MemoryDependenceAnalysis.cpp revision 9f47fb66370e5513bb9f737923e8cb476088acec
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/CaptureTracking.h" 26#include "llvm/Analysis/Dominators.h" 27#include "llvm/Analysis/InstructionSimplify.h" 28#include "llvm/Analysis/MemoryBuiltins.h" 29#include "llvm/Analysis/PHITransAddr.h" 30#include "llvm/Analysis/ValueTracking.h" 31#include "llvm/ADT/Statistic.h" 32#include "llvm/ADT/STLExtras.h" 33#include "llvm/Support/PredIteratorCache.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/Target/TargetData.h" 36using namespace llvm; 37 38STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 39STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 40STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 41 42STATISTIC(NumCacheNonLocalPtr, 43 "Number of fully cached non-local ptr responses"); 44STATISTIC(NumCacheDirtyNonLocalPtr, 45 "Number of cached, but dirty, non-local ptr responses"); 46STATISTIC(NumUncacheNonLocalPtr, 47 "Number of uncached non-local ptr responses"); 48STATISTIC(NumCacheCompleteNonLocalPtr, 49 "Number of block queries that were completely cached"); 50 51// Limit for the number of instructions to scan in a block. 52// FIXME: Figure out what a sane value is for this. 53// (500 is relatively insane.) 54static const int BlockScanLimit = 500; 55 56char MemoryDependenceAnalysis::ID = 0; 57 58// Register this pass... 59INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep", 60 "Memory Dependence Analysis", false, true) 61INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 62INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep", 63 "Memory Dependence Analysis", false, true) 64 65MemoryDependenceAnalysis::MemoryDependenceAnalysis() 66: FunctionPass(ID), PredCache(0) { 67 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry()); 68} 69MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 70} 71 72/// Clean up memory in between runs 73void MemoryDependenceAnalysis::releaseMemory() { 74 LocalDeps.clear(); 75 NonLocalDeps.clear(); 76 NonLocalPointerDeps.clear(); 77 ReverseLocalDeps.clear(); 78 ReverseNonLocalDeps.clear(); 79 ReverseNonLocalPtrDeps.clear(); 80 PredCache->clear(); 81} 82 83 84 85/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 86/// 87void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 88 AU.setPreservesAll(); 89 AU.addRequiredTransitive<AliasAnalysis>(); 90} 91 92bool MemoryDependenceAnalysis::runOnFunction(Function &) { 93 AA = &getAnalysis<AliasAnalysis>(); 94 TD = getAnalysisIfAvailable<TargetData>(); 95 DT = getAnalysisIfAvailable<DominatorTree>(); 96 if (PredCache == 0) 97 PredCache.reset(new PredIteratorCache()); 98 return false; 99} 100 101/// RemoveFromReverseMap - This is a helper function that removes Val from 102/// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry. 103template <typename KeyTy> 104static void RemoveFromReverseMap(DenseMap<Instruction*, 105 SmallPtrSet<KeyTy, 4> > &ReverseMap, 106 Instruction *Inst, KeyTy Val) { 107 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator 108 InstIt = ReverseMap.find(Inst); 109 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); 110 bool Found = InstIt->second.erase(Val); 111 assert(Found && "Invalid reverse map!"); (void)Found; 112 if (InstIt->second.empty()) 113 ReverseMap.erase(InstIt); 114} 115 116/// GetLocation - If the given instruction references a specific memory 117/// location, fill in Loc with the details, otherwise set Loc.Ptr to null. 118/// Return a ModRefInfo value describing the general behavior of the 119/// instruction. 120static 121AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst, 122 AliasAnalysis::Location &Loc, 123 AliasAnalysis *AA) { 124 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 125 if (LI->isUnordered()) { 126 Loc = AA->getLocation(LI); 127 return AliasAnalysis::Ref; 128 } else if (LI->getOrdering() == Monotonic) { 129 Loc = AA->getLocation(LI); 130 return AliasAnalysis::ModRef; 131 } 132 Loc = AliasAnalysis::Location(); 133 return AliasAnalysis::ModRef; 134 } 135 136 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 137 if (SI->isUnordered()) { 138 Loc = AA->getLocation(SI); 139 return AliasAnalysis::Mod; 140 } else if (SI->getOrdering() == Monotonic) { 141 Loc = AA->getLocation(SI); 142 return AliasAnalysis::ModRef; 143 } 144 Loc = AliasAnalysis::Location(); 145 return AliasAnalysis::ModRef; 146 } 147 148 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 149 Loc = AA->getLocation(V); 150 return AliasAnalysis::ModRef; 151 } 152 153 if (const CallInst *CI = isFreeCall(Inst)) { 154 // calls to free() deallocate the entire structure 155 Loc = AliasAnalysis::Location(CI->getArgOperand(0)); 156 return AliasAnalysis::Mod; 157 } 158 159 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) 160 switch (II->getIntrinsicID()) { 161 case Intrinsic::lifetime_start: 162 case Intrinsic::lifetime_end: 163 case Intrinsic::invariant_start: 164 Loc = AliasAnalysis::Location(II->getArgOperand(1), 165 cast<ConstantInt>(II->getArgOperand(0)) 166 ->getZExtValue(), 167 II->getMetadata(LLVMContext::MD_tbaa)); 168 // These intrinsics don't really modify the memory, but returning Mod 169 // will allow them to be handled conservatively. 170 return AliasAnalysis::Mod; 171 case Intrinsic::invariant_end: 172 Loc = AliasAnalysis::Location(II->getArgOperand(2), 173 cast<ConstantInt>(II->getArgOperand(1)) 174 ->getZExtValue(), 175 II->getMetadata(LLVMContext::MD_tbaa)); 176 // These intrinsics don't really modify the memory, but returning Mod 177 // will allow them to be handled conservatively. 178 return AliasAnalysis::Mod; 179 default: 180 break; 181 } 182 183 // Otherwise, just do the coarse-grained thing that always works. 184 if (Inst->mayWriteToMemory()) 185 return AliasAnalysis::ModRef; 186 if (Inst->mayReadFromMemory()) 187 return AliasAnalysis::Ref; 188 return AliasAnalysis::NoModRef; 189} 190 191/// getCallSiteDependencyFrom - Private helper for finding the local 192/// dependencies of a call site. 193MemDepResult MemoryDependenceAnalysis:: 194getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 195 BasicBlock::iterator ScanIt, BasicBlock *BB) { 196 unsigned Limit = BlockScanLimit; 197 198 // Walk backwards through the block, looking for dependencies 199 while (ScanIt != BB->begin()) { 200 // Limit the amount of scanning we do so we don't end up with quadratic 201 // running time on extreme testcases. 202 --Limit; 203 if (!Limit) 204 return MemDepResult::getUnknown(); 205 206 Instruction *Inst = --ScanIt; 207 208 // If this inst is a memory op, get the pointer it accessed 209 AliasAnalysis::Location Loc; 210 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA); 211 if (Loc.Ptr) { 212 // A simple instruction. 213 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef) 214 return MemDepResult::getClobber(Inst); 215 continue; 216 } 217 218 if (CallSite InstCS = cast<Value>(Inst)) { 219 // Debug intrinsics don't cause dependences. 220 if (isa<DbgInfoIntrinsic>(Inst)) continue; 221 // If these two calls do not interfere, look past it. 222 switch (AA->getModRefInfo(CS, InstCS)) { 223 case AliasAnalysis::NoModRef: 224 // If the two calls are the same, return InstCS as a Def, so that 225 // CS can be found redundant and eliminated. 226 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) && 227 CS.getInstruction()->isIdenticalToWhenDefined(Inst)) 228 return MemDepResult::getDef(Inst); 229 230 // Otherwise if the two calls don't interact (e.g. InstCS is readnone) 231 // keep scanning. 