MemoryDependenceAnalysis.cpp revision 7d4ff60b55772c731688cd27f9252e31ae964f84
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/ADT/STLExtras.h" 20#include "llvm/ADT/Statistic.h" 21#include "llvm/Analysis/AliasAnalysis.h" 22#include "llvm/Analysis/Dominators.h" 23#include "llvm/Analysis/InstructionSimplify.h" 24#include "llvm/Analysis/MemoryBuiltins.h" 25#include "llvm/Analysis/PHITransAddr.h" 26#include "llvm/Analysis/ValueTracking.h" 27#include "llvm/IR/DataLayout.h" 28#include "llvm/IR/Function.h" 29#include "llvm/IR/Instructions.h" 30#include "llvm/IR/IntrinsicInst.h" 31#include "llvm/IR/LLVMContext.h" 32#include "llvm/Support/Debug.h" 33#include "llvm/Support/PredIteratorCache.h" 34using namespace llvm; 35 36STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 37STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 38STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 39 40STATISTIC(NumCacheNonLocalPtr, 41 "Number of fully cached non-local ptr responses"); 42STATISTIC(NumCacheDirtyNonLocalPtr, 43 "Number of cached, but dirty, non-local ptr responses"); 44STATISTIC(NumUncacheNonLocalPtr, 45 "Number of uncached non-local ptr responses"); 46STATISTIC(NumCacheCompleteNonLocalPtr, 47 "Number of block queries that were completely cached"); 48 49// Limit for the number of instructions to scan in a block. 50// FIXME: Figure out what a sane value is for this. 51// (500 is relatively insane.) 52static const int BlockScanLimit = 500; 53 54char MemoryDependenceAnalysis::ID = 0; 55 56// Register this pass... 57INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep", 58 "Memory Dependence Analysis", false, true) 59INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 60INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep", 61 "Memory Dependence Analysis", false, true) 62 63MemoryDependenceAnalysis::MemoryDependenceAnalysis() 64: FunctionPass(ID), PredCache(0) { 65 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry()); 66} 67MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { 68} 69 70/// Clean up memory in between runs 71void MemoryDependenceAnalysis::releaseMemory() { 72 LocalDeps.clear(); 73 NonLocalDeps.clear(); 74 NonLocalPointerDeps.clear(); 75 ReverseLocalDeps.clear(); 76 ReverseNonLocalDeps.clear(); 77 ReverseNonLocalPtrDeps.clear(); 78 PredCache->clear(); 79} 80 81 82 83/// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. 84/// 85void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 86 AU.setPreservesAll(); 87 AU.addRequiredTransitive<AliasAnalysis>(); 88} 89 90bool MemoryDependenceAnalysis::runOnFunction(Function &) { 91 AA = &getAnalysis<AliasAnalysis>(); 92 TD = getAnalysisIfAvailable<DataLayout>(); 93 DT = getAnalysisIfAvailable<DominatorTree>(); 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. 101template <typename KeyTy> 102static 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. 118static 119AliasAnalysis::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 } 127 if (LI->getOrdering() == Monotonic) { 128 Loc = AA->getLocation(LI); 129 return AliasAnalysis::ModRef; 130 } 131 Loc = AliasAnalysis::Location(); 132 return AliasAnalysis::ModRef; 133 } 134 135 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 136 if (SI->isUnordered()) { 137 Loc = AA->getLocation(SI); 138 return AliasAnalysis::Mod; 139 } 140 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, AA->getTargetLibraryInfo())) { 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 continue; 233 default: 234 return MemDepResult::getClobber(Inst); 235 } 236 } 237 238 // If we could not obtain a pointer for the instruction and the instruction 239 // touches memory then assume that this is a dependency. 240 if (MR != AliasAnalysis::NoModRef) 241 return MemDepResult::getClobber(Inst); 242 } 243 244 // No dependence found. If this is the entry block of the function, it is 245 // unknown, otherwise it is non-local. 246 if (BB != &BB->getParent()->getEntryBlock()) 247 return MemDepResult::getNonLocal(); 248 return MemDepResult::getNonFuncLocal(); 249} 250 251/// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that 252/// would fully overlap MemLoc if done as a wider legal integer load. 253/// 254/// MemLocBase, MemLocOffset are lazily computed here the first time the 255/// base/offs of memloc is needed. 256static bool 257isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc, 258 const Value *&MemLocBase, 259 int64_t &MemLocOffs, 260 const LoadInst *LI, 261 const DataLayout *TD) { 262 // If we have no target data, we can't do this. 263 if (TD == 0) return false; 264 265 // If we haven't already computed the base/offset of MemLoc, do so now. 266 if (MemLocBase == 0) 267 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, TD); 268 269 unsigned Size = MemoryDependenceAnalysis:: 270 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size, 271 LI, *TD); 272 return Size != 0; 273} 274 275/// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that 276/// looks at a memory location for a load (specified by MemLocBase, Offs, 277/// and Size) and compares it against a load. If the specified load could 278/// be safely widened to a larger integer load that is 1) still efficient, 279/// 2) safe for the target, and 3) would provide the specified memory 280/// location value, then this function returns the size in bytes of the 281/// load width to use. If not, this returns zero. 282unsigned MemoryDependenceAnalysis:: 283getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs, 284 unsigned MemLocSize, const LoadInst *LI, 285 const DataLayout &TD) { 286 // We can only extend simple integer loads. 287 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0; 288 289 // Load widening is hostile to ThreadSanitizer: it may cause false positives 290 // or make the reports more cryptic (access sizes are wrong). 291 if (LI->getParent()->getParent()->getAttributes(). 292 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeThread)) 293 return 0; 294 295 // Get the base of this load. 296 int64_t LIOffs = 0; 297 const Value *LIBase = 298 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, &TD); 299 300 // If the two pointers are not based on the same pointer, we can't tell that 301 // they are related. 302 if (LIBase != MemLocBase) return 0; 303 304 // Okay, the two values are based on the same pointer, but returned as 305 // no-alias. This happens when we have things like two byte loads at "P+1" 306 // and "P+3". Check to see if increasing the size of the "LI" load up to its 307 // alignment (or the largest native integer type) will allow us to load all 308 // the bits required by MemLoc. 