), 1442 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1443 assert(I->first != D && "Inst occurs in rev NLPD map"); 1444 1445 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1446 E = I->second.end(); II != E; ++II) 1447 assert(*II != ValueIsLoadPair(D, false) && 1448 *II != ValueIsLoadPair(D, true) && 1449 "Inst occurs in ReverseNonLocalPtrDeps map"); 1450 } 1451 1452} 1453