232 break; 233 default: 234 return MemDepResult::getClobber(Inst); 235 } 236 } 237 } 238 239 // No dependence found. If this is the entry block of the function, it is 240 // unknown, otherwise it is non-local. 241 if (BB != &BB->getParent()->getEntryBlock()) 242 return MemDepResult::getNonLocal(); 243 return MemDepResult::getNonFuncLocal(); 244} 245 246/// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that 247/// would fully overlap MemLoc if done as a wider legal integer load. 248/// 249/// MemLocBase, MemLocOffset are lazily computed here the first time the 250/// base/offs of memloc is needed. 251static bool 252isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc, 253 const Value *&MemLocBase, 254 int64_t &MemLocOffs, 255 const LoadInst *LI, 256 const TargetData *TD) { 257 // If we have no target data, we can't do this. 258 if (TD == 0) return false; 259 260 // If we haven't already computed the base/offset of MemLoc, do so now. 261 if (MemLocBase == 0) 262 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, *TD); 263 264 unsigned Size = MemoryDependenceAnalysis:: 265 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size, 266 LI, *TD); 267 return Size != 0; 268} 269 270/// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that 271/// looks at a memory location for a load (specified by MemLocBase, Offs, 272/// and Size) and compares it against a load. If the specified load could 273/// be safely widened to a larger integer load that is 1) still efficient, 274/// 2) safe for the target, and 3) would provide the specified memory 275/// location value, then this function returns the size in bytes of the 276/// load width to use. If not, this returns zero. 277unsigned MemoryDependenceAnalysis:: 278getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs, 279 unsigned MemLocSize, const LoadInst *LI, 280 const TargetData &TD) { 281 // We can only extend simple integer loads. 282 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0; 283 284 // Get the base of this load. 285 int64_t LIOffs = 0; 286 const Value *LIBase = 287 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, TD); 288 289 // If the two pointers are not based on the same pointer, we can't tell that 290 // they are related. 291 if (LIBase != MemLocBase) return 0; 292 293 // Okay, the two values are based on the same pointer, but returned as 294 // no-alias. This happens when we have things like two byte loads at "P+1" 295 // and "P+3". Check to see if increasing the size of the "LI" load up to its 296 // alignment (or the largest native integer type) will allow us to load all 297 // the bits required by MemLoc. 298 299 // If MemLoc is before LI, then no widening of LI will help us out. 300 if (MemLocOffs < LIOffs) return 0; 301 302 // Get the alignment of the load in bytes. We assume that it is safe to load 303 // any legal integer up to this size without a problem. For example, if we're 304 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can 305 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it 306 // to i16. 307 unsigned LoadAlign = LI->getAlignment(); 308 309 int64_t MemLocEnd = MemLocOffs+MemLocSize; 310 311 // If no amount of rounding up will let MemLoc fit into LI, then bail out. 312 if (LIOffs+LoadAlign < MemLocEnd) return 0; 313 314 // This is the size of the load to try. Start with the next larger power of 315 // two. 316 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U; 317 NewLoadByteSize = NextPowerOf2(NewLoadByteSize); 318 319 while (1) { 320 // If this load size is bigger than our known alignment or would not fit 321 // into a native integer register, then we fail. 322 if (NewLoadByteSize > LoadAlign || 323 !TD.fitsInLegalInteger(NewLoadByteSize*8)) 324 return 0; 325 326 // If a load of this width would include all of MemLoc, then we succeed. 327 if (LIOffs+NewLoadByteSize >= MemLocEnd) 328 return NewLoadByteSize; 329 330 NewLoadByteSize <<= 1; 331 } 332 333 return 0; 334} 335 336namespace { 337 /// Only find pointer captures which happen before the given instruction. Uses 338 /// the dominator tree to determine whether one instruction is before another. 339 struct CapturesBefore : public CaptureTracker { 340 CapturesBefore(const Instruction *I, DominatorTree *DT) 341 : BeforeHere(I), DT(DT), Captured(false) {} 342 343 void tooManyUses() { Captured = true; } 344 345 bool shouldExplore(Use *U) { 346 Instruction *I = cast<Instruction>(U->getUser()); 347 if (BeforeHere != I && DT->dominates(BeforeHere, I)) 348 return false; 349 return true; 350 } 351 352 bool captured(Instruction *I) { 353 if (BeforeHere != I && DT->dominates(BeforeHere, I)) 354 return false; 355 Captured = true; 356 return true; 357 } 358 359 const Instruction *BeforeHere; 360 DominatorTree *DT; 361 362 bool Captured; 363 }; 364} 365 366AliasAnalysis::ModRefResult 367MemoryDependenceAnalysis::getModRefInfo(const Instruction *Inst, 368 const AliasAnalysis::Location &MemLoc) { 369 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc); 370 if (MR != AliasAnalysis::ModRef) return MR; 371 372 // FIXME: this is really just shoring-up a deficiency in alias analysis. 373 // BasicAA isn't willing to spend linear time determining whether an alloca 374 // was captured before or after this particular call, while we are. However, 375 // with a smarter AA in place, this test is just wasting compile time. 376 if (!DT) return AliasAnalysis::ModRef; 377 const Value *Object = GetUnderlyingObject(MemLoc.Ptr, TD); 378 if (!isIdentifiedObject(Object) || isa<GlobalValue>(Object)) 379 return AliasAnalysis::ModRef; 380 ImmutableCallSite CS(Inst); 381 if (!CS.getInstruction()) return AliasAnalysis::ModRef; 382 383 CapturesBefore CB(Inst, DT); 384 llvm::PointerMayBeCaptured(Object, &CB); 385 386 if (isa<Constant>(Object) || CS.getInstruction() == Object || CB.Captured) 387 return AliasAnalysis::ModRef; 388 389 unsigned ArgNo = 0; 390 for (ImmutableCallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end(); 391 CI != CE; ++CI, ++ArgNo) { 392 // Only look at the no-capture or byval pointer arguments. If this 393 // pointer were passed to arguments that were neither of these, then it 394 // couldn't be no-capture. 395 if (!(*CI)->getType()->isPointerTy() || 396 (!CS.doesNotCapture(ArgNo) && !CS.isByValArgument(ArgNo))) 397 continue; 398 399 // If this is a no-capture pointer argument, see if we can tell that it 400 // is impossible to alias the pointer we're checking. If not, we have to 401 // assume that the call could touch the pointer, even though it doesn't 402 // escape. 403 if (!AA->isNoAlias(AliasAnalysis::Location(*CI), 404 AliasAnalysis::Location(Object))) { 405 return AliasAnalysis::ModRef; 406 } 407 } 408 return AliasAnalysis::NoModRef; 409} 410 411/// getPointerDependencyFrom - Return the instruction on which a memory 412/// location depends. If isLoad is true, this routine ignores may-aliases with 413/// read-only operations. If isLoad is false, this routine ignores may-aliases 414/// with reads from read-only locations. 415MemDepResult MemoryDependenceAnalysis:: 416getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad, 417 BasicBlock::iterator ScanIt, BasicBlock *BB) { 418 419 const Value *MemLocBase = 0; 420 int64_t MemLocOffset = 0; 421 422 unsigned Limit = BlockScanLimit; 423 424 // Walk backwards through the basic block, looking for dependencies. 