309 310 // If MemLoc is before LI, then no widening of LI will help us out. 311 if (MemLocOffs < LIOffs) return 0; 312 313 // Get the alignment of the load in bytes. We assume that it is safe to load 314 // any legal integer up to this size without a problem. For example, if we're 315 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can 316 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it 317 // to i16. 318 unsigned LoadAlign = LI->getAlignment(); 319 320 int64_t MemLocEnd = MemLocOffs+MemLocSize; 321 322 // If no amount of rounding up will let MemLoc fit into LI, then bail out. 323 if (LIOffs+LoadAlign < MemLocEnd) return 0; 324 325 // This is the size of the load to try. Start with the next larger power of 326 // two. 327 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U; 328 NewLoadByteSize = NextPowerOf2(NewLoadByteSize); 329 330 while (1) { 331 // If this load size is bigger than our known alignment or would not fit 332 // into a native integer register, then we fail. 333 if (NewLoadByteSize > LoadAlign || 334 !TD.fitsInLegalInteger(NewLoadByteSize*8)) 335 return 0; 336 337 if (LIOffs+NewLoadByteSize > MemLocEnd && 338 LI->getParent()->getParent()->getAttributes(). 339 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeAddress)) 340 // We will be reading past the location accessed by the original program. 341 // While this is safe in a regular build, Address Safety analysis tools 342 // may start reporting false warnings. So, don't do widening. 343 return 0; 344 345 // If a load of this width would include all of MemLoc, then we succeed. 346 if (LIOffs+NewLoadByteSize >= MemLocEnd) 347 return NewLoadByteSize; 348 349 NewLoadByteSize <<= 1; 350 } 351} 352 353/// getPointerDependencyFrom - Return the instruction on which a memory 354/// location depends. If isLoad is true, this routine ignores may-aliases with 355/// read-only operations. If isLoad is false, this routine ignores may-aliases 356/// with reads from read-only locations. If possible, pass the query 357/// instruction as well; this function may take advantage of the metadata 358/// annotated to the query instruction to refine the result. 359MemDepResult MemoryDependenceAnalysis:: 360getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad, 361 BasicBlock::iterator ScanIt, BasicBlock *BB, 362 Instruction *QueryInst) { 363 364 const Value *MemLocBase = 0; 365 int64_t MemLocOffset = 0; 366 unsigned Limit = BlockScanLimit; 367 bool isInvariantLoad = false; 368 if (isLoad && QueryInst) { 369 LoadInst *LI = dyn_cast<LoadInst>(QueryInst); 370 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != 0) 371 isInvariantLoad = true; 372 } 373 374 // Walk backwards through the basic block, looking for dependencies. 375 while (ScanIt != BB->begin()) { 376 // Limit the amount of scanning we do so we don't end up with quadratic 377 // running time on extreme testcases. 378 --Limit; 379 if (!Limit) 380 return MemDepResult::getUnknown(); 381 382 Instruction *Inst = --ScanIt; 383 384 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 385 // Debug intrinsics don't (and can't) cause dependences. 386 if (isa<DbgInfoIntrinsic>(II)) continue; 387 388 // If we reach a lifetime begin or end marker, then the query ends here 389 // because the value is undefined. 390 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 391 // FIXME: This only considers queries directly on the invariant-tagged 392 // pointer, not on query pointers that are indexed off of them. It'd 393 // be nice to handle that at some point (the right approach is to use 394 // GetPointerBaseWithConstantOffset). 395 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)), 396 MemLoc)) 397 return MemDepResult::getDef(II); 398 continue; 399 } 400 } 401 402 // Values depend on loads if the pointers are must aliased. This means that 403 // a load depends on another must aliased load from the same value. 404 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 405 // Atomic loads have complications involved. 406 // FIXME: This is overly conservative. 407 if (!LI->isUnordered()) 408 return MemDepResult::getClobber(LI); 409 410 AliasAnalysis::Location LoadLoc = AA->getLocation(LI); 411 412 // If we found a pointer, check if it could be the same as our pointer. 413 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc); 414 415 if (isLoad) { 416 if (R == AliasAnalysis::NoAlias) { 417 // If this is an over-aligned integer load (for example, 418 // "load i8* %P, align 4") see if it would obviously overlap with the 419 // queried location if widened to a larger load (e.g. if the queried 420 // location is 1 byte at P+1). If so, return it as a load/load 421 // clobber result, allowing the client to decide to widen the load if 422 // it wants to. 423 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) 424 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() && 425 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase, 426 MemLocOffset, LI, TD)) 427 return MemDepResult::getClobber(Inst); 428 429 continue; 430 } 431 432 // Must aliased loads are defs of each other. 433 if (R == AliasAnalysis::MustAlias) 434 return MemDepResult::getDef(Inst); 435 436#if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads 437 // in terms of clobbering loads, but since it does this by looking 438 // at the clobbering load directly, it doesn't know about any 439 // phi translation that may have happened along the way. 440 441 // If we have a partial alias, then return this as a clobber for the 442 // client to handle. 443 if (R == AliasAnalysis::PartialAlias) 444 return MemDepResult::getClobber(Inst); 445#endif 446 447 // Random may-alias loads don't depend on each other without a 448 // dependence. 449 continue; 450 } 451 452 // Stores don't depend on other no-aliased accesses. 453 if (R == AliasAnalysis::NoAlias) 454 continue; 455 456 // Stores don't alias loads from read-only memory. 457 if (AA->pointsToConstantMemory(LoadLoc)) 458 continue; 459 460 // Stores depend on may/must aliased loads. 461 return MemDepResult::getDef(Inst); 462 } 463 464 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 465 // Atomic stores have complications involved. 466 // FIXME: This is overly conservative. 467 if (!SI->isUnordered()) 468 return MemDepResult::getClobber(SI); 469 470 // If alias analysis can tell that this store is guaranteed to not modify 471 // the query pointer, ignore it. Use getModRefInfo to handle cases where 472 // the query pointer points to constant memory etc. 473 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef) 474 continue; 475 476 // Ok, this store might clobber the query pointer. Check to see if it is 477 // a must alias: in this case, we want to return this as a def. 478 AliasAnalysis::Location StoreLoc = AA->getLocation(SI); 479 480 // If we found a pointer, check if it could be the same as our pointer. 