1 //===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation -------------===// 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 #include "llvm/Analysis/MemoryDependenceAnalysis.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/Statistic.h" 20 #include "llvm/Analysis/AliasAnalysis.h" 21 #include "llvm/Analysis/AssumptionCache.h" 22 #include "llvm/Analysis/InstructionSimplify.h" 23 #include "llvm/Analysis/MemoryBuiltins.h" 24 #include "llvm/Analysis/PHITransAddr.h" 25 #include "llvm/Analysis/ValueTracking.h" 26 #include "llvm/IR/DataLayout.h" 27 #include "llvm/IR/Dominators.h" 28 #include "llvm/IR/Function.h" 29 #include "llvm/IR/Instructions.h" 30 #include "llvm/IR/IntrinsicInst.h" 31 #include "llvm/IR/LLVMContext.h" 32 #include "llvm/IR/PredIteratorCache.h" 33 #include "llvm/Support/Debug.h" 34 using namespace llvm; 35 36 #define DEBUG_TYPE "memdep" 37 38 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 39 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 40 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 41 42 STATISTIC(NumCacheNonLocalPtr, 43 "Number of fully cached non-local ptr responses"); 44 STATISTIC(NumCacheDirtyNonLocalPtr, 45 "Number of cached, but dirty, non-local ptr responses"); 46 STATISTIC(NumUncacheNonLocalPtr, 47 "Number of uncached non-local ptr responses"); 48 STATISTIC(NumCacheCompleteNonLocalPtr, 49 "Number of block queries that were completely cached"); 50 51 // Limit for the number of instructions to scan in a block. 52 static const unsigned int BlockScanLimit = 100; 53 54 // Limit on the number of memdep results to process. 55 static const unsigned int NumResultsLimit = 100; 56 57 char MemoryDependenceAnalysis::ID = 0; 58 59 // Register this pass... 60 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep", 61 "Memory Dependence Analysis", false, true) 62 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 63 INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 64 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep", 65 "Memory Dependence Analysis", false, true) 66 67 MemoryDependenceAnalysis::MemoryDependenceAnalysis() 68 : FunctionPass(ID), PredCache() { 69 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry()); 70 } 71 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 72 } 73 74 /// Clean up memory in between runs 75 void MemoryDependenceAnalysis::releaseMemory() { 76 LocalDeps.clear(); 77 NonLocalDeps.clear(); 78 NonLocalPointerDeps.clear(); 79 ReverseLocalDeps.clear(); 80 ReverseNonLocalDeps.clear(); 81 ReverseNonLocalPtrDeps.clear(); 82 PredCache->clear(); 83 } 84 85 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 86 /// 87 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 88 AU.setPreservesAll(); 89 AU.addRequired<AssumptionCacheTracker>(); 90 AU.addRequiredTransitive<AliasAnalysis>(); 91 } 92 93 bool MemoryDependenceAnalysis::runOnFunction(Function &F) { 94 AA = &getAnalysis<AliasAnalysis>(); 95 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 96 DominatorTreeWrapperPass *DTWP = 97 getAnalysisIfAvailable<DominatorTreeWrapperPass>(); 98 DT = DTWP ? &DTWP->getDomTree() : nullptr; 99 if (!PredCache) 100 PredCache.reset(new PredIteratorCache()); 101 return false; 102 } 103 104 /// RemoveFromReverseMap - This is a helper function that removes Val from 105 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry. 106 template <typename KeyTy> 107 static void RemoveFromReverseMap(DenseMap<Instruction*, 108 SmallPtrSet<KeyTy, 4> > &ReverseMap, 109 Instruction *Inst, KeyTy Val) { 110 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator 111 InstIt = ReverseMap.find(Inst); 112 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); 113 bool Found = InstIt->second.erase(Val); 114 assert(Found && "Invalid reverse map!"); (void)Found; 115 if (InstIt->second.empty()) 116 ReverseMap.erase(InstIt); 117 } 118 119 /// GetLocation - If the given instruction references a specific memory 120 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null. 121 /// Return a ModRefInfo value describing the general behavior of the 122 /// instruction. 123 static 124 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst, 125 AliasAnalysis::Location &Loc, 126 AliasAnalysis *AA) { 127 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 128 if (LI->isUnordered()) { 129 Loc = AA->getLocation(LI); 130 return AliasAnalysis::Ref; 131 } 132 if (LI->getOrdering() == Monotonic) { 133 Loc = AA->getLocation(LI); 134 return AliasAnalysis::ModRef; 135 } 136 Loc = AliasAnalysis::Location(); 137 return AliasAnalysis::ModRef; 138 } 139 140 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 141 if (SI->isUnordered()) { 142 Loc = AA->getLocation(SI); 143 return AliasAnalysis::Mod; 144 } 145 if (SI->getOrdering() == Monotonic) { 146 Loc = AA->getLocation(SI); 147 return AliasAnalysis::ModRef; 148 } 149 Loc = AliasAnalysis::Location(); 150 return AliasAnalysis::ModRef; 151 } 152 153 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 154 Loc = AA->getLocation(V); 155 return AliasAnalysis::ModRef; 156 } 157 158 if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) { 159 // calls to free() deallocate the entire structure 160 Loc = AliasAnalysis::Location(CI->getArgOperand(0)); 161 return AliasAnalysis::Mod; 162 } 163 164 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 165 AAMDNodes AAInfo; 166 167 switch (II->getIntrinsicID()) { 168 case Intrinsic::lifetime_start: 169 case Intrinsic::lifetime_end: 170 case Intrinsic::invariant_start: 171 II->getAAMetadata(AAInfo); 172 Loc = AliasAnalysis::Location(II->getArgOperand(1), 173 cast<ConstantInt>(II->getArgOperand(0)) 174 ->getZExtValue(), AAInfo); 175 // These intrinsics don't really modify the memory, but returning Mod 176 // will allow them to be handled conservatively. 177 return AliasAnalysis::Mod; 178 case Intrinsic::invariant_end: 179 II->getAAMetadata(AAInfo); 180 Loc = AliasAnalysis::Location(II->getArgOperand(2), 181 cast<ConstantInt>(II->getArgOperand(1)) 182 ->getZExtValue(), AAInfo); 183 // These intrinsics don't really modify the memory, but returning Mod 184 // will allow them to be handled conservatively. 185 return AliasAnalysis::Mod; 186 default: 187 break; 188 } 189 } 190 191 // Otherwise, just do the coarse-grained thing that always works. 192 if (Inst->mayWriteToMemory()) 193 return AliasAnalysis::ModRef; 194 if (Inst->mayReadFromMemory()) 195 return AliasAnalysis::Ref; 196 return AliasAnalysis::NoModRef; 197 } 198 199 /// getCallSiteDependencyFrom - Private helper for finding the local 200 /// dependencies of a call site. 201 MemDepResult MemoryDependenceAnalysis:: 202 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 203 BasicBlock::iterator ScanIt, BasicBlock *BB) { 204 unsigned Limit = BlockScanLimit; 205 206 // Walk backwards through the block, looking for dependencies 207 while (ScanIt != BB->begin()) { 208 // Limit the amount of scanning we do so we don't end up with quadratic 209 // running time on extreme testcases. 210 --Limit; 211 if (!Limit) 212 return MemDepResult::getUnknown(); 213 214 Instruction *Inst = --ScanIt; 215 216 // If this inst is a memory op, get the pointer it accessed 217 AliasAnalysis::Location Loc; 218 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA); 219 if (Loc.Ptr) { 220 // A simple instruction. 221 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef) 222 return MemDepResult::getClobber(Inst); 223 continue; 224 } 225 226 if (auto InstCS = CallSite(Inst)) { 227 // Debug intrinsics don't cause dependences. 228 if (isa<DbgInfoIntrinsic>(Inst)) continue; 229 // If these two calls do not interfere, look past it. 230 switch (AA->getModRefInfo(CS, InstCS)) { 231 case AliasAnalysis::NoModRef: 232 // If the two calls are the same, return InstCS as a Def, so that 233 // CS can be found redundant and eliminated. 234 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) && 235 CS.getInstruction()->isIdenticalToWhenDefined(Inst)) 236 return MemDepResult::getDef(Inst); 237 238 // Otherwise if the two calls don't interact (e.g. InstCS is readnone) 239 // keep scanning. 240 continue; 241 default: 242 return MemDepResult::getClobber(Inst); 243 } 244 } 245 246 // If we could not obtain a pointer for the instruction and the instruction 247 // touches memory then assume that this is a dependency. 