425 while (ScanIt != BB->begin()) { 426 // Limit the amount of scanning we do so we don't end up with quadratic 427 // running time on extreme testcases. 428 --Limit; 429 if (!Limit) 430 return MemDepResult::getUnknown(); 431 432 Instruction *Inst = --ScanIt; 433 434 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 435 // Debug intrinsics don't (and can't) cause dependences. 436 if (isa<DbgInfoIntrinsic>(II)) continue; 437 438 // 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 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 455 // Atomic loads have complications involved. 456 // FIXME: This is overly conservative. 457 if (!LI->isUnordered()) 458 return MemDepResult::getClobber(LI); 459 460 AliasAnalysis::Location LoadLoc = AA->getLocation(LI); 461 462 // If we found a pointer, check if it could be the same as our pointer. 463 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc); 464 465 if (isLoad) { 466 if (R == AliasAnalysis::NoAlias) { 467 // If this is an over-aligned integer load (for example, 468 // "load i8* %P, align 4") see if it would obviously overlap with the 469 // queried location if widened to a larger load (e.g. if the queried 470 // location is 1 byte at P+1). If so, return it as a load/load 471 // clobber result, allowing the client to decide to widen the load if 472 // it wants to. 473 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) 474 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() && 475 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase, 476 MemLocOffset, LI, TD)) 477 return MemDepResult::getClobber(Inst); 478 479 continue; 480 } 481 482 // Must aliased loads are defs of each other. 483 if (R == AliasAnalysis::MustAlias) 484 return MemDepResult::getDef(Inst); 485 486#if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads 487 // in terms of clobbering loads, but since it does this by looking 488 // at the clobbering load directly, it doesn't know about any 489 // phi translation that may have happened along the way. 490 491 // If we have a partial alias, then return this as a clobber for the 492 // client to handle. 493 if (R == AliasAnalysis::PartialAlias) 494 return MemDepResult::getClobber(Inst); 495#endif 496 497 // Random may-alias loads don't depend on each other without a 498 // dependence. 499 continue; 500 } 501 502 // Stores don't depend on other no-aliased accesses. 503 if (R == AliasAnalysis::NoAlias) 504 continue; 505 506 // Stores don't alias loads from read-only memory. 507 if (AA->pointsToConstantMemory(LoadLoc)) 508 continue; 509 510 // Stores depend on may/must aliased loads. 511 return MemDepResult::getDef(Inst); 512 } 513 514 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 515 // Atomic stores have complications involved. 516 // FIXME: This is overly conservative. 517 if (!SI->isUnordered()) 518 return MemDepResult::getClobber(SI); 519 520 // If alias analysis can tell that this store is guaranteed to not modify 521 // the query pointer, ignore it. Use getModRefInfo to handle cases where 522 // the query pointer points to constant memory etc. 523 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef) 524 continue; 525 526 // Ok, this store might clobber the query pointer. Check to see if it is 527 // a must alias: in this case, we want to return this as a def. 528 AliasAnalysis::Location StoreLoc = AA->getLocation(SI); 529 530 // If we found a pointer, check if it could be the same as our pointer. 531 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc); 532 533 if (R == AliasAnalysis::NoAlias) 534 continue; 535 if (R == AliasAnalysis::MustAlias) 536 return MemDepResult::getDef(Inst); 537 return MemDepResult::getClobber(Inst); 538 } 539 540 // If this is an allocation, and if we know that the accessed pointer is to 541 // the allocation, return Def. This means that there is no dependence and 542 // the access can be optimized based on that. For example, a load could 543 // turn into undef. 544 // Note: Only determine this to be a malloc if Inst is the malloc call, not 545 // a subsequent bitcast of the malloc call result. There can be stores to 546 // the malloced memory between the malloc call and its bitcast uses, and we 547 // need to continue scanning until the malloc call. 548 if (isa<AllocaInst>(Inst) || 549 (isa<CallInst>(Inst) && extractMallocCall(Inst))) { 550 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD); 551 552 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr)) 553 return MemDepResult::getDef(Inst); 554 continue; 555 } 556 557 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 558 switch (getModRefInfo(Inst, MemLoc)) { 559 case AliasAnalysis::NoModRef: 560 // If the call has no effect on the queried pointer, just ignore it. 561 continue; 562 case AliasAnalysis::Mod: 563 return MemDepResult::getClobber(Inst); 564 case AliasAnalysis::Ref: 565 // If the call is known to never store to the pointer, and if this is a 566 // load query, we can safely ignore it (scan past it). 567 if (isLoad) 568 continue; 569 default: 570 // Otherwise, there is a potential dependence. Return a clobber. 571 return MemDepResult::getClobber(Inst); 572 } 573 } 574 575 // No dependence found. If this is the entry block of the function, it is 576 // unknown, otherwise it is non-local. 577 if (BB != &BB->getParent()->getEntryBlock()) 578 return MemDepResult::getNonLocal(); 579 return MemDepResult::getNonFuncLocal(); 580} 581 582/// getDependency - Return the instruction on which a memory operation 583/// depends. 584MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 585 Instruction *ScanPos = QueryInst; 586 587 // Check for a cached result 588 MemDepResult &LocalCache = LocalDeps[QueryInst]; 589 590 // If the cached entry is non-dirty, just return it. Note that this depends 591 // on MemDepResult's default constructing to 'dirty'. 592 if (!LocalCache.isDirty()) 593 return LocalCache; 594 595 // Otherwise, if we have a dirty entry, we know we can start the scan at that 596 // instruction, which may save us some work. 597 if (Instruction *Inst = LocalCache.getInst()) { 598 ScanPos = Inst; 599 600 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 601 } 602 603 BasicBlock *QueryParent = QueryInst->getParent(); 604 605 // Do the scan. 606 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 607 // No dependence found. If this is the entry block of the function, it is 608 // unknown, otherwise it is non-local. 609 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 610 LocalCache = MemDepResult::getNonLocal(); 611 else 612 LocalCache = MemDepResult::getNonFuncLocal(); 613 } else { 614 AliasAnalysis::Location MemLoc; 615 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA); 616 if (MemLoc.Ptr) { 617 // If we can do a pointer scan, make it happen. 618 bool isLoad = !(MR & AliasAnalysis::Mod); 619 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst)) 620 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start; 621 622 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos, 623 QueryParent); 624 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 625 CallSite QueryCS(QueryInst); 626 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 627 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 628 QueryParent); 629 } else 630 // Non-memory instruction. 631 LocalCache = MemDepResult::getUnknown(); 632 } 633 634 // Remember the result! 