481 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc); 482 483 if (R == AliasAnalysis::NoAlias) 484 continue; 485 if (R == AliasAnalysis::MustAlias) 486 return MemDepResult::getDef(Inst); 487 if (isInvariantLoad) 488 continue; 489 return MemDepResult::getClobber(Inst); 490 } 491 492 // If this is an allocation, and if we know that the accessed pointer is to 493 // the allocation, return Def. This means that there is no dependence and 494 // the access can be optimized based on that. For example, a load could 495 // turn into undef. 496 // Note: Only determine this to be a malloc if Inst is the malloc call, not 497 // a subsequent bitcast of the malloc call result. There can be stores to 498 // the malloced memory between the malloc call and its bitcast uses, and we 499 // need to continue scanning until the malloc call. 500 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo(); 501 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) { 502 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD); 503 504 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr)) 505 return MemDepResult::getDef(Inst); 506 // Be conservative if the accessed pointer may alias the allocation. 507 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias) 508 return MemDepResult::getClobber(Inst); 509 // If the allocation is not aliased and does not read memory (like 510 // strdup), it is safe to ignore. 511 if (isa<AllocaInst>(Inst) || 512 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI)) 513 continue; 514 } 515 516 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 517 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc); 518 // If necessary, perform additional analysis. 519 if (MR == AliasAnalysis::ModRef) 520 MR = AA->callCapturesBefore(Inst, MemLoc, DT); 521 switch (MR) { 522 case AliasAnalysis::NoModRef: 523 // If the call has no effect on the queried pointer, just ignore it. 524 continue; 525 case AliasAnalysis::Mod: 526 return MemDepResult::getClobber(Inst); 527 case AliasAnalysis::Ref: 528 // If the call is known to never store to the pointer, and if this is a 529 // load query, we can safely ignore it (scan past it). 530 if (isLoad) 531 continue; 532 default: 533 // Otherwise, there is a potential dependence. Return a clobber. 534 return MemDepResult::getClobber(Inst); 535 } 536 } 537 538 // No dependence found. If this is the entry block of the function, it is 539 // unknown, otherwise it is non-local. 540 if (BB != &BB->getParent()->getEntryBlock()) 541 return MemDepResult::getNonLocal(); 542 return MemDepResult::getNonFuncLocal(); 543} 544 545/// getDependency - Return the instruction on which a memory operation 546/// depends. 547MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 548 Instruction *ScanPos = QueryInst; 549 550 // Check for a cached result 551 MemDepResult &LocalCache = LocalDeps[QueryInst]; 552 553 // If the cached entry is non-dirty, just return it. Note that this depends 554 // on MemDepResult's default constructing to 'dirty'. 555 if (!LocalCache.isDirty()) 556 return LocalCache; 557 558 // Otherwise, if we have a dirty entry, we know we can start the scan at that 559 // instruction, which may save us some work. 560 if (Instruction *Inst = LocalCache.getInst()) { 561 ScanPos = Inst; 562 563 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 564 } 565 566 BasicBlock *QueryParent = QueryInst->getParent(); 567 568 // Do the scan. 569 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 570 // No dependence found. If this is the entry block of the function, it is 571 // unknown, otherwise it is non-local. 572 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 573 LocalCache = MemDepResult::getNonLocal(); 574 else 575 LocalCache = MemDepResult::getNonFuncLocal(); 576 } else { 577 AliasAnalysis::Location MemLoc; 578 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA); 579 if (MemLoc.Ptr) { 580 // If we can do a pointer scan, make it happen. 581 bool isLoad = !(MR & AliasAnalysis::Mod); 582 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst)) 583 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start; 584 585 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos, 586 QueryParent, QueryInst); 587 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 588 CallSite QueryCS(QueryInst); 589 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 590 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 591 QueryParent); 592 } else 593 // Non-memory instruction. 594 LocalCache = MemDepResult::getUnknown(); 595 } 596 597 // Remember the result! 598 if (Instruction *I = LocalCache.getInst()) 599 ReverseLocalDeps[I].insert(QueryInst); 600 601 return LocalCache; 602} 603 604#ifndef NDEBUG 605/// AssertSorted - This method is used when -debug is specified to verify that 606/// cache arrays are properly kept sorted. 607static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 608 int Count = -1) { 609 if (Count == -1) Count = Cache.size(); 610 if (Count == 0) return; 611 612 for (unsigned i = 1; i != unsigned(Count); ++i) 613 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 614} 615#endif 616 617/// getNonLocalCallDependency - Perform a full dependency query for the 618/// specified call, returning the set of blocks that the value is 619/// potentially live across. The returned set of results will include a 620/// "NonLocal" result for all blocks where the value is live across. 621/// 622/// This method assumes the instruction returns a "NonLocal" dependency 623/// within its own block. 624/// 625/// This returns a reference to an internal data structure that may be 626/// invalidated on the next non-local query or when an instruction is 627/// removed. Clients must copy this data if they want it around longer than 628/// that. 629const MemoryDependenceAnalysis::NonLocalDepInfo & 630MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 631 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 632 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 633 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 634 NonLocalDepInfo &Cache = CacheP.first; 635 636 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 637 /// the cached case, this can happen due to instructions being deleted etc. In 638 /// the uncached case, this starts out as the set of predecessors we care 639 /// about. 640 SmallVector<BasicBlock*, 32> DirtyBlocks; 641 642 if (!Cache.empty()) { 643 // Okay, we have a cache entry. If we know it is not dirty, just return it 644 // with no computation. 