248 if (MR != AliasAnalysis::NoModRef) 249 return MemDepResult::getClobber(Inst); 250 } 251 252 // No dependence found. If this is the entry block of the function, it is 253 // unknown, otherwise it is non-local. 254 if (BB != &BB->getParent()->getEntryBlock()) 255 return MemDepResult::getNonLocal(); 256 return MemDepResult::getNonFuncLocal(); 257 } 258 259 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that 260 /// would fully overlap MemLoc if done as a wider legal integer load. 261 /// 262 /// MemLocBase, MemLocOffset are lazily computed here the first time the 263 /// base/offs of memloc is needed. 264 static bool isLoadLoadClobberIfExtendedToFullWidth( 265 const AliasAnalysis::Location &MemLoc, const Value *&MemLocBase, 266 int64_t &MemLocOffs, const LoadInst *LI) { 267 const DataLayout &DL = LI->getModule()->getDataLayout(); 268 269 // If we haven't already computed the base/offset of MemLoc, do so now. 270 if (!MemLocBase) 271 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, DL); 272 273 unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize( 274 MemLocBase, MemLocOffs, MemLoc.Size, LI); 275 return Size != 0; 276 } 277 278 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that 279 /// looks at a memory location for a load (specified by MemLocBase, Offs, 280 /// and Size) and compares it against a load. If the specified load could 281 /// be safely widened to a larger integer load that is 1) still efficient, 282 /// 2) safe for the target, and 3) would provide the specified memory 283 /// location value, then this function returns the size in bytes of the 284 /// load width to use. If not, this returns zero. 285 unsigned MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize( 286 const Value *MemLocBase, int64_t MemLocOffs, unsigned MemLocSize, 287 const LoadInst *LI) { 288 // We can only extend simple integer loads. 289 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0; 290 291 // Load widening is hostile to ThreadSanitizer: it may cause false positives 292 // or make the reports more cryptic (access sizes are wrong). 293 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread)) 294 return 0; 295 296 const DataLayout &DL = LI->getModule()->getDataLayout(); 297 298 // Get the base of this load. 299 int64_t LIOffs = 0; 300 const Value *LIBase = 301 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, DL); 302 303 // If the two pointers are not based on the same pointer, we can't tell that 304 // they are related. 305 if (LIBase != MemLocBase) return 0; 306 307 // Okay, the two values are based on the same pointer, but returned as 308 // no-alias. This happens when we have things like two byte loads at "P+1" 309 // and "P+3". Check to see if increasing the size of the "LI" load up to its 310 // alignment (or the largest native integer type) will allow us to load all 311 // the bits required by MemLoc. 312 313 // If MemLoc is before LI, then no widening of LI will help us out. 314 if (MemLocOffs < LIOffs) return 0; 315 316 // Get the alignment of the load in bytes. We assume that it is safe to load 317 // any legal integer up to this size without a problem. For example, if we're 318 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can 319 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it 320 // to i16. 321 unsigned LoadAlign = LI->getAlignment(); 322 323 int64_t MemLocEnd = MemLocOffs+MemLocSize; 324 325 // If no amount of rounding up will let MemLoc fit into LI, then bail out. 326 if (LIOffs+LoadAlign < MemLocEnd) return 0; 327 328 // This is the size of the load to try. Start with the next larger power of 329 // two. 330 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U; 331 NewLoadByteSize = NextPowerOf2(NewLoadByteSize); 332 333 while (1) { 334 // If this load size is bigger than our known alignment or would not fit 335 // into a native integer register, then we fail. 336 if (NewLoadByteSize > LoadAlign || 337 !DL.fitsInLegalInteger(NewLoadByteSize*8)) 338 return 0; 339 340 if (LIOffs + NewLoadByteSize > MemLocEnd && 341 LI->getParent()->getParent()->hasFnAttribute( 342 Attribute::SanitizeAddress)) 343 // We will be reading past the location accessed by the original program. 344 // While this is safe in a regular build, Address Safety analysis tools 345 // may start reporting false warnings. So, don't do widening. 346 return 0; 347 348 // If a load of this width would include all of MemLoc, then we succeed. 349 if (LIOffs+NewLoadByteSize >= MemLocEnd) 350 return NewLoadByteSize; 351 352 NewLoadByteSize <<= 1; 353 } 354 } 355 356 static bool isVolatile(Instruction *Inst) { 357 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) 358 return LI->isVolatile(); 359 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) 360 return SI->isVolatile(); 361 else if (AtomicCmpXchgInst *AI = dyn_cast<AtomicCmpXchgInst>(Inst)) 362 return AI->isVolatile(); 363 return false; 364 } 365 366 367 /// getPointerDependencyFrom - Return the instruction on which a memory 368 /// location depends. If isLoad is true, this routine ignores may-aliases with 369 /// read-only operations. If isLoad is false, this routine ignores may-aliases 370 /// with reads from read-only locations. If possible, pass the query 371 /// instruction as well; this function may take advantage of the metadata 372 /// annotated to the query instruction to refine the result. 373 MemDepResult MemoryDependenceAnalysis:: 374 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad, 375 BasicBlock::iterator ScanIt, BasicBlock *BB, 376 Instruction *QueryInst) { 377 378 const Value *MemLocBase = nullptr; 379 int64_t MemLocOffset = 0; 380 unsigned Limit = BlockScanLimit; 381 bool isInvariantLoad = false; 382 383 // We must be careful with atomic accesses, as they may allow another thread 384 // to touch this location, cloberring it. We are conservative: if the 385 // QueryInst is not a simple (non-atomic) memory access, we automatically 386 // return getClobber. 387 // If it is simple, we know based on the results of 388 // "Compiler testing via a theory of sound optimisations in the C11/C++11 389 // memory model" in PLDI 2013, that a non-atomic location can only be 390 // clobbered between a pair of a release and an acquire action, with no 391 // access to the location in between. 392 // Here is an example for giving the general intuition behind this rule. 393 // In the following code: 394 // store x 0; 395 // release action; [1] 396 // acquire action; [4] 397 // %val = load x; 398 // It is unsafe to replace %val by 0 because another thread may be running: 399 // acquire action; [2] 400 // store x 42; 401 // release action; [3] 402 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val 403 // being 42. A key property of this program however is that if either 404 // 1 or 4 were missing, there would be a race between the store of 42 405 // either the store of 0 or the load (making the whole progam racy). 406 // The paper mentionned above shows that the same property is respected 407 // by every program that can detect any optimisation of that kind: either 408 // it is racy (undefined) or there is a release followed by an acquire 409 // between the pair of accesses under consideration. 410 411 // If the load is invariant, we "know" that it doesn't alias *any* write. We 412 // do want to respect mustalias results since defs are useful for value 413 // forwarding, but any mayalias write can be assumed to be noalias. 414 // Arguably, this logic should be pushed inside AliasAnalysis itself. 415 if (isLoad && QueryInst) { 416 LoadInst *LI = dyn_cast<LoadInst>(QueryInst); 417 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr) 418 isInvariantLoad = true; 419 } 420 421 const DataLayout &DL = BB->getModule()->getDataLayout(); 422 423 // Walk backwards through the basic block, looking for dependencies. 424 while (ScanIt != BB->begin()) { 425 Instruction *Inst = --ScanIt; 426 427 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) 428 // Debug intrinsics don't (and can't) cause dependencies. 429 if (isa<DbgInfoIntrinsic>(II)) continue; 430 431 // Limit the amount of scanning we do so we don't end up with quadratic 432 // running time on extreme testcases. 433 --Limit; 434 if (!Limit) 435 return MemDepResult::getUnknown(); 436 437 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 438 // If we reach a lifetime begin or end marker, then the query ends here 439 // because the value is undefined. 