635 if (Instruction *I = LocalCache.getInst()) 636 ReverseLocalDeps[I].insert(QueryInst); 637 638 return LocalCache; 639} 640 641#ifndef NDEBUG 642/// AssertSorted - This method is used when -debug is specified to verify that 643/// cache arrays are properly kept sorted. 644static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 645 int Count = -1) { 646 if (Count == -1) Count = Cache.size(); 647 if (Count == 0) return; 648 649 for (unsigned i = 1; i != unsigned(Count); ++i) 650 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 651} 652#endif 653 654/// getNonLocalCallDependency - Perform a full dependency query for the 655/// specified call, returning the set of blocks that the value is 656/// potentially live across. The returned set of results will include a 657/// "NonLocal" result for all blocks where the value is live across. 658/// 659/// This method assumes the instruction returns a "NonLocal" dependency 660/// within its own block. 661/// 662/// This returns a reference to an internal data structure that may be 663/// invalidated on the next non-local query or when an instruction is 664/// removed. Clients must copy this data if they want it around longer than 665/// that. 666const MemoryDependenceAnalysis::NonLocalDepInfo & 667MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 668 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 669 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 670 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 671 NonLocalDepInfo &Cache = CacheP.first; 672 673 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 674 /// the cached case, this can happen due to instructions being deleted etc. In 675 /// the uncached case, this starts out as the set of predecessors we care 676 /// about. 677 SmallVector<BasicBlock*, 32> DirtyBlocks; 678 679 if (!Cache.empty()) { 680 // Okay, we have a cache entry. If we know it is not dirty, just return it 681 // with no computation. 682 if (!CacheP.second) { 683 ++NumCacheNonLocal; 684 return Cache; 685 } 686 687 // If we already have a partially computed set of results, scan them to 688 // determine what is dirty, seeding our initial DirtyBlocks worklist. 689 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 690 I != E; ++I) 691 if (I->getResult().isDirty()) 692 DirtyBlocks.push_back(I->getBB()); 693 694 // Sort the cache so that we can do fast binary search lookups below. 695 std::sort(Cache.begin(), Cache.end()); 696 697 ++NumCacheDirtyNonLocal; 698 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 699 // << Cache.size() << " cached: " << *QueryInst; 700 } else { 701 // Seed DirtyBlocks with each of the preds of QueryInst's block. 702 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 703 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 704 DirtyBlocks.push_back(*PI); 705 ++NumUncacheNonLocal; 706 } 707 708 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 709 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 710 711 SmallPtrSet<BasicBlock*, 64> Visited; 712 713 unsigned NumSortedEntries = Cache.size(); 714 DEBUG(AssertSorted(Cache)); 715 716 // Iterate while we still have blocks to update. 717 while (!DirtyBlocks.empty()) { 718 BasicBlock *DirtyBB = DirtyBlocks.back(); 719 DirtyBlocks.pop_back(); 720 721 // Already processed this block? 722 if (!Visited.insert(DirtyBB)) 723 continue; 724 725 // Do a binary search to see if we already have an entry for this block in 726 // the cache set. If so, find it. 727 DEBUG(AssertSorted(Cache, NumSortedEntries)); 728 NonLocalDepInfo::iterator Entry = 729 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 730 NonLocalDepEntry(DirtyBB)); 731 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB) 732 --Entry; 733 734 NonLocalDepEntry *ExistingResult = 0; 735 if (Entry != Cache.begin()+NumSortedEntries && 736 Entry->getBB() == DirtyBB) { 737 // If we already have an entry, and if it isn't already dirty, the block 738 // is done. 739 if (!Entry->getResult().isDirty()) 740 continue; 741 742 // Otherwise, remember this slot so we can update the value. 743 ExistingResult = &*Entry; 744 } 745 746 // If the dirty entry has a pointer, start scanning from it so we don't have 747 // to rescan the entire block. 748 BasicBlock::iterator ScanPos = DirtyBB->end(); 749 if (ExistingResult) { 750 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 751 ScanPos = Inst; 752 // We're removing QueryInst's use of Inst. 753 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 754 QueryCS.getInstruction()); 755 } 756 } 757 758 // Find out if this block has a local dependency for QueryInst. 759 MemDepResult Dep; 760 761 if (ScanPos != DirtyBB->begin()) { 762 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 763 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 764 // No dependence found. If this is the entry block of the function, it is 765 // a clobber, otherwise it is unknown. 766 Dep = MemDepResult::getNonLocal(); 767 } else { 768 Dep = MemDepResult::getNonFuncLocal(); 769 } 770 771 // If we had a dirty entry for the block, update it. Otherwise, just add 772 // a new entry. 773 if (ExistingResult) 774 ExistingResult->setResult(Dep); 775 else 776 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 777 778 // If the block has a dependency (i.e. it isn't completely transparent to 779 // the value), remember the association! 780 if (!Dep.isNonLocal()) { 781 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 782 // update this when we remove instructions. 783 if (Instruction *Inst = Dep.getInst()) 784 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 785 } else { 786 787 // If the block *is* completely transparent to the load, we need to check 788 // the predecessors of this block. Add them to our worklist. 789 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 790 DirtyBlocks.push_back(*PI); 791 } 792 } 793 794 return Cache; 795} 796 797/// getNonLocalPointerDependency - Perform a full dependency query for an 798/// access to the specified (non-volatile) memory location, returning the 799/// set of instructions that either define or clobber the value. 800/// 801/// This method assumes the pointer has a "NonLocal" dependency within its 802/// own block. 803/// 804void MemoryDependenceAnalysis:: 805getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad, 806 BasicBlock *FromBB, 807 SmallVectorImpl<NonLocalDepResult> &Result) { 808 assert(Loc.Ptr->getType()->isPointerTy() && 809 "Can't get pointer deps of a non-pointer!"); 810 Result.clear(); 811 812 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD); 813 814 // This is the set of blocks we've inspected, and the pointer we consider in 815 // each block. Because of critical edges, we currently bail out if querying 816 // a block with multiple different pointers. This can happen during PHI 817 // translation. 818 DenseMap<BasicBlock*, Value*> Visited; 819 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB, 820 Result, Visited, true)) 821 return; 822 Result.clear(); 823 Result.push_back(NonLocalDepResult(FromBB, 824 MemDepResult::getUnknown(), 825 const_cast<Value *>(Loc.Ptr))); 826} 827 828/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 829/// Pointer/PointeeSize using either cached information in Cache or by doing a 830/// lookup (which may use dirty cache info if available). If we do a lookup, 831/// add the result to the cache. 