645 if (!CacheP.second) { 646 ++NumCacheNonLocal; 647 return Cache; 648 } 649 650 // If we already have a partially computed set of results, scan them to 651 // determine what is dirty, seeding our initial DirtyBlocks worklist. 652 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 653 I != E; ++I) 654 if (I->getResult().isDirty()) 655 DirtyBlocks.push_back(I->getBB()); 656 657 // Sort the cache so that we can do fast binary search lookups below. 658 std::sort(Cache.begin(), Cache.end()); 659 660 ++NumCacheDirtyNonLocal; 661 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 662 // << Cache.size() << " cached: " << *QueryInst; 663 } else { 664 // Seed DirtyBlocks with each of the preds of QueryInst's block. 665 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 666 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 667 DirtyBlocks.push_back(*PI); 668 ++NumUncacheNonLocal; 669 } 670 671 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 672 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 673 674 SmallPtrSet<BasicBlock*, 64> Visited; 675 676 unsigned NumSortedEntries = Cache.size(); 677 DEBUG(AssertSorted(Cache)); 678 679 // Iterate while we still have blocks to update. 680 while (!DirtyBlocks.empty()) { 681 BasicBlock *DirtyBB = DirtyBlocks.back(); 682 DirtyBlocks.pop_back(); 683 684 // Already processed this block? 685 if (!Visited.insert(DirtyBB)) 686 continue; 687 688 // Do a binary search to see if we already have an entry for this block in 689 // the cache set. If so, find it. 690 DEBUG(AssertSorted(Cache, NumSortedEntries)); 691 NonLocalDepInfo::iterator Entry = 692 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 693 NonLocalDepEntry(DirtyBB)); 694 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB) 695 --Entry; 696 697 NonLocalDepEntry *ExistingResult = 0; 698 if (Entry != Cache.begin()+NumSortedEntries && 699 Entry->getBB() == DirtyBB) { 700 // If we already have an entry, and if it isn't already dirty, the block 701 // is done. 702 if (!Entry->getResult().isDirty()) 703 continue; 704 705 // Otherwise, remember this slot so we can update the value. 706 ExistingResult = &*Entry; 707 } 708 709 // If the dirty entry has a pointer, start scanning from it so we don't have 710 // to rescan the entire block. 711 BasicBlock::iterator ScanPos = DirtyBB->end(); 712 if (ExistingResult) { 713 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 714 ScanPos = Inst; 715 // We're removing QueryInst's use of Inst. 716 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 717 QueryCS.getInstruction()); 718 } 719 } 720 721 // Find out if this block has a local dependency for QueryInst. 722 MemDepResult Dep; 723 724 if (ScanPos != DirtyBB->begin()) { 725 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 726 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 727 // No dependence found. If this is the entry block of the function, it is 728 // a clobber, otherwise it is unknown. 729 Dep = MemDepResult::getNonLocal(); 730 } else { 731 Dep = MemDepResult::getNonFuncLocal(); 732 } 733 734 // If we had a dirty entry for the block, update it. Otherwise, just add 735 // a new entry. 736 if (ExistingResult) 737 ExistingResult->setResult(Dep); 738 else 739 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 740 741 // If the block has a dependency (i.e. it isn't completely transparent to 742 // the value), remember the association! 743 if (!Dep.isNonLocal()) { 744 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 745 // update this when we remove instructions. 746 if (Instruction *Inst = Dep.getInst()) 747 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 748 } else { 749 750 // If the block *is* completely transparent to the load, we need to check 751 // the predecessors of this block. Add them to our worklist. 752 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 753 DirtyBlocks.push_back(*PI); 754 } 755 } 756 757 return Cache; 758} 759 760/// getNonLocalPointerDependency - Perform a full dependency query for an 761/// access to the specified (non-volatile) memory location, returning the 762/// set of instructions that either define or clobber the value. 763/// 764/// This method assumes the pointer has a "NonLocal" dependency within its 765/// own block. 766/// 767void MemoryDependenceAnalysis:: 768getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad, 769 BasicBlock *FromBB, 770 SmallVectorImpl<NonLocalDepResult> &Result) { 771 assert(Loc.Ptr->getType()->isPointerTy() && 772 "Can't get pointer deps of a non-pointer!"); 773 Result.clear(); 774 775 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD); 776 777 // This is the set of blocks we've inspected, and the pointer we consider in 778 // each block. Because of critical edges, we currently bail out if querying 779 // a block with multiple different pointers. This can happen during PHI 780 // translation. 781 DenseMap<BasicBlock*, Value*> Visited; 782 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB, 783 Result, Visited, true)) 784 return; 785 Result.clear(); 786 Result.push_back(NonLocalDepResult(FromBB, 787 MemDepResult::getUnknown(), 788 const_cast<Value *>(Loc.Ptr))); 789} 790 791/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 792/// Pointer/PointeeSize using either cached information in Cache or by doing a 793/// lookup (which may use dirty cache info if available). If we do a lookup, 794/// add the result to the cache. 795MemDepResult MemoryDependenceAnalysis:: 796GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc, 797 bool isLoad, BasicBlock *BB, 798 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 799 800 // Do a binary search to see if we already have an entry for this block in 801 // the cache set. If so, find it. 802 NonLocalDepInfo::iterator Entry = 803 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 804 NonLocalDepEntry(BB)); 805 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 806 --Entry; 807 808 NonLocalDepEntry *ExistingResult = 0; 809 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 810 ExistingResult = &*Entry; 811 812 // If we have a cached entry, and it is non-dirty, use it as the value for 813 // this dependency. 814 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 815 ++NumCacheNonLocalPtr; 816 return ExistingResult->getResult(); 817 } 818 819 // Otherwise, we have to scan for the value. If we have a dirty cache 820 // entry, start scanning from its position, otherwise we scan from the end 821 // of the block. 