440 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 441 // FIXME: This only considers queries directly on the invariant-tagged 442 // pointer, not on query pointers that are indexed off of them. It'd 443 // be nice to handle that at some point (the right approach is to use 444 // GetPointerBaseWithConstantOffset). 445 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)), 446 MemLoc)) 447 return MemDepResult::getDef(II); 448 continue; 449 } 450 } 451 452 // Values depend on loads if the pointers are must aliased. This means that 453 // a load depends on another must aliased load from the same value. 454 // One exception is atomic loads: a value can depend on an atomic load that it 455 // does not alias with when this atomic load indicates that another thread may 456 // be accessing the location. 457 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 458 459 // While volatile access cannot be eliminated, they do not have to clobber 460 // non-aliasing locations, as normal accesses, for example, can be safely 461 // reordered with volatile accesses. 462 if (LI->isVolatile()) { 463 if (!QueryInst) 464 // Original QueryInst *may* be volatile 465 return MemDepResult::getClobber(LI); 466 if (isVolatile(QueryInst)) 467 // Ordering required if QueryInst is itself volatile 468 return MemDepResult::getClobber(LI); 469 // Otherwise, volatile doesn't imply any special ordering 470 } 471 472 // Atomic loads have complications involved. 473 // A Monotonic (or higher) load is OK if the query inst is itself not atomic. 474 // FIXME: This is overly conservative. 475 if (LI->isAtomic() && LI->getOrdering() > Unordered) { 476 if (!QueryInst) 477 return MemDepResult::getClobber(LI); 478 if (LI->getOrdering() != Monotonic) 479 return MemDepResult::getClobber(LI); 480 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) { 481 if (!QueryLI->isSimple()) 482 return MemDepResult::getClobber(LI); 483 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) { 484 if (!QuerySI->isSimple()) 485 return MemDepResult::getClobber(LI); 486 } else if (QueryInst->mayReadOrWriteMemory()) { 487 return MemDepResult::getClobber(LI); 488 } 489 } 490 491 AliasAnalysis::Location LoadLoc = AA->getLocation(LI); 492 493 // If we found a pointer, check if it could be the same as our pointer. 494 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc); 495 496 if (isLoad) { 497 if (R == AliasAnalysis::NoAlias) { 498 // If this is an over-aligned integer load (for example, 499 // "load i8* %P, align 4") see if it would obviously overlap with the 500 // queried location if widened to a larger load (e.g. if the queried 501 // location is 1 byte at P+1). If so, return it as a load/load 502 // clobber result, allowing the client to decide to widen the load if 503 // it wants to. 504 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) { 505 if (LI->getAlignment() * 8 > ITy->getPrimitiveSizeInBits() && 506 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase, 507 MemLocOffset, LI)) 508 return MemDepResult::getClobber(Inst); 509 } 510 continue; 511 } 512 513 // Must aliased loads are defs of each other. 514 if (R == AliasAnalysis::MustAlias) 515 return MemDepResult::getDef(Inst); 516 517 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads 518 // in terms of clobbering loads, but since it does this by looking 519 // at the clobbering load directly, it doesn't know about any 520 // phi translation that may have happened along the way. 521 522 // If we have a partial alias, then return this as a clobber for the 523 // client to handle. 524 if (R == AliasAnalysis::PartialAlias) 525 return MemDepResult::getClobber(Inst); 526 #endif 527 528 // Random may-alias loads don't depend on each other without a 529 // dependence. 530 continue; 531 } 532 533 // Stores don't depend on other no-aliased accesses. 534 if (R == AliasAnalysis::NoAlias) 535 continue; 536 537 // Stores don't alias loads from read-only memory. 538 if (AA->pointsToConstantMemory(LoadLoc)) 539 continue; 540 541 // Stores depend on may/must aliased loads. 542 return MemDepResult::getDef(Inst); 543 } 544 545 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 546 // Atomic stores have complications involved. 547 // A Monotonic store is OK if the query inst is itself not atomic. 548 // FIXME: This is overly conservative. 549 if (!SI->isUnordered()) { 550 if (!QueryInst) 551 return MemDepResult::getClobber(SI); 552 if (SI->getOrdering() != Monotonic) 553 return MemDepResult::getClobber(SI); 554 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) { 555 if (!QueryLI->isSimple()) 556 return MemDepResult::getClobber(SI); 557 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) { 558 if (!QuerySI->isSimple()) 559 return MemDepResult::getClobber(SI); 560 } else if (QueryInst->mayReadOrWriteMemory()) { 561 return MemDepResult::getClobber(SI); 562 } 563 } 564 565 // FIXME: this is overly conservative. 566 // While volatile access cannot be eliminated, they do not have to clobber 567 // non-aliasing locations, as normal accesses can for example be reordered 568 // with volatile accesses. 569 if (SI->isVolatile()) 570 return MemDepResult::getClobber(SI); 571 572 // If alias analysis can tell that this store is guaranteed to not modify 573 // the query pointer, ignore it. Use getModRefInfo to handle cases where 574 // the query pointer points to constant memory etc. 575 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef) 576 continue; 577 578 // Ok, this store might clobber the query pointer. Check to see if it is 579 // a must alias: in this case, we want to return this as a def. 580 AliasAnalysis::Location StoreLoc = AA->getLocation(SI); 581 582 // If we found a pointer, check if it could be the same as our pointer. 583 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc); 584 585 if (R == AliasAnalysis::NoAlias) 586 continue; 587 if (R == AliasAnalysis::MustAlias) 588 return MemDepResult::getDef(Inst); 589 if (isInvariantLoad) 590 continue; 591 return MemDepResult::getClobber(Inst); 592 } 593 594 // If this is an allocation, and if we know that the accessed pointer is to 595 // the allocation, return Def. This means that there is no dependence and 596 // the access can be optimized based on that. For example, a load could 597 // turn into undef. 598 // Note: Only determine this to be a malloc if Inst is the malloc call, not 599 // a subsequent bitcast of the malloc call result. There can be stores to 600 // the malloced memory between the malloc call and its bitcast uses, and we 601 // need to continue scanning until the malloc call. 602 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo(); 603 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) { 604 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL); 605 606 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr)) 607 return MemDepResult::getDef(Inst); 608 if (isInvariantLoad) 609 continue; 610 // Be conservative if the accessed pointer may alias the allocation. 611 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias) 612 return MemDepResult::getClobber(Inst); 613 // If the allocation is not aliased and does not read memory (like 614 // strdup), it is safe to ignore. 615 if (isa<AllocaInst>(Inst) || 616 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI)) 617 continue; 618 } 619 620 if (isInvariantLoad) 621 continue; 622 623 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 624 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc); 625 // If necessary, perform additional analysis. 626 if (MR == AliasAnalysis::ModRef) 627 MR = AA->callCapturesBefore(Inst, MemLoc, DT); 628 switch (MR) { 629 case AliasAnalysis::NoModRef: 630 // If the call has no effect on the queried pointer, just ignore it. 631 continue; 632 case AliasAnalysis::Mod: 633 return MemDepResult::getClobber(Inst); 634 case AliasAnalysis::Ref: 635 // If the call is known to never store to the pointer, and if this is a 636 // load query, we can safely ignore it (scan past it). 637 if (isLoad) 638 continue; 639 default: 640 // Otherwise, there is a potential dependence. Return a clobber. 641 return MemDepResult::getClobber(Inst); 642 } 643 } 644 645 // No dependence found. If this is the entry block of the function, it is 646 // unknown, otherwise it is non-local. 