832MemDepResult MemoryDependenceAnalysis:: 833GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc, 834 bool isLoad, BasicBlock *BB, 835 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 836 837 // Do a binary search to see if we already have an entry for this block in 838 // the cache set. If so, find it. 839 NonLocalDepInfo::iterator Entry = 840 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 841 NonLocalDepEntry(BB)); 842 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 843 --Entry; 844 845 NonLocalDepEntry *ExistingResult = 0; 846 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 847 ExistingResult = &*Entry; 848 849 // If we have a cached entry, and it is non-dirty, use it as the value for 850 // this dependency. 851 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 852 ++NumCacheNonLocalPtr; 853 return ExistingResult->getResult(); 854 } 855 856 // Otherwise, we have to scan for the value. If we have a dirty cache 857 // entry, start scanning from its position, otherwise we scan from the end 858 // of the block. 859 BasicBlock::iterator ScanPos = BB->end(); 860 if (ExistingResult && ExistingResult->getResult().getInst()) { 861 assert(ExistingResult->getResult().getInst()->getParent() == BB && 862 "Instruction invalidated?"); 863 ++NumCacheDirtyNonLocalPtr; 864 ScanPos = ExistingResult->getResult().getInst(); 865 866 // Eliminating the dirty entry from 'Cache', so update the reverse info. 867 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 868 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 869 } else { 870 ++NumUncacheNonLocalPtr; 871 } 872 873 // Scan the block for the dependency. 874 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB); 875 876 // If we had a dirty entry for the block, update it. Otherwise, just add 877 // a new entry. 878 if (ExistingResult) 879 ExistingResult->setResult(Dep); 880 else 881 Cache->push_back(NonLocalDepEntry(BB, Dep)); 882 883 // If the block has a dependency (i.e. it isn't completely transparent to 884 // the value), remember the reverse association because we just added it 885 // to Cache! 886 if (!Dep.isDef() && !Dep.isClobber()) 887 return Dep; 888 889 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 890 // update MemDep when we remove instructions. 891 Instruction *Inst = Dep.getInst(); 892 assert(Inst && "Didn't depend on anything?"); 893 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 894 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 895 return Dep; 896} 897 898/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 899/// number of elements in the array that are already properly ordered. This is 900/// optimized for the case when only a few entries are added. 901static void 902SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 903 unsigned NumSortedEntries) { 904 switch (Cache.size() - NumSortedEntries) { 905 case 0: 906 // done, no new entries. 907 break; 908 case 2: { 909 // Two new entries, insert the last one into place. 910 NonLocalDepEntry Val = Cache.back(); 911 Cache.pop_back(); 912 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 913 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 914 Cache.insert(Entry, Val); 915 // FALL THROUGH. 916 } 917 case 1: 918 // One new entry, Just insert the new value at the appropriate position. 919 if (Cache.size() != 1) { 920 NonLocalDepEntry Val = Cache.back(); 921 Cache.pop_back(); 922 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 923 std::upper_bound(Cache.begin(), Cache.end(), Val); 924 Cache.insert(Entry, Val); 925 } 926 break; 927 default: 928 // Added many values, do a full scale sort. 929 std::sort(Cache.begin(), Cache.end()); 930 break; 931 } 932} 933 934/// getNonLocalPointerDepFromBB - Perform a dependency query based on 935/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 936/// results to the results vector and keep track of which blocks are visited in 937/// 'Visited'. 938/// 939/// This has special behavior for the first block queries (when SkipFirstBlock 940/// is true). In this special case, it ignores the contents of the specified 941/// block and starts returning dependence info for its predecessors. 942/// 943/// This function returns false on success, or true to indicate that it could 944/// not compute dependence information for some reason. This should be treated 945/// as a clobber dependence on the first instruction in the predecessor block. 946bool MemoryDependenceAnalysis:: 947getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, 948 const AliasAnalysis::Location &Loc, 949 bool isLoad, BasicBlock *StartBB, 950 SmallVectorImpl<NonLocalDepResult> &Result, 951 DenseMap<BasicBlock*, Value*> &Visited, 952 bool SkipFirstBlock) { 953 954 // Look up the cached info for Pointer. 955 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 956 957 // Set up a temporary NLPI value. If the map doesn't yet have an entry for 958 // CacheKey, this value will be inserted as the associated value. Otherwise, 959 // it'll be ignored, and we'll have to check to see if the cached size and 960 // tbaa tag are consistent with the current query. 961 NonLocalPointerInfo InitialNLPI; 962 InitialNLPI.Size = Loc.Size; 963 InitialNLPI.TBAATag = Loc.TBAATag; 964 965 // Get the NLPI for CacheKey, inserting one into the map if it doesn't 966 // already have one. 967 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair = 968 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI)); 969 NonLocalPointerInfo *CacheInfo = &Pair.first->second; 970 971 // If we already have a cache entry for this CacheKey, we may need to do some 972 // work to reconcile the cache entry and the current query. 973 if (!Pair.second) { 974 if (CacheInfo->Size < Loc.Size) { 975 // The query's Size is greater than the cached one. Throw out the 976 // cached data and procede with the query at the greater size. 977 CacheInfo->Pair = BBSkipFirstBlockPair(); 978 CacheInfo->Size = Loc.Size; 979 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 980 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 981 if (Instruction *Inst = DI->getResult().getInst()) 982 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 983 CacheInfo->NonLocalDeps.clear(); 984 } else if (CacheInfo->Size > Loc.Size) { 985 // This query's Size is less than the cached one. Conservatively restart 986 // the query using the greater size. 987 return getNonLocalPointerDepFromBB(Pointer, 988 Loc.getWithNewSize(CacheInfo->Size), 989 isLoad, StartBB, Result, Visited, 990 SkipFirstBlock); 991 } 992 993 // If the query's TBAATag is inconsistent with the cached one, 994 // conservatively throw out the cached data and restart the query with 995 // no tag if needed. 996 if (CacheInfo->TBAATag != Loc.TBAATag) { 997 if (CacheInfo->TBAATag) { 998 CacheInfo->Pair = BBSkipFirstBlockPair(); 999 CacheInfo->TBAATag = 0; 1000 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 1001 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 1002 if (Instruction *Inst = DI->getResult().getInst()) 1003 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 1004 CacheInfo->NonLocalDeps.clear(); 1005 } 1006 if (Loc.TBAATag) 1007 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(), 1008 isLoad, StartBB, Result, Visited, 1009 SkipFirstBlock); 1010 } 1011 } 1012 1013 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps; 1014 1015 // If we have valid cached information for exactly the block we are 1016 // investigating, just return it with no recomputation. 