822 BasicBlock::iterator ScanPos = BB->end(); 823 if (ExistingResult && ExistingResult->getResult().getInst()) { 824 assert(ExistingResult->getResult().getInst()->getParent() == BB && 825 "Instruction invalidated?"); 826 ++NumCacheDirtyNonLocalPtr; 827 ScanPos = ExistingResult->getResult().getInst(); 828 829 // Eliminating the dirty entry from 'Cache', so update the reverse info. 830 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 831 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 832 } else { 833 ++NumUncacheNonLocalPtr; 834 } 835 836 // Scan the block for the dependency. 837 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB); 838 839 // If we had a dirty entry for the block, update it. Otherwise, just add 840 // a new entry. 841 if (ExistingResult) 842 ExistingResult->setResult(Dep); 843 else 844 Cache->push_back(NonLocalDepEntry(BB, Dep)); 845 846 // If the block has a dependency (i.e. it isn't completely transparent to 847 // the value), remember the reverse association because we just added it 848 // to Cache! 849 if (!Dep.isDef() && !Dep.isClobber()) 850 return Dep; 851 852 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 853 // update MemDep when we remove instructions. 854 Instruction *Inst = Dep.getInst(); 855 assert(Inst && "Didn't depend on anything?"); 856 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 857 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 858 return Dep; 859} 860 861/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 862/// number of elements in the array that are already properly ordered. This is 863/// optimized for the case when only a few entries are added. 864static void 865SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 866 unsigned NumSortedEntries) { 867 switch (Cache.size() - NumSortedEntries) { 868 case 0: 869 // done, no new entries. 870 break; 871 case 2: { 872 // Two new entries, insert the last one into place. 873 NonLocalDepEntry Val = Cache.back(); 874 Cache.pop_back(); 875 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 876 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 877 Cache.insert(Entry, Val); 878 // FALL THROUGH. 879 } 880 case 1: 881 // One new entry, Just insert the new value at the appropriate position. 882 if (Cache.size() != 1) { 883 NonLocalDepEntry Val = Cache.back(); 884 Cache.pop_back(); 885 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 886 std::upper_bound(Cache.begin(), Cache.end(), Val); 887 Cache.insert(Entry, Val); 888 } 889 break; 890 default: 891 // Added many values, do a full scale sort. 892 std::sort(Cache.begin(), Cache.end()); 893 break; 894 } 895} 896 897/// getNonLocalPointerDepFromBB - Perform a dependency query based on 898/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 899/// results to the results vector and keep track of which blocks are visited in 900/// 'Visited'. 901/// 902/// This has special behavior for the first block queries (when SkipFirstBlock 903/// is true). In this special case, it ignores the contents of the specified 904/// block and starts returning dependence info for its predecessors. 905/// 906/// This function returns false on success, or true to indicate that it could 907/// not compute dependence information for some reason. This should be treated 908/// as a clobber dependence on the first instruction in the predecessor block. 909bool MemoryDependenceAnalysis:: 910getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, 911 const AliasAnalysis::Location &Loc, 912 bool isLoad, BasicBlock *StartBB, 913 SmallVectorImpl<NonLocalDepResult> &Result, 914 DenseMap<BasicBlock*, Value*> &Visited, 915 bool SkipFirstBlock) { 916 917 // Look up the cached info for Pointer. 918 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 919 920 // Set up a temporary NLPI value. If the map doesn't yet have an entry for 921 // CacheKey, this value will be inserted as the associated value. Otherwise, 922 // it'll be ignored, and we'll have to check to see if the cached size and 923 // tbaa tag are consistent with the current query. 924 NonLocalPointerInfo InitialNLPI; 925 InitialNLPI.Size = Loc.Size; 926 InitialNLPI.TBAATag = Loc.TBAATag; 927 928 // Get the NLPI for CacheKey, inserting one into the map if it doesn't 929 // already have one. 930 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair = 931 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI)); 932 NonLocalPointerInfo *CacheInfo = &Pair.first->second; 933 934 // If we already have a cache entry for this CacheKey, we may need to do some 935 // work to reconcile the cache entry and the current query. 936 if (!Pair.second) { 937 if (CacheInfo->Size < Loc.Size) { 938 // The query's Size is greater than the cached one. Throw out the 939 // cached data and proceed with the query at the greater size. 940 CacheInfo->Pair = BBSkipFirstBlockPair(); 941 CacheInfo->Size = Loc.Size; 942 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 943 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 944 if (Instruction *Inst = DI->getResult().getInst()) 945 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 946 CacheInfo->NonLocalDeps.clear(); 947 } else if (CacheInfo->Size > Loc.Size) { 948 // This query's Size is less than the cached one. Conservatively restart 949 // the query using the greater size. 950 return getNonLocalPointerDepFromBB(Pointer, 951 Loc.getWithNewSize(CacheInfo->Size), 952 isLoad, StartBB, Result, Visited, 953 SkipFirstBlock); 954 } 955 956 // If the query's TBAATag is inconsistent with the cached one, 957 // conservatively throw out the cached data and restart the query with 958 // no tag if needed. 959 if (CacheInfo->TBAATag != Loc.TBAATag) { 960 if (CacheInfo->TBAATag) { 961 CacheInfo->Pair = BBSkipFirstBlockPair(); 962 CacheInfo->TBAATag = 0; 963 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 964 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 965 if (Instruction *Inst = DI->getResult().getInst()) 966 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 967 CacheInfo->NonLocalDeps.clear(); 968 } 969 if (Loc.TBAATag) 970 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(), 971 isLoad, StartBB, Result, Visited, 972 SkipFirstBlock); 973 } 974 } 975 976 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps; 977 978 // If we have valid cached information for exactly the block we are 979 // investigating, just return it with no recomputation. 980 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 981 // We have a fully cached result for this query then we can just return the 982 // cached results and populate the visited set. However, we have to verify 983 // that we don't already have conflicting results for these blocks. Check 984 // to ensure that if a block in the results set is in the visited set that 985 // it was for the same pointer query. 