647 if (BB != &BB->getParent()->getEntryBlock()) 648 return MemDepResult::getNonLocal(); 649 return MemDepResult::getNonFuncLocal(); 650 } 651 652 /// getDependency - Return the instruction on which a memory operation 653 /// depends. 654 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 655 Instruction *ScanPos = QueryInst; 656 657 // Check for a cached result 658 MemDepResult &LocalCache = LocalDeps[QueryInst]; 659 660 // If the cached entry is non-dirty, just return it. Note that this depends 661 // on MemDepResult's default constructing to 'dirty'. 662 if (!LocalCache.isDirty()) 663 return LocalCache; 664 665 // Otherwise, if we have a dirty entry, we know we can start the scan at that 666 // instruction, which may save us some work. 667 if (Instruction *Inst = LocalCache.getInst()) { 668 ScanPos = Inst; 669 670 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 671 } 672 673 BasicBlock *QueryParent = QueryInst->getParent(); 674 675 // Do the scan. 676 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 677 // No dependence found. If this is the entry block of the function, it is 678 // unknown, otherwise it is non-local. 679 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 680 LocalCache = MemDepResult::getNonLocal(); 681 else 682 LocalCache = MemDepResult::getNonFuncLocal(); 683 } else { 684 AliasAnalysis::Location MemLoc; 685 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA); 686 if (MemLoc.Ptr) { 687 // If we can do a pointer scan, make it happen. 688 bool isLoad = !(MR & AliasAnalysis::Mod); 689 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst)) 690 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start; 691 692 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos, 693 QueryParent, QueryInst); 694 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 695 CallSite QueryCS(QueryInst); 696 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 697 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 698 QueryParent); 699 } else 700 // Non-memory instruction. 701 LocalCache = MemDepResult::getUnknown(); 702 } 703 704 // Remember the result! 705 if (Instruction *I = LocalCache.getInst()) 706 ReverseLocalDeps[I].insert(QueryInst); 707 708 return LocalCache; 709 } 710 711 #ifndef NDEBUG 712 /// AssertSorted - This method is used when -debug is specified to verify that 713 /// cache arrays are properly kept sorted. 714 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 715 int Count = -1) { 716 if (Count == -1) Count = Cache.size(); 717 if (Count == 0) return; 718 719 for (unsigned i = 1; i != unsigned(Count); ++i) 720 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 721 } 722 #endif 723 724 /// getNonLocalCallDependency - Perform a full dependency query for the 725 /// specified call, returning the set of blocks that the value is 726 /// potentially live across. The returned set of results will include a 727 /// "NonLocal" result for all blocks where the value is live across. 728 /// 729 /// This method assumes the instruction returns a "NonLocal" dependency 730 /// within its own block. 731 /// 732 /// This returns a reference to an internal data structure that may be 733 /// invalidated on the next non-local query or when an instruction is 734 /// removed. Clients must copy this data if they want it around longer than 735 /// that. 736 const MemoryDependenceAnalysis::NonLocalDepInfo & 737 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 738 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 739 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 740 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 741 NonLocalDepInfo &Cache = CacheP.first; 742 743 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 744 /// the cached case, this can happen due to instructions being deleted etc. In 745 /// the uncached case, this starts out as the set of predecessors we care 746 /// about. 747 SmallVector<BasicBlock*, 32> DirtyBlocks; 748 749 if (!Cache.empty()) { 750 // Okay, we have a cache entry. If we know it is not dirty, just return it 751 // with no computation. 752 if (!CacheP.second) { 753 ++NumCacheNonLocal; 754 return Cache; 755 } 756 757 // If we already have a partially computed set of results, scan them to 758 // determine what is dirty, seeding our initial DirtyBlocks worklist. 759 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 760 I != E; ++I) 761 if (I->getResult().isDirty()) 762 DirtyBlocks.push_back(I->getBB()); 763 764 // Sort the cache so that we can do fast binary search lookups below. 765 std::sort(Cache.begin(), Cache.end()); 766 767 ++NumCacheDirtyNonLocal; 768 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 769 // << Cache.size() << " cached: " << *QueryInst; 770 } else { 771 // Seed DirtyBlocks with each of the preds of QueryInst's block. 772 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 773 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 774 DirtyBlocks.push_back(*PI); 775 ++NumUncacheNonLocal; 776 } 777 778 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 779 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 780 781 SmallPtrSet<BasicBlock*, 64> Visited; 782 783 unsigned NumSortedEntries = Cache.size(); 784 DEBUG(AssertSorted(Cache)); 785 786 // Iterate while we still have blocks to update. 787 while (!DirtyBlocks.empty()) { 788 BasicBlock *DirtyBB = DirtyBlocks.back(); 789 DirtyBlocks.pop_back(); 790 791 // Already processed this block? 792 if (!Visited.insert(DirtyBB).second) 793 continue; 794 795 // Do a binary search to see if we already have an entry for this block in 796 // the cache set. If so, find it. 797 DEBUG(AssertSorted(Cache, NumSortedEntries)); 798 NonLocalDepInfo::iterator Entry = 799 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 800 NonLocalDepEntry(DirtyBB)); 801 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB) 802 --Entry; 803 804 NonLocalDepEntry *ExistingResult = nullptr; 805 if (Entry != Cache.begin()+NumSortedEntries && 806 Entry->getBB() == DirtyBB) { 807 // If we already have an entry, and if it isn't already dirty, the block 808 // is done. 809 if (!Entry->getResult().isDirty()) 810 continue; 811 812 // Otherwise, remember this slot so we can update the value. 813 ExistingResult = &*Entry; 814 } 815 816 // If the dirty entry has a pointer, start scanning from it so we don't have 817 // to rescan the entire block. 818 BasicBlock::iterator ScanPos = DirtyBB->end(); 819 if (ExistingResult) { 820 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 821 ScanPos = Inst; 822 // We're removing QueryInst's use of Inst. 823 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 824 QueryCS.getInstruction()); 825 } 826 } 827 828 // Find out if this block has a local dependency for QueryInst. 829 MemDepResult Dep; 830 831 if (ScanPos != DirtyBB->begin()) { 832 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 833 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 834 // No dependence found. If this is the entry block of the function, it is 835 // a clobber, otherwise it is unknown. 836 Dep = MemDepResult::getNonLocal(); 837 } else { 838 Dep = MemDepResult::getNonFuncLocal(); 839 } 840 841 // If we had a dirty entry for the block, update it. Otherwise, just add 842 // a new entry. 843 if (ExistingResult) 844 ExistingResult->setResult(Dep); 845 else 846 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 847 848 // If the block has a dependency (i.e. it isn't completely transparent to 849 // the value), remember the association! 850 if (!Dep.isNonLocal()) { 851 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 852 // update this when we remove instructions. 853 if (Instruction *Inst = Dep.getInst()) 854 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 855 } else { 856 857 // If the block *is* completely transparent to the load, we need to check 858 // the predecessors of this block. Add them to our worklist. 859 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 860 DirtyBlocks.push_back(*PI); 861 } 862 } 863 864 return Cache; 865 } 866 867 /// getNonLocalPointerDependency - Perform a full dependency query for an 868 /// access to the specified (non-volatile) memory location, returning the 869 /// set of instructions that either define or clobber the value. 870 /// 871 /// This method assumes the pointer has a "NonLocal" dependency within its 872 /// own block. 