1017 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 1018 // We have a fully cached result for this query then we can just return the 1019 // cached results and populate the visited set. However, we have to verify 1020 // that we don't already have conflicting results for these blocks. Check 1021 // to ensure that if a block in the results set is in the visited set that 1022 // it was for the same pointer query. 1023 if (!Visited.empty()) { 1024 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 1025 I != E; ++I) { 1026 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 1027 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 1028 continue; 1029 1030 // We have a pointer mismatch in a block. Just return clobber, saying 1031 // that something was clobbered in this result. We could also do a 1032 // non-fully cached query, but there is little point in doing this. 1033 return true; 1034 } 1035 } 1036 1037 Value *Addr = Pointer.getAddr(); 1038 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 1039 I != E; ++I) { 1040 Visited.insert(std::make_pair(I->getBB(), Addr)); 1041 if (!I->getResult().isNonLocal()) 1042 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 1043 } 1044 ++NumCacheCompleteNonLocalPtr; 1045 return false; 1046 } 1047 1048 // Otherwise, either this is a new block, a block with an invalid cache 1049 // pointer or one that we're about to invalidate by putting more info into it 1050 // than its valid cache info. If empty, the result will be valid cache info, 1051 // otherwise it isn't. 1052 if (Cache->empty()) 1053 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 1054 else 1055 CacheInfo->Pair = BBSkipFirstBlockPair(); 1056 1057 SmallVector<BasicBlock*, 32> Worklist; 1058 Worklist.push_back(StartBB); 1059 1060 // PredList used inside loop. 1061 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList; 1062 1063 // Keep track of the entries that we know are sorted. Previously cached 1064 // entries will all be sorted. The entries we add we only sort on demand (we 1065 // don't insert every element into its sorted position). We know that we 1066 // won't get any reuse from currently inserted values, because we don't 1067 // revisit blocks after we insert info for them. 1068 unsigned NumSortedEntries = Cache->size(); 1069 DEBUG(AssertSorted(*Cache)); 1070 1071 while (!Worklist.empty()) { 1072 BasicBlock *BB = Worklist.pop_back_val(); 1073 1074 // Skip the first block if we have it. 1075 if (!SkipFirstBlock) { 1076 // Analyze the dependency of *Pointer in FromBB. See if we already have 1077 // been here. 1078 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 1079 1080 // Get the dependency info for Pointer in BB. If we have cached 1081 // information, we will use it, otherwise we compute it. 1082 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 1083 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache, 1084 NumSortedEntries); 1085 1086 // If we got a Def or Clobber, add this to the list of results. 1087 if (!Dep.isNonLocal()) { 1088 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 1089 continue; 1090 } 1091 } 1092 1093 // If 'Pointer' is an instruction defined in this block, then we need to do 1094 // phi translation to change it into a value live in the predecessor block. 1095 // If not, we just add the predecessors to the worklist and scan them with 1096 // the same Pointer. 1097 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 1098 SkipFirstBlock = false; 1099 SmallVector<BasicBlock*, 16> NewBlocks; 1100 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1101 // Verify that we haven't looked at this block yet. 1102 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1103 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 1104 if (InsertRes.second) { 1105 // First time we've looked at *PI. 1106 NewBlocks.push_back(*PI); 1107 continue; 1108 } 1109 1110 // If we have seen this block before, but it was with a different 1111 // pointer then we have a phi translation failure and we have to treat 1112 // this as a clobber. 1113 if (InsertRes.first->second != Pointer.getAddr()) { 1114 // Make sure to clean up the Visited map before continuing on to 1115 // PredTranslationFailure. 1116 for (unsigned i = 0; i < NewBlocks.size(); i++) 1117 Visited.erase(NewBlocks[i]); 1118 goto PredTranslationFailure; 1119 } 1120 } 1121 Worklist.append(NewBlocks.begin(), NewBlocks.end()); 1122 continue; 1123 } 1124 1125 // We do need to do phi translation, if we know ahead of time we can't phi 1126 // translate this value, don't even try. 1127 if (!Pointer.IsPotentiallyPHITranslatable()) 1128 goto PredTranslationFailure; 1129 1130 // We may have added values to the cache list before this PHI translation. 1131 // If so, we haven't done anything to ensure that the cache remains sorted. 1132 // Sort it now (if needed) so that recursive invocations of 1133 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 1134 // value will only see properly sorted cache arrays. 1135 if (Cache && NumSortedEntries != Cache->size()) { 1136 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1137 NumSortedEntries = Cache->size(); 1138 } 1139 Cache = 0; 1140 1141 PredList.clear(); 1142 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1143 BasicBlock *Pred = *PI; 1144 PredList.push_back(std::make_pair(Pred, Pointer)); 1145 1146 // Get the PHI translated pointer in this predecessor. This can fail if 1147 // not translatable, in which case the getAddr() returns null. 1148 PHITransAddr &PredPointer = PredList.back().second; 1149 PredPointer.PHITranslateValue(BB, Pred, 0); 1150 1151 Value *PredPtrVal = PredPointer.getAddr(); 1152 1153 // Check to see if we have already visited this pred block with another 1154 // pointer. If so, we can't do this lookup. This failure can occur 1155 // with PHI translation when a critical edge exists and the PHI node in 1156 // the successor translates to a pointer value different than the 1157 // pointer the block was first analyzed with. 1158 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1159 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 1160 1161 if (!InsertRes.second) { 1162 // We found the pred; take it off the list of preds to visit. 1163 PredList.pop_back(); 1164 1165 // If the predecessor was visited with PredPtr, then we already did 1166 // the analysis and can ignore it. 1167 if (InsertRes.first->second == PredPtrVal) 1168 continue; 1169 1170 // Otherwise, the block was previously analyzed with a different 1171 // pointer. We can't represent the result of this case, so we just 1172 // treat this as a phi translation failure. 1173 1174 // Make sure to clean up the Visited map before continuing on to 1175 // PredTranslationFailure. 1176 for (unsigned i = 0; i < PredList.size(); i++) 1177 Visited.erase(PredList[i].first); 1178 1179 goto PredTranslationFailure; 1180 } 1181 } 1182 1183 // Actually process results here; this need to be a separate loop to avoid 1184 // calling getNonLocalPointerDepFromBB for blocks we don't want to return 1185 // any results for. (getNonLocalPointerDepFromBB will modify our 1186 // datastructures in ways the code after the PredTranslationFailure label 1187 // doesn't expect.) 1188 for (unsigned i = 0; i < PredList.size(); i++) { 1189 BasicBlock *Pred = PredList[i].first; 1190 PHITransAddr &PredPointer = PredList[i].second; 1191 Value *PredPtrVal = PredPointer.getAddr(); 1192 1193 bool CanTranslate = true; 1194 // If PHI translation was unable to find an available pointer in this 1195 // predecessor, then we have to assume that the pointer is clobbered in 1196 // that predecessor. We can still do PRE of the load, which would insert 1197 // a computation of the pointer in this predecessor. 