986 if (!Visited.empty()) { 987 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 988 I != E; ++I) { 989 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 990 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 991 continue; 992 993 // We have a pointer mismatch in a block. Just return clobber, saying 994 // that something was clobbered in this result. We could also do a 995 // non-fully cached query, but there is little point in doing this. 996 return true; 997 } 998 } 999 1000 Value *Addr = Pointer.getAddr(); 1001 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 1002 I != E; ++I) { 1003 Visited.insert(std::make_pair(I->getBB(), Addr)); 1004 if (!I->getResult().isNonLocal() && DT->isReachableFromEntry(I->getBB())) 1005 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 1006 } 1007 ++NumCacheCompleteNonLocalPtr; 1008 return false; 1009 } 1010 1011 // Otherwise, either this is a new block, a block with an invalid cache 1012 // pointer or one that we're about to invalidate by putting more info into it 1013 // than its valid cache info. If empty, the result will be valid cache info, 1014 // otherwise it isn't. 1015 if (Cache->empty()) 1016 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 1017 else 1018 CacheInfo->Pair = BBSkipFirstBlockPair(); 1019 1020 SmallVector<BasicBlock*, 32> Worklist; 1021 Worklist.push_back(StartBB); 1022 1023 // PredList used inside loop. 1024 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList; 1025 1026 // Keep track of the entries that we know are sorted. Previously cached 1027 // entries will all be sorted. The entries we add we only sort on demand (we 1028 // don't insert every element into its sorted position). We know that we 1029 // won't get any reuse from currently inserted values, because we don't 1030 // revisit blocks after we insert info for them. 1031 unsigned NumSortedEntries = Cache->size(); 1032 DEBUG(AssertSorted(*Cache)); 1033 1034 while (!Worklist.empty()) { 1035 BasicBlock *BB = Worklist.pop_back_val(); 1036 1037 // Skip the first block if we have it. 1038 if (!SkipFirstBlock) { 1039 // Analyze the dependency of *Pointer in FromBB. See if we already have 1040 // been here. 1041 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 1042 1043 // Get the dependency info for Pointer in BB. If we have cached 1044 // information, we will use it, otherwise we compute it. 1045 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 1046 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache, 1047 NumSortedEntries); 1048 1049 // If we got a Def or Clobber, add this to the list of results. 1050 if (!Dep.isNonLocal() && DT->isReachableFromEntry(BB)) { 1051 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 1052 continue; 1053 } 1054 } 1055 1056 // If 'Pointer' is an instruction defined in this block, then we need to do 1057 // phi translation to change it into a value live in the predecessor block. 1058 // If not, we just add the predecessors to the worklist and scan them with 1059 // the same Pointer. 1060 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 1061 SkipFirstBlock = false; 1062 SmallVector<BasicBlock*, 16> NewBlocks; 1063 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1064 // Verify that we haven't looked at this block yet. 1065 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1066 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 1067 if (InsertRes.second) { 1068 // First time we've looked at *PI. 1069 NewBlocks.push_back(*PI); 1070 continue; 1071 } 1072 1073 // If we have seen this block before, but it was with a different 1074 // pointer then we have a phi translation failure and we have to treat 1075 // this as a clobber. 1076 if (InsertRes.first->second != Pointer.getAddr()) { 1077 // Make sure to clean up the Visited map before continuing on to 1078 // PredTranslationFailure. 1079 for (unsigned i = 0; i < NewBlocks.size(); i++) 1080 Visited.erase(NewBlocks[i]); 1081 goto PredTranslationFailure; 1082 } 1083 } 1084 Worklist.append(NewBlocks.begin(), NewBlocks.end()); 1085 continue; 1086 } 1087 1088 // We do need to do phi translation, if we know ahead of time we can't phi 1089 // translate this value, don't even try. 1090 if (!Pointer.IsPotentiallyPHITranslatable()) 1091 goto PredTranslationFailure; 1092 1093 // We may have added values to the cache list before this PHI translation. 1094 // If so, we haven't done anything to ensure that the cache remains sorted. 1095 // Sort it now (if needed) so that recursive invocations of 1096 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 1097 // value will only see properly sorted cache arrays. 1098 if (Cache && NumSortedEntries != Cache->size()) { 1099 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1100 NumSortedEntries = Cache->size(); 1101 } 1102 Cache = 0; 1103 1104 PredList.clear(); 1105 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1106 BasicBlock *Pred = *PI; 1107 PredList.push_back(std::make_pair(Pred, Pointer)); 1108 1109 // Get the PHI translated pointer in this predecessor. This can fail if 1110 // not translatable, in which case the getAddr() returns null. 1111 PHITransAddr &PredPointer = PredList.back().second; 1112 PredPointer.PHITranslateValue(BB, Pred, 0); 1113 1114 Value *PredPtrVal = PredPointer.getAddr(); 1115 1116 // Check to see if we have already visited this pred block with another 1117 // pointer. If so, we can't do this lookup. This failure can occur 1118 // with PHI translation when a critical edge exists and the PHI node in 1119 // the successor translates to a pointer value different than the 1120 // pointer the block was first analyzed with. 1121 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1122 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 1123 1124 if (!InsertRes.second) { 1125 // We found the pred; take it off the list of preds to visit. 1126 PredList.pop_back(); 1127 1128 // If the predecessor was visited with PredPtr, then we already did 1129 // the analysis and can ignore it. 1130 if (InsertRes.first->second == PredPtrVal) 1131 continue; 1132 1133 // Otherwise, the block was previously analyzed with a different 1134 // pointer. We can't represent the result of this case, so we just 1135 // treat this as a phi translation failure. 1136 1137 // Make sure to clean up the Visited map before continuing on to 1138 // PredTranslationFailure. 