873 /// 874 void MemoryDependenceAnalysis:: 875 getNonLocalPointerDependency(Instruction *QueryInst, 876 SmallVectorImpl<NonLocalDepResult> &Result) { 877 const AliasAnalysis::Location Loc = AA->getLocation(QueryInst); 878 bool isLoad = isa<LoadInst>(QueryInst); 879 BasicBlock *FromBB = QueryInst->getParent(); 880 assert(FromBB); 881 882 assert(Loc.Ptr->getType()->isPointerTy() && 883 "Can't get pointer deps of a non-pointer!"); 884 Result.clear(); 885 886 // This routine does not expect to deal with volatile instructions. 887 // Doing so would require piping through the QueryInst all the way through. 888 // TODO: volatiles can't be elided, but they can be reordered with other 889 // non-volatile accesses. 890 891 // We currently give up on any instruction which is ordered, but we do handle 892 // atomic instructions which are unordered. 893 // TODO: Handle ordered instructions 894 auto isOrdered = [](Instruction *Inst) { 895 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 896 return !LI->isUnordered(); 897 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 898 return !SI->isUnordered(); 899 } 900 return false; 901 }; 902 if (isVolatile(QueryInst) || isOrdered(QueryInst)) { 903 Result.push_back(NonLocalDepResult(FromBB, 904 MemDepResult::getUnknown(), 905 const_cast<Value *>(Loc.Ptr))); 906 return; 907 } 908 const DataLayout &DL = FromBB->getModule()->getDataLayout(); 909 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, AC); 910 911 // This is the set of blocks we've inspected, and the pointer we consider in 912 // each block. Because of critical edges, we currently bail out if querying 913 // a block with multiple different pointers. This can happen during PHI 914 // translation. 915 DenseMap<BasicBlock*, Value*> Visited; 916 if (!getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB, 917 Result, Visited, true)) 918 return; 919 Result.clear(); 920 Result.push_back(NonLocalDepResult(FromBB, 921 MemDepResult::getUnknown(), 922 const_cast<Value *>(Loc.Ptr))); 923 } 924 925 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with 926 /// Pointer/PointeeSize using either cached information in Cache or by doing a 927 /// lookup (which may use dirty cache info if available). If we do a lookup, 928 /// add the result to the cache. 929 MemDepResult MemoryDependenceAnalysis:: 930 GetNonLocalInfoForBlock(Instruction *QueryInst, 931 const AliasAnalysis::Location &Loc, 932 bool isLoad, BasicBlock *BB, 933 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 934 935 // Do a binary search to see if we already have an entry for this block in 936 // the cache set. If so, find it. 937 NonLocalDepInfo::iterator Entry = 938 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 939 NonLocalDepEntry(BB)); 940 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 941 --Entry; 942 943 NonLocalDepEntry *ExistingResult = nullptr; 944 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 945 ExistingResult = &*Entry; 946 947 // If we have a cached entry, and it is non-dirty, use it as the value for 948 // this dependency. 949 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 950 ++NumCacheNonLocalPtr; 951 return ExistingResult->getResult(); 952 } 953 954 // Otherwise, we have to scan for the value. If we have a dirty cache 955 // entry, start scanning from its position, otherwise we scan from the end 956 // of the block. 957 BasicBlock::iterator ScanPos = BB->end(); 958 if (ExistingResult && ExistingResult->getResult().getInst()) { 959 assert(ExistingResult->getResult().getInst()->getParent() == BB && 960 "Instruction invalidated?"); 961 ++NumCacheDirtyNonLocalPtr; 962 ScanPos = ExistingResult->getResult().getInst(); 963 964 // Eliminating the dirty entry from 'Cache', so update the reverse info. 965 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 966 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 967 } else { 968 ++NumUncacheNonLocalPtr; 969 } 970 971 // Scan the block for the dependency. 972 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB, 973 QueryInst); 974 975 // If we had a dirty entry for the block, update it. Otherwise, just add 976 // a new entry. 977 if (ExistingResult) 978 ExistingResult->setResult(Dep); 979 else 980 Cache->push_back(NonLocalDepEntry(BB, Dep)); 981 982 // If the block has a dependency (i.e. it isn't completely transparent to 983 // the value), remember the reverse association because we just added it 984 // to Cache! 985 if (!Dep.isDef() && !Dep.isClobber()) 986 return Dep; 987 988 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 989 // update MemDep when we remove instructions. 990 Instruction *Inst = Dep.getInst(); 991 assert(Inst && "Didn't depend on anything?"); 992 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 993 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 994 return Dep; 995 } 996 997 /// SortNonLocalDepInfoCache - Sort the NonLocalDepInfo cache, given a certain 998 /// number of elements in the array that are already properly ordered. This is 999 /// optimized for the case when only a few entries are added. 1000 static void 1001 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 1002 unsigned NumSortedEntries) { 1003 switch (Cache.size() - NumSortedEntries) { 1004 case 0: 1005 // done, no new entries. 1006 break; 1007 case 2: { 1008 // Two new entries, insert the last one into place. 1009 NonLocalDepEntry Val = Cache.back(); 1010 Cache.pop_back(); 1011 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 1012 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 1013 Cache.insert(Entry, Val); 1014 // FALL THROUGH. 1015 } 1016 case 1: 1017 // One new entry, Just insert the new value at the appropriate position. 1018 if (Cache.size() != 1) { 1019 NonLocalDepEntry Val = Cache.back(); 1020 Cache.pop_back(); 1021 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 1022 std::upper_bound(Cache.begin(), Cache.end(), Val); 1023 Cache.insert(Entry, Val); 1024 } 1025 break; 1026 default: 1027 // Added many values, do a full scale sort. 1028 std::sort(Cache.begin(), Cache.end()); 1029 break; 1030 } 1031 } 1032 1033 /// getNonLocalPointerDepFromBB - Perform a dependency query based on 1034 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 1035 /// results to the results vector and keep track of which blocks are visited in 1036 /// 'Visited'. 1037 /// 1038 /// This has special behavior for the first block queries (when SkipFirstBlock 1039 /// is true). In this special case, it ignores the contents of the specified 1040 /// block and starts returning dependence info for its predecessors. 1041 /// 1042 /// This function returns false on success, or true to indicate that it could 1043 /// not compute dependence information for some reason. This should be treated 1044 /// as a clobber dependence on the first instruction in the predecessor block. 1045 bool MemoryDependenceAnalysis:: 1046 getNonLocalPointerDepFromBB(Instruction *QueryInst, 1047 const PHITransAddr &Pointer, 1048 const AliasAnalysis::Location &Loc, 1049 bool isLoad, BasicBlock *StartBB, 1050 SmallVectorImpl<NonLocalDepResult> &Result, 1051 DenseMap<BasicBlock*, Value*> &Visited, 1052 bool SkipFirstBlock) { 1053 // Look up the cached info for Pointer. 1054 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 1055 1056 // Set up a temporary NLPI value. If the map doesn't yet have an entry for 1057 // CacheKey, this value will be inserted as the associated value. Otherwise, 1058 // it'll be ignored, and we'll have to check to see if the cached size and 1059 // aa tags are consistent with the current query. 1060 NonLocalPointerInfo InitialNLPI; 1061 InitialNLPI.Size = Loc.Size; 1062 InitialNLPI.AATags = Loc.AATags; 1063 1064 // Get the NLPI for CacheKey, inserting one into the map if it doesn't 1065 // already have one. 1066 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair = 1067 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI)); 1068 NonLocalPointerInfo *CacheInfo = &Pair.first->second; 1069 1070 // If we already have a cache entry for this CacheKey, we may need to do some 1071 // work to reconcile the cache entry and the current query. 