1198 if (PredPtrVal == 0) 1199 CanTranslate = false; 1200 1201 // FIXME: it is entirely possible that PHI translating will end up with 1202 // the same value. Consider PHI translating something like: 1203 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 1204 // to recurse here, pedantically speaking. 1205 1206 // If getNonLocalPointerDepFromBB fails here, that means the cached 1207 // result conflicted with the Visited list; we have to conservatively 1208 // assume it is unknown, but this also does not block PRE of the load. 1209 if (!CanTranslate || 1210 getNonLocalPointerDepFromBB(PredPointer, 1211 Loc.getWithNewPtr(PredPtrVal), 1212 isLoad, Pred, 1213 Result, Visited)) { 1214 // Add the entry to the Result list. 1215 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal); 1216 Result.push_back(Entry); 1217 1218 // Since we had a phi translation failure, the cache for CacheKey won't 1219 // include all of the entries that we need to immediately satisfy future 1220 // queries. Mark this in NonLocalPointerDeps by setting the 1221 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 1222 // cached value to do more work but not miss the phi trans failure. 1223 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey]; 1224 NLPI.Pair = BBSkipFirstBlockPair(); 1225 continue; 1226 } 1227 } 1228 1229 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 1230 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1231 Cache = &CacheInfo->NonLocalDeps; 1232 NumSortedEntries = Cache->size(); 1233 1234 // Since we did phi translation, the "Cache" set won't contain all of the 1235 // results for the query. This is ok (we can still use it to accelerate 1236 // specific block queries) but we can't do the fastpath "return all 1237 // results from the set" Clear out the indicator for this. 1238 CacheInfo->Pair = BBSkipFirstBlockPair(); 1239 SkipFirstBlock = false; 1240 continue; 1241 1242 PredTranslationFailure: 1243 // The following code is "failure"; we can't produce a sane translation 1244 // for the given block. It assumes that we haven't modified any of 1245 // our datastructures while processing the current block. 1246 1247 if (Cache == 0) { 1248 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1249 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1250 Cache = &CacheInfo->NonLocalDeps; 1251 NumSortedEntries = Cache->size(); 1252 } 1253 1254 // Since we failed phi translation, the "Cache" set won't contain all of the 1255 // results for the query. This is ok (we can still use it to accelerate 1256 // specific block queries) but we can't do the fastpath "return all 1257 // results from the set". Clear out the indicator for this. 1258 CacheInfo->Pair = BBSkipFirstBlockPair(); 1259 1260 // If *nothing* works, mark the pointer as unknown. 1261 // 1262 // If this is the magic first block, return this as a clobber of the whole 1263 // incoming value. Since we can't phi translate to one of the predecessors, 1264 // we have to bail out. 1265 if (SkipFirstBlock) 1266 return true; 1267 1268 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 1269 assert(I != Cache->rend() && "Didn't find current block??"); 1270 if (I->getBB() != BB) 1271 continue; 1272 1273 assert(I->getResult().isNonLocal() && 1274 "Should only be here with transparent block"); 1275 I->setResult(MemDepResult::getUnknown()); 1276 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 1277 Pointer.getAddr())); 1278 break; 1279 } 1280 } 1281 1282 // Okay, we're done now. If we added new values to the cache, re-sort it. 1283 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1284 DEBUG(AssertSorted(*Cache)); 1285 return false; 1286} 1287 1288/// RemoveCachedNonLocalPointerDependencies - If P exists in 1289/// CachedNonLocalPointerInfo, remove it. 1290void MemoryDependenceAnalysis:: 1291RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 1292 CachedNonLocalPointerInfo::iterator It = 1293 NonLocalPointerDeps.find(P); 1294 if (It == NonLocalPointerDeps.end()) return; 1295 1296 // Remove all of the entries in the BB->val map. This involves removing 1297 // instructions from the reverse map. 1298 NonLocalDepInfo &PInfo = It->second.NonLocalDeps; 1299 1300 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 1301 Instruction *Target = PInfo[i].getResult().getInst(); 1302 if (Target == 0) continue; // Ignore non-local dep results. 1303 assert(Target->getParent() == PInfo[i].getBB()); 1304 1305 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1306 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1307 } 1308 1309 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1310 NonLocalPointerDeps.erase(It); 1311} 1312 1313 1314/// invalidateCachedPointerInfo - This method is used to invalidate cached 1315/// information about the specified pointer, because it may be too 1316/// conservative in memdep. This is an optional call that can be used when 1317/// the client detects an equivalence between the pointer and some other 1318/// value and replaces the other value with ptr. This can make Ptr available 1319/// in more places that cached info does not necessarily keep. 1320void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1321 // If Ptr isn't really a pointer, just ignore it. 1322 if (!Ptr->getType()->isPointerTy()) return; 1323 // Flush store info for the pointer. 1324 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1325 // Flush load info for the pointer. 1326 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1327} 1328 1329/// invalidateCachedPredecessors - Clear the PredIteratorCache info. 1330/// This needs to be done when the CFG changes, e.g., due to splitting 1331/// critical edges. 1332void MemoryDependenceAnalysis::invalidateCachedPredecessors() { 1333 PredCache->clear(); 1334} 1335 1336/// removeInstruction - Remove an instruction from the dependence analysis, 1337/// updating the dependence of instructions that previously depended on it. 1338/// This method attempts to keep the cache coherent using the reverse map. 1339void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1340 // Walk through the Non-local dependencies, removing this one as the value 1341 // for any cached queries. 1342 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1343 if (NLDI != NonLocalDeps.end()) { 1344 NonLocalDepInfo &BlockMap = NLDI->second.first; 1345 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1346 DI != DE; ++DI) 1347 if (Instruction *Inst = DI->getResult().getInst()) 1348 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1349 NonLocalDeps.erase(NLDI); 1350 } 1351 1352 // If we have a cached local dependence query for this instruction, remove it. 1353 // 1354 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1355 if (LocalDepEntry != LocalDeps.end()) { 1356 // Remove us from DepInst's reverse set now that the local dep info is gone. 1357 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1358 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1359 1360 // Remove this local dependency info. 1361 LocalDeps.erase(LocalDepEntry); 1362 } 1363 1364 // If we have any cached pointer dependencies on this instruction, remove 1365 // them. If the instruction has non-pointer type, then it can't be a pointer 1366 // base. 1367 1368 // Remove it from both the load info and the store info. The instruction 1369 // can't be in either of these maps if it is non-pointer. 