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