1072 if (!Pair.second) { 1073 if (CacheInfo->Size < Loc.Size) { 1074 // The query's Size is greater than the cached one. Throw out the 1075 // cached data and proceed with the query at the greater size. 1076 CacheInfo->Pair = BBSkipFirstBlockPair(); 1077 CacheInfo->Size = Loc.Size; 1078 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 1079 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 1080 if (Instruction *Inst = DI->getResult().getInst()) 1081 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 1082 CacheInfo->NonLocalDeps.clear(); 1083 } else if (CacheInfo->Size > Loc.Size) { 1084 // This query's Size is less than the cached one. Conservatively restart 1085 // the query using the greater size. 1086 return getNonLocalPointerDepFromBB(QueryInst, Pointer, 1087 Loc.getWithNewSize(CacheInfo->Size), 1088 isLoad, StartBB, Result, Visited, 1089 SkipFirstBlock); 1090 } 1091 1092 // If the query's AATags are inconsistent with the cached one, 1093 // conservatively throw out the cached data and restart the query with 1094 // no tag if needed. 1095 if (CacheInfo->AATags != Loc.AATags) { 1096 if (CacheInfo->AATags) { 1097 CacheInfo->Pair = BBSkipFirstBlockPair(); 1098 CacheInfo->AATags = AAMDNodes(); 1099 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 1100 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 1101 if (Instruction *Inst = DI->getResult().getInst()) 1102 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 1103 CacheInfo->NonLocalDeps.clear(); 1104 } 1105 if (Loc.AATags) 1106 return getNonLocalPointerDepFromBB(QueryInst, 1107 Pointer, Loc.getWithoutAATags(), 1108 isLoad, StartBB, Result, Visited, 1109 SkipFirstBlock); 1110 } 1111 } 1112 1113 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps; 1114 1115 // If we have valid cached information for exactly the block we are 1116 // investigating, just return it with no recomputation. 1117 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 1118 // We have a fully cached result for this query then we can just return the 1119 // cached results and populate the visited set. However, we have to verify 1120 // that we don't already have conflicting results for these blocks. Check 1121 // to ensure that if a block in the results set is in the visited set that 1122 // it was for the same pointer query. 1123 if (!Visited.empty()) { 1124 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 1125 I != E; ++I) { 1126 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 1127 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 1128 continue; 1129 1130 // We have a pointer mismatch in a block. Just return clobber, saying 1131 // that something was clobbered in this result. We could also do a 1132 // non-fully cached query, but there is little point in doing this. 1133 return true; 1134 } 1135 } 1136 1137 Value *Addr = Pointer.getAddr(); 1138 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 1139 I != E; ++I) { 1140 Visited.insert(std::make_pair(I->getBB(), Addr)); 1141 if (I->getResult().isNonLocal()) { 1142 continue; 1143 } 1144 1145 if (!DT) { 1146 Result.push_back(NonLocalDepResult(I->getBB(), 1147 MemDepResult::getUnknown(), 1148 Addr)); 1149 } else if (DT->isReachableFromEntry(I->getBB())) { 1150 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 1151 } 1152 } 1153 ++NumCacheCompleteNonLocalPtr; 1154 return false; 1155 } 1156 1157 // Otherwise, either this is a new block, a block with an invalid cache 1158 // pointer or one that we're about to invalidate by putting more info into it 1159 // than its valid cache info. If empty, the result will be valid cache info, 1160 // otherwise it isn't. 1161 if (Cache->empty()) 1162 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 1163 else 1164 CacheInfo->Pair = BBSkipFirstBlockPair(); 1165 1166 SmallVector<BasicBlock*, 32> Worklist; 1167 Worklist.push_back(StartBB); 1168 1169 // PredList used inside loop. 1170 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList; 1171 1172 // Keep track of the entries that we know are sorted. Previously cached 1173 // entries will all be sorted. The entries we add we only sort on demand (we 1174 // don't insert every element into its sorted position). We know that we 1175 // won't get any reuse from currently inserted values, because we don't 1176 // revisit blocks after we insert info for them. 1177 unsigned NumSortedEntries = Cache->size(); 1178 DEBUG(AssertSorted(*Cache)); 1179 1180 while (!Worklist.empty()) { 1181 BasicBlock *BB = Worklist.pop_back_val(); 1182 1183 // If we do process a large number of blocks it becomes very expensive and 1184 // likely it isn't worth worrying about 1185 if (Result.size() > NumResultsLimit) { 1186 Worklist.clear(); 1187 // Sort it now (if needed) so that recursive invocations of 1188 // getNonLocalPointerDepFromBB and other routines that could reuse the 1189 // cache value will only see properly sorted cache arrays. 1190 if (Cache && NumSortedEntries != Cache->size()) { 1191 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1192 } 1193 // Since we bail out, the "Cache" set won't contain all of the 1194 // results for the query. This is ok (we can still use it to accelerate 1195 // specific block queries) but we can't do the fastpath "return all 1196 // results from the set". Clear out the indicator for this. 1197 CacheInfo->Pair = BBSkipFirstBlockPair(); 1198 return true; 1199 } 1200 1201 // Skip the first block if we have it. 1202 if (!SkipFirstBlock) { 1203 // Analyze the dependency of *Pointer in FromBB. See if we already have 1204 // been here. 1205 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 1206 1207 // Get the dependency info for Pointer in BB. If we have cached 1208 // information, we will use it, otherwise we compute it. 1209 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 1210 MemDepResult Dep = GetNonLocalInfoForBlock(QueryInst, 1211 Loc, isLoad, BB, Cache, 1212 NumSortedEntries); 1213 1214 // If we got a Def or Clobber, add this to the list of results. 1215 if (!Dep.isNonLocal()) { 1216 if (!DT) { 1217 Result.push_back(NonLocalDepResult(BB, 1218 MemDepResult::getUnknown(), 1219 Pointer.getAddr())); 1220 continue; 1221 } else if (DT->isReachableFromEntry(BB)) { 1222 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 1223 continue; 1224 } 1225 } 1226 } 1227 1228 // If 'Pointer' is an instruction defined in this block, then we need to do 1229 // phi translation to change it into a value live in the predecessor block. 1230 // If not, we just add the predecessors to the worklist and scan them with 1231 // the same Pointer. 1232 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 1233 SkipFirstBlock = false; 1234 SmallVector<BasicBlock*, 16> NewBlocks; 1235 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1236 // Verify that we haven't looked at this block yet. 1237 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1238 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 1239 if (InsertRes.second) { 1240 // First time we've looked at *PI. 1241 NewBlocks.push_back(*PI); 1242 continue; 1243 } 1244 1245 // If we have seen this block before, but it was with a different 1246 // pointer then we have a phi translation failure and we have to treat 1247 // this as a clobber. 1248 if (InsertRes.first->second != Pointer.getAddr()) { 1249 // Make sure to clean up the Visited map before continuing on to 1250 // PredTranslationFailure. 1251 for (unsigned i = 0; i < NewBlocks.size(); i++) 1252 Visited.erase(NewBlocks[i]); 1253 goto PredTranslationFailure; 1254 } 1255 } 1256 Worklist.append(NewBlocks.begin(), NewBlocks.end()); 1257 continue; 1258 } 1259 1260 // We do need to do phi translation, if we know ahead of time we can't phi 1261 // translate this value, don't even try. 1262 if (!Pointer.IsPotentiallyPHITranslatable()) 1263 goto PredTranslationFailure; 1264 1265 // We may have added values to the cache list before this PHI translation. 1266 // If so, we haven't done anything to ensure that the cache remains sorted. 1267 // Sort it now (if needed) so that recursive invocations of 1268 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 1269 // value will only see properly sorted cache arrays. 