1139 for (unsigned i = 0, n = PredList.size(); i < n; ++i) 1140 Visited.erase(PredList[i].first); 1141 1142 goto PredTranslationFailure; 1143 } 1144 } 1145 1146 // Actually process results here; this need to be a separate loop to avoid 1147 // calling getNonLocalPointerDepFromBB for blocks we don't want to return 1148 // any results for. (getNonLocalPointerDepFromBB will modify our 1149 // datastructures in ways the code after the PredTranslationFailure label 1150 // doesn't expect.) 1151 for (unsigned i = 0, n = PredList.size(); i < n; ++i) { 1152 BasicBlock *Pred = PredList[i].first; 1153 PHITransAddr &PredPointer = PredList[i].second; 1154 Value *PredPtrVal = PredPointer.getAddr(); 1155 1156 bool CanTranslate = true; 1157 // If PHI translation was unable to find an available pointer in this 1158 // predecessor, then we have to assume that the pointer is clobbered in 1159 // that predecessor. We can still do PRE of the load, which would insert 1160 // a computation of the pointer in this predecessor. 1161 if (PredPtrVal == 0) 1162 CanTranslate = false; 1163 1164 // FIXME: it is entirely possible that PHI translating will end up with 1165 // the same value. Consider PHI translating something like: 1166 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 1167 // to recurse here, pedantically speaking. 1168 1169 // If getNonLocalPointerDepFromBB fails here, that means the cached 1170 // result conflicted with the Visited list; we have to conservatively 1171 // assume it is unknown, but this also does not block PRE of the load. 1172 if (!CanTranslate || 1173 getNonLocalPointerDepFromBB(PredPointer, 1174 Loc.getWithNewPtr(PredPtrVal), 1175 isLoad, Pred, 1176 Result, Visited)) { 1177 // Add the entry to the Result list. 1178 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal); 1179 Result.push_back(Entry); 1180 1181 // Since we had a phi translation failure, the cache for CacheKey won't 1182 // include all of the entries that we need to immediately satisfy future 1183 // queries. Mark this in NonLocalPointerDeps by setting the 1184 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 1185 // cached value to do more work but not miss the phi trans failure. 1186 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey]; 1187 NLPI.Pair = BBSkipFirstBlockPair(); 1188 continue; 1189 } 1190 } 1191 1192 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 1193 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1194 Cache = &CacheInfo->NonLocalDeps; 1195 NumSortedEntries = Cache->size(); 1196 1197 // Since we did phi translation, the "Cache" set won't contain all of the 1198 // results for the query. This is ok (we can still use it to accelerate 1199 // specific block queries) but we can't do the fastpath "return all 1200 // results from the set" Clear out the indicator for this. 1201 CacheInfo->Pair = BBSkipFirstBlockPair(); 1202 SkipFirstBlock = false; 1203 continue; 1204 1205 PredTranslationFailure: 1206 // The following code is "failure"; we can't produce a sane translation 1207 // for the given block. It assumes that we haven't modified any of 1208 // our datastructures while processing the current block. 1209 1210 if (Cache == 0) { 1211 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1212 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1213 Cache = &CacheInfo->NonLocalDeps; 1214 NumSortedEntries = Cache->size(); 1215 } 1216 1217 // Since we failed phi translation, the "Cache" set won't contain all of the 1218 // results for the query. This is ok (we can still use it to accelerate 1219 // specific block queries) but we can't do the fastpath "return all 1220 // results from the set". Clear out the indicator for this. 1221 CacheInfo->Pair = BBSkipFirstBlockPair(); 1222 1223 // If *nothing* works, mark the pointer as unknown. 1224 // 1225 // If this is the magic first block, return this as a clobber of the whole 1226 // incoming value. Since we can't phi translate to one of the predecessors, 1227 // we have to bail out. 1228 if (SkipFirstBlock) 1229 return true; 1230 1231 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 1232 assert(I != Cache->rend() && "Didn't find current block??"); 1233 if (I->getBB() != BB) 1234 continue; 1235 1236 assert(I->getResult().isNonLocal() && 1237 "Should only be here with transparent block"); 1238 I->setResult(MemDepResult::getUnknown()); 1239 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 1240 Pointer.getAddr())); 1241 break; 1242 } 1243 } 1244 1245 // Okay, we're done now. If we added new values to the cache, re-sort it. 1246 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1247 DEBUG(AssertSorted(*Cache)); 1248 return false; 1249} 1250 1251/// RemoveCachedNonLocalPointerDependencies - If P exists in 1252/// CachedNonLocalPointerInfo, remove it. 1253void MemoryDependenceAnalysis:: 1254RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 1255 CachedNonLocalPointerInfo::iterator It = 1256 NonLocalPointerDeps.find(P); 1257 if (It == NonLocalPointerDeps.end()) return; 1258 1259 // Remove all of the entries in the BB->val map. This involves removing 1260 // instructions from the reverse map. 1261 NonLocalDepInfo &PInfo = It->second.NonLocalDeps; 1262 1263 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 1264 Instruction *Target = PInfo[i].getResult().getInst(); 1265 if (Target == 0) continue; // Ignore non-local dep results. 1266 assert(Target->getParent() == PInfo[i].getBB()); 1267 1268 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1269 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1270 } 1271 1272 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1273 NonLocalPointerDeps.erase(It); 1274} 1275 1276 1277/// invalidateCachedPointerInfo - This method is used to invalidate cached 1278/// information about the specified pointer, because it may be too 1279/// conservative in memdep. This is an optional call that can be used when 1280/// the client detects an equivalence between the pointer and some other 1281/// value and replaces the other value with ptr. This can make Ptr available 1282/// in more places that cached info does not necessarily keep. 1283void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1284 // If Ptr isn't really a pointer, just ignore it. 1285 if (!Ptr->getType()->isPointerTy()) return; 1286 // Flush store info for the pointer. 1287 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1288 // Flush load info for the pointer. 1289 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1290} 1291 1292/// invalidateCachedPredecessors - Clear the PredIteratorCache info. 1293/// This needs to be done when the CFG changes, e.g., due to splitting 1294/// critical edges. 