1370 if (RemInst->getType()->isPointerTy()) { 1371 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1372 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1373 } 1374 1375 // Loop over all of the things that depend on the instruction we're removing. 1376 // 1377 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1378 1379 // If we find RemInst as a clobber or Def in any of the maps for other values, 1380 // we need to replace its entry with a dirty version of the instruction after 1381 // it. If RemInst is a terminator, we use a null dirty value. 1382 // 1383 // Using a dirty version of the instruction after RemInst saves having to scan 1384 // the entire block to get to this point. 1385 MemDepResult NewDirtyVal; 1386 if (!RemInst->isTerminator()) 1387 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1388 1389 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1390 if (ReverseDepIt != ReverseLocalDeps.end()) { 1391 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1392 // RemInst can't be the terminator if it has local stuff depending on it. 1393 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1394 "Nothing can locally depend on a terminator"); 1395 1396 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1397 E = ReverseDeps.end(); I != E; ++I) { 1398 Instruction *InstDependingOnRemInst = *I; 1399 assert(InstDependingOnRemInst != RemInst && 1400 "Already removed our local dep info"); 1401 1402 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1403 1404 // Make sure to remember that new things depend on NewDepInst. 1405 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1406 "a local dep on this if it is a terminator!"); 1407 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1408 InstDependingOnRemInst)); 1409 } 1410 1411 ReverseLocalDeps.erase(ReverseDepIt); 1412 1413 // Add new reverse deps after scanning the set, to avoid invalidating the 1414 // 'ReverseDeps' reference. 1415 while (!ReverseDepsToAdd.empty()) { 1416 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1417 .insert(ReverseDepsToAdd.back().second); 1418 ReverseDepsToAdd.pop_back(); 1419 } 1420 } 1421 1422 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1423 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1424 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1425 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1426 I != E; ++I) { 1427 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1428 1429 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1430 // The information is now dirty! 1431 INLD.second = true; 1432 1433 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1434 DE = INLD.first.end(); DI != DE; ++DI) { 1435 if (DI->getResult().getInst() != RemInst) continue; 1436 1437 // Convert to a dirty entry for the subsequent instruction. 1438 DI->setResult(NewDirtyVal); 1439 1440 if (Instruction *NextI = NewDirtyVal.getInst()) 1441 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1442 } 1443 } 1444 1445 ReverseNonLocalDeps.erase(ReverseDepIt); 1446 1447 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1448 while (!ReverseDepsToAdd.empty()) { 1449 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1450 .insert(ReverseDepsToAdd.back().second); 1451 ReverseDepsToAdd.pop_back(); 1452 } 1453 } 1454 1455 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1456 // value in the NonLocalPointerDeps info. 1457 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1458 ReverseNonLocalPtrDeps.find(RemInst); 1459 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1460 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1461 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1462 1463 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1464 E = Set.end(); I != E; ++I) { 1465 ValueIsLoadPair P = *I; 1466 assert(P.getPointer() != RemInst && 1467 "Already removed NonLocalPointerDeps info for RemInst"); 1468 1469 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps; 1470 1471 // The cache is not valid for any specific block anymore. 1472 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair(); 1473 1474 // Update any entries for RemInst to use the instruction after it. 1475 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1476 DI != DE; ++DI) { 1477 if (DI->getResult().getInst() != RemInst) continue; 1478 1479 // Convert to a dirty entry for the subsequent instruction. 1480 DI->setResult(NewDirtyVal); 1481 1482 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1483 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1484 } 1485 1486 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1487 // subsequent value may invalidate the sortedness. 1488 std::sort(NLPDI.begin(), NLPDI.end()); 1489 } 1490 1491 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1492 1493 while (!ReversePtrDepsToAdd.empty()) { 1494 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1495 .insert(ReversePtrDepsToAdd.back().second); 1496 ReversePtrDepsToAdd.pop_back(); 1497 } 1498 } 1499 1500 1501 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1502 AA->deleteValue(RemInst); 1503 DEBUG(verifyRemoved(RemInst)); 1504} 1505/// verifyRemoved - Verify that the specified instruction does not occur 1506/// in our internal data structures. 1507void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1508 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1509 E = LocalDeps.end(); I != E; ++I) { 1510 assert(I->first != D && "Inst occurs in data structures"); 1511 assert(I->second.getInst() != D && 1512 "Inst occurs in data structures"); 1513 } 1514 1515 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1516 E = NonLocalPointerDeps.end(); I != E; ++I) { 1517 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1518 const NonLocalDepInfo &Val = I->second.NonLocalDeps; 1519 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1520 II != E; ++II) 1521 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1522 } 1523 1524 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1525 E = NonLocalDeps.end(); I != E; ++I) { 1526 assert(I->first != D && "Inst occurs in data structures"); 1527 const PerInstNLInfo &INLD = I->second; 1528 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1529 EE = INLD.first.end(); II != EE; ++II) 1530 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1531 } 1532 1533 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1534 E = ReverseLocalDeps.end(); I != E; ++I) { 1535 assert(I->first != D && "Inst occurs in data structures"); 1536 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1537 EE = I->second.end(); II != EE; ++II) 1538 assert(*II != D && "Inst occurs in data structures"); 1539 } 1540 1541 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1542 E = ReverseNonLocalDeps.end(); 1543 I != E; ++I) { 1544 assert(I->first != D && "Inst occurs in data structures"); 1545 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1546 EE = I->second.end(); II != EE; ++II) 1547 assert(*II != D && "Inst occurs in data structures"); 1548 } 1549 1550 for (ReverseNonLocalPtrDepTy::const_iterator 1551 I = ReverseNonLocalPtrDeps.begin(), 1552 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1553 assert(I->first != D && "Inst occurs in rev NLPD map"); 1554 1555 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1556 E = I->second.end(); II != E; ++II) 1557 assert(*II != ValueIsLoadPair(D, false) && 1558 *II != ValueIsLoadPair(D, true) && 1559 "Inst occurs in ReverseNonLocalPtrDeps map"); 1560 } 1561 1562} 1563