1270 if (Cache && NumSortedEntries != Cache->size()) { 1271 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1272 NumSortedEntries = Cache->size(); 1273 } 1274 Cache = nullptr; 1275 1276 PredList.clear(); 1277 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1278 BasicBlock *Pred = *PI; 1279 PredList.push_back(std::make_pair(Pred, Pointer)); 1280 1281 // Get the PHI translated pointer in this predecessor. This can fail if 1282 // not translatable, in which case the getAddr() returns null. 1283 PHITransAddr &PredPointer = PredList.back().second; 1284 PredPointer.PHITranslateValue(BB, Pred, nullptr); 1285 1286 Value *PredPtrVal = PredPointer.getAddr(); 1287 1288 // Check to see if we have already visited this pred block with another 1289 // pointer. If so, we can't do this lookup. This failure can occur 1290 // with PHI translation when a critical edge exists and the PHI node in 1291 // the successor translates to a pointer value different than the 1292 // pointer the block was first analyzed with. 1293 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1294 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 1295 1296 if (!InsertRes.second) { 1297 // We found the pred; take it off the list of preds to visit. 1298 PredList.pop_back(); 1299 1300 // If the predecessor was visited with PredPtr, then we already did 1301 // the analysis and can ignore it. 1302 if (InsertRes.first->second == PredPtrVal) 1303 continue; 1304 1305 // Otherwise, the block was previously analyzed with a different 1306 // pointer. We can't represent the result of this case, so we just 1307 // treat this as a phi translation failure. 1308 1309 // Make sure to clean up the Visited map before continuing on to 1310 // PredTranslationFailure. 1311 for (unsigned i = 0, n = PredList.size(); i < n; ++i) 1312 Visited.erase(PredList[i].first); 1313 1314 goto PredTranslationFailure; 1315 } 1316 } 1317 1318 // Actually process results here; this need to be a separate loop to avoid 1319 // calling getNonLocalPointerDepFromBB for blocks we don't want to return 1320 // any results for. (getNonLocalPointerDepFromBB will modify our 1321 // datastructures in ways the code after the PredTranslationFailure label 1322 // doesn't expect.) 1323 for (unsigned i = 0, n = PredList.size(); i < n; ++i) { 1324 BasicBlock *Pred = PredList[i].first; 1325 PHITransAddr &PredPointer = PredList[i].second; 1326 Value *PredPtrVal = PredPointer.getAddr(); 1327 1328 bool CanTranslate = true; 1329 // If PHI translation was unable to find an available pointer in this 1330 // predecessor, then we have to assume that the pointer is clobbered in 1331 // that predecessor. We can still do PRE of the load, which would insert 1332 // a computation of the pointer in this predecessor. 1333 if (!PredPtrVal) 1334 CanTranslate = false; 1335 1336 // FIXME: it is entirely possible that PHI translating will end up with 1337 // the same value. Consider PHI translating something like: 1338 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 1339 // to recurse here, pedantically speaking. 1340 1341 // If getNonLocalPointerDepFromBB fails here, that means the cached 1342 // result conflicted with the Visited list; we have to conservatively 1343 // assume it is unknown, but this also does not block PRE of the load. 1344 if (!CanTranslate || 1345 getNonLocalPointerDepFromBB(QueryInst, PredPointer, 1346 Loc.getWithNewPtr(PredPtrVal), 1347 isLoad, Pred, 1348 Result, Visited)) { 1349 // Add the entry to the Result list. 1350 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal); 1351 Result.push_back(Entry); 1352 1353 // Since we had a phi translation failure, the cache for CacheKey won't 1354 // include all of the entries that we need to immediately satisfy future 1355 // queries. Mark this in NonLocalPointerDeps by setting the 1356 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 1357 // cached value to do more work but not miss the phi trans failure. 1358 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey]; 1359 NLPI.Pair = BBSkipFirstBlockPair(); 1360 continue; 1361 } 1362 } 1363 1364 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 1365 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1366 Cache = &CacheInfo->NonLocalDeps; 1367 NumSortedEntries = Cache->size(); 1368 1369 // Since we did phi translation, the "Cache" set won't contain all of the 1370 // results for the query. This is ok (we can still use it to accelerate 1371 // specific block queries) but we can't do the fastpath "return all 1372 // results from the set" Clear out the indicator for this. 1373 CacheInfo->Pair = BBSkipFirstBlockPair(); 1374 SkipFirstBlock = false; 1375 continue; 1376 1377 PredTranslationFailure: 1378 // The following code is "failure"; we can't produce a sane translation 1379 // for the given block. It assumes that we haven't modified any of 1380 // our datastructures while processing the current block. 1381 1382 if (!Cache) { 1383 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1384 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1385 Cache = &CacheInfo->NonLocalDeps; 1386 NumSortedEntries = Cache->size(); 1387 } 1388 1389 // Since we failed phi translation, the "Cache" set won't contain all of the 1390 // results for the query. This is ok (we can still use it to accelerate 1391 // specific block queries) but we can't do the fastpath "return all 1392 // results from the set". Clear out the indicator for this. 1393 CacheInfo->Pair = BBSkipFirstBlockPair(); 1394 1395 // If *nothing* works, mark the pointer as unknown. 1396 // 1397 // If this is the magic first block, return this as a clobber of the whole 1398 // incoming value. Since we can't phi translate to one of the predecessors, 1399 // we have to bail out. 1400 if (SkipFirstBlock) 1401 return true; 1402 1403 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 1404 assert(I != Cache->rend() && "Didn't find current block??"); 1405 if (I->getBB() != BB) 1406 continue; 1407 1408 assert((I->getResult().isNonLocal() || !DT->isReachableFromEntry(BB)) && 1409 "Should only be here with transparent block"); 1410 I->setResult(MemDepResult::getUnknown()); 1411 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 1412 Pointer.getAddr())); 1413 break; 1414 } 1415 } 1416 1417 // Okay, we're done now. If we added new values to the cache, re-sort it. 1418 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1419 DEBUG(AssertSorted(*Cache)); 1420 return false; 1421 } 1422 1423 /// RemoveCachedNonLocalPointerDependencies - If P exists in 1424 /// CachedNonLocalPointerInfo, remove it. 1425 void MemoryDependenceAnalysis:: 1426 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 1427 CachedNonLocalPointerInfo::iterator It = 1428 NonLocalPointerDeps.find(P); 1429 if (It == NonLocalPointerDeps.end()) return; 1430 1431 // Remove all of the entries in the BB->val map. This involves removing 1432 // instructions from the reverse map. 1433 NonLocalDepInfo &PInfo = It->second.NonLocalDeps; 1434 1435 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 1436 Instruction *Target = PInfo[i].getResult().getInst(); 1437 if (!Target) continue; // Ignore non-local dep results. 1438 assert(Target->getParent() == PInfo[i].getBB()); 1439 1440 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1441 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1442 } 1443 1444 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1445 NonLocalPointerDeps.erase(It); 1446 } 1447 1448 1449 /// invalidateCachedPointerInfo - This method is used to invalidate cached 1450 /// information about the specified pointer, because it may be too 1451 /// conservative in memdep. This is an optional call that can be used when 1452 /// the client detects an equivalence between the pointer and some other 1453 /// value and replaces the other value with ptr. This can make Ptr available 1454 /// in more places that cached info does not necessarily keep. 1455 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1456 // If Ptr isn't really a pointer, just ignore it. 1457 if (!Ptr->getType()->isPointerTy()) return; 1458 // Flush store info for the pointer. 1459 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1460 // Flush load info for the pointer. 1461 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1462 } 1463 1464 /// invalidateCachedPredecessors - Clear the PredIteratorCache info. 1465 /// This needs to be done when the CFG changes, e.g., due to splitting 1466 /// critical edges. 1467 void MemoryDependenceAnalysis::invalidateCachedPredecessors() { 1468 PredCache->clear(); 1469 } 1470 1471 /// removeInstruction - Remove an instruction from the dependence analysis, 1472 /// updating the dependence of instructions that previously depended on it. 1473 /// This method attempts to keep the cache coherent using the reverse map. 1474 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1475 // Walk through the Non-local dependencies, removing this one as the value 1476 // for any cached queries. 1477 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1478 if (NLDI != NonLocalDeps.end()) { 1479 NonLocalDepInfo &BlockMap = NLDI->second.first; 1480 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1481 DI != DE; ++DI) 1482 if (Instruction *Inst = DI->getResult().getInst()) 1483 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1484 NonLocalDeps.erase(NLDI); 1485 } 1486 1487 // If we have a cached local dependence query for this instruction, remove it. 1488 // 1489 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1490 if (LocalDepEntry != LocalDeps.end()) { 1491 // Remove us from DepInst's reverse set now that the local dep info is gone. 1492 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1493 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1494 1495 // Remove this local dependency info. 1496 LocalDeps.erase(LocalDepEntry); 1497 } 1498 1499 // If we have any cached pointer dependencies on this instruction, remove 1500 // them. If the instruction has non-pointer type, then it can't be a pointer 1501 // base. 1502 1503 // Remove it from both the load info and the store info. The instruction 1504 // can't be in either of these maps if it is non-pointer. 1505 if (RemInst->getType()->isPointerTy()) { 1506 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1507 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1508 } 1509 1510 // Loop over all of the things that depend on the instruction we're removing. 1511 // 1512 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1513 1514 // If we find RemInst as a clobber or Def in any of the maps for other values, 1515 // we need to replace its entry with a dirty version of the instruction after 1516 // it. If RemInst is a terminator, we use a null dirty value. 1517 // 1518 // Using a dirty version of the instruction after RemInst saves having to scan 1519 // the entire block to get to this point. 1520 MemDepResult NewDirtyVal; 1521 if (!RemInst->isTerminator()) 1522 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1523 1524 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1525 if (ReverseDepIt != ReverseLocalDeps.end()) { 1526 // RemInst can't be the terminator if it has local stuff depending on it. 1527 assert(!ReverseDepIt->second.empty() && !isa<TerminatorInst>(RemInst) && 1528 "Nothing can locally depend on a terminator"); 1529 1530 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) { 1531 assert(InstDependingOnRemInst != RemInst && 1532 "Already removed our local dep info"); 1533 1534 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1535 1536 // Make sure to remember that new things depend on NewDepInst. 1537 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1538 "a local dep on this if it is a terminator!"); 1539 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1540 InstDependingOnRemInst)); 1541 } 1542 1543 ReverseLocalDeps.erase(ReverseDepIt); 1544 1545 // Add new reverse deps after scanning the set, to avoid invalidating the 1546 // 'ReverseDeps' reference. 1547 while (!ReverseDepsToAdd.empty()) { 1548 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1549 .insert(ReverseDepsToAdd.back().second); 1550 ReverseDepsToAdd.pop_back(); 1551 } 1552 } 1553 1554 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1555 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1556 for (Instruction *I : ReverseDepIt->second) { 1557 assert(I != RemInst && "Already removed NonLocalDep info for RemInst"); 1558 1559 PerInstNLInfo &INLD = NonLocalDeps[I]; 1560 // The information is now dirty! 1561 INLD.second = true; 1562 1563 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1564 DE = INLD.first.end(); DI != DE; ++DI) { 1565 if (DI->getResult().getInst() != RemInst) continue; 1566 1567 // Convert to a dirty entry for the subsequent instruction. 1568 DI->setResult(NewDirtyVal); 1569 1570 if (Instruction *NextI = NewDirtyVal.getInst()) 1571 ReverseDepsToAdd.push_back(std::make_pair(NextI, I)); 1572 } 1573 } 1574 1575 ReverseNonLocalDeps.erase(ReverseDepIt); 1576 1577 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1578 while (!ReverseDepsToAdd.empty()) { 1579 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1580 .insert(ReverseDepsToAdd.back().second); 1581 ReverseDepsToAdd.pop_back(); 1582 } 1583 } 1584 1585 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1586 // value in the NonLocalPointerDeps info. 1587 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1588 ReverseNonLocalPtrDeps.find(RemInst); 1589 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1590 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1591 1592 for (ValueIsLoadPair P : ReversePtrDepIt->second) { 1593 assert(P.getPointer() != RemInst && 1594 "Already removed NonLocalPointerDeps info for RemInst"); 1595 1596 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps; 1597 1598 // The cache is not valid for any specific block anymore. 1599 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair(); 1600 1601 // Update any entries for RemInst to use the instruction after it. 1602 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1603 DI != DE; ++DI) { 1604 if (DI->getResult().getInst() != RemInst) continue; 1605 1606 // Convert to a dirty entry for the subsequent instruction. 1607 DI->setResult(NewDirtyVal); 1608 1609 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1610 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1611 } 1612 1613 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1614 // subsequent value may invalidate the sortedness. 1615 std::sort(NLPDI.begin(), NLPDI.end()); 1616 } 1617 1618 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1619 1620 while (!ReversePtrDepsToAdd.empty()) { 1621 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1622 .insert(ReversePtrDepsToAdd.back().second); 1623 ReversePtrDepsToAdd.pop_back(); 1624 } 1625 } 1626 1627 1628 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1629 AA->deleteValue(RemInst); 1630 DEBUG(verifyRemoved(RemInst)); 1631 } 1632 /// verifyRemoved - Verify that the specified instruction does not occur 1633 /// in our internal data structures. This function verifies by asserting in 1634 /// debug builds. 1635 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1636 #ifndef NDEBUG 1637 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1638 E = LocalDeps.end(); I != E; ++I) { 1639 assert(I->first != D && "Inst occurs in data structures"); 1640 assert(I->second.getInst() != D && 1641 "Inst occurs in data structures"); 1642 } 1643 1644 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1645 E = NonLocalPointerDeps.end(); I != E; ++I) { 1646 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1647 const NonLocalDepInfo &Val = I->second.NonLocalDeps; 1648 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1649 II != E; ++II) 1650 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1651 } 1652 1653 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1654 E = NonLocalDeps.end(); I != E; ++I) { 1655 assert(I->first != D && "Inst occurs in data structures"); 1656 const PerInstNLInfo &INLD = I->second; 1657 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1658 EE = INLD.first.end(); II != EE; ++II) 1659 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1660 } 1661 1662 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1663 E = ReverseLocalDeps.end(); I != E; ++I) { 1664 assert(I->first != D && "Inst occurs in data structures"); 1665 for (Instruction *Inst : I->second) 1666 assert(Inst != D && "Inst occurs in data structures"); 1667 } 1668 1669 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1670 E = ReverseNonLocalDeps.end(); 1671 I != E; ++I) { 1672 assert(I->first != D && "Inst occurs in data structures"); 1673 for (Instruction *Inst : I->second) 1674 assert(Inst != D && "Inst occurs in data structures"); 1675 } 1676 1677 for (ReverseNonLocalPtrDepTy::const_iterator 1678 I = ReverseNonLocalPtrDeps.begin(), 1679 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1680 assert(I->first != D && "Inst occurs in rev NLPD map"); 1681 1682 for (ValueIsLoadPair P : I->second) 1683 assert(P != ValueIsLoadPair(D, false) && 1684 P != ValueIsLoadPair(D, true) && 1685 "Inst occurs in ReverseNonLocalPtrDeps map"); 1686 } 1687 #endif 1688 } 1689