1295void MemoryDependenceAnalysis::invalidateCachedPredecessors() { 1296 PredCache->clear(); 1297} 1298 1299/// removeInstruction - Remove an instruction from the dependence analysis, 1300/// updating the dependence of instructions that previously depended on it. 1301/// This method attempts to keep the cache coherent using the reverse map. 1302void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1303 // Walk through the Non-local dependencies, removing this one as the value 1304 // for any cached queries. 1305 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1306 if (NLDI != NonLocalDeps.end()) { 1307 NonLocalDepInfo &BlockMap = NLDI->second.first; 1308 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1309 DI != DE; ++DI) 1310 if (Instruction *Inst = DI->getResult().getInst()) 1311 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1312 NonLocalDeps.erase(NLDI); 1313 } 1314 1315 // If we have a cached local dependence query for this instruction, remove it. 1316 // 1317 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1318 if (LocalDepEntry != LocalDeps.end()) { 1319 // Remove us from DepInst's reverse set now that the local dep info is gone. 1320 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1321 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1322 1323 // Remove this local dependency info. 1324 LocalDeps.erase(LocalDepEntry); 1325 } 1326 1327 // If we have any cached pointer dependencies on this instruction, remove 1328 // them. If the instruction has non-pointer type, then it can't be a pointer 1329 // base. 1330 1331 // Remove it from both the load info and the store info. The instruction 1332 // can't be in either of these maps if it is non-pointer. 1333 if (RemInst->getType()->isPointerTy()) { 1334 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1335 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1336 } 1337 1338 // Loop over all of the things that depend on the instruction we're removing. 1339 // 1340 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1341 1342 // If we find RemInst as a clobber or Def in any of the maps for other values, 1343 // we need to replace its entry with a dirty version of the instruction after 1344 // it. If RemInst is a terminator, we use a null dirty value. 1345 // 1346 // Using a dirty version of the instruction after RemInst saves having to scan 1347 // the entire block to get to this point. 1348 MemDepResult NewDirtyVal; 1349 if (!RemInst->isTerminator()) 1350 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1351 1352 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1353 if (ReverseDepIt != ReverseLocalDeps.end()) { 1354 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1355 // RemInst can't be the terminator if it has local stuff depending on it. 1356 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1357 "Nothing can locally depend on a terminator"); 1358 1359 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1360 E = ReverseDeps.end(); I != E; ++I) { 1361 Instruction *InstDependingOnRemInst = *I; 1362 assert(InstDependingOnRemInst != RemInst && 1363 "Already removed our local dep info"); 1364 1365 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1366 1367 // Make sure to remember that new things depend on NewDepInst. 1368 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1369 "a local dep on this if it is a terminator!"); 1370 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1371 InstDependingOnRemInst)); 1372 } 1373 1374 ReverseLocalDeps.erase(ReverseDepIt); 1375 1376 // Add new reverse deps after scanning the set, to avoid invalidating the 1377 // 'ReverseDeps' reference. 1378 while (!ReverseDepsToAdd.empty()) { 1379 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1380 .insert(ReverseDepsToAdd.back().second); 1381 ReverseDepsToAdd.pop_back(); 1382 } 1383 } 1384 1385 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1386 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1387 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1388 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1389 I != E; ++I) { 1390 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1391 1392 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1393 // The information is now dirty! 1394 INLD.second = true; 1395 1396 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1397 DE = INLD.first.end(); DI != DE; ++DI) { 1398 if (DI->getResult().getInst() != RemInst) continue; 1399 1400 // Convert to a dirty entry for the subsequent instruction. 1401 DI->setResult(NewDirtyVal); 1402 1403 if (Instruction *NextI = NewDirtyVal.getInst()) 1404 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1405 } 1406 } 1407 1408 ReverseNonLocalDeps.erase(ReverseDepIt); 1409 1410 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1411 while (!ReverseDepsToAdd.empty()) { 1412 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1413 .insert(ReverseDepsToAdd.back().second); 1414 ReverseDepsToAdd.pop_back(); 1415 } 1416 } 1417 1418 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1419 // value in the NonLocalPointerDeps info. 1420 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1421 ReverseNonLocalPtrDeps.find(RemInst); 1422 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1423 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1424 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1425 1426 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1427 E = Set.end(); I != E; ++I) { 1428 ValueIsLoadPair P = *I; 1429 assert(P.getPointer() != RemInst && 1430 "Already removed NonLocalPointerDeps info for RemInst"); 1431 1432 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps; 1433 1434 // The cache is not valid for any specific block anymore. 1435 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair(); 1436 1437 // Update any entries for RemInst to use the instruction after it. 1438 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1439 DI != DE; ++DI) { 1440 if (DI->getResult().getInst() != RemInst) continue; 1441 1442 // Convert to a dirty entry for the subsequent instruction. 1443 DI->setResult(NewDirtyVal); 1444 1445 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1446 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1447 } 1448 1449 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1450 // subsequent value may invalidate the sortedness. 1451 std::sort(NLPDI.begin(), NLPDI.end()); 1452 } 1453 1454 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1455 1456 while (!ReversePtrDepsToAdd.empty()) { 1457 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1458 .insert(ReversePtrDepsToAdd.back().second); 1459 ReversePtrDepsToAdd.pop_back(); 1460 } 1461 } 1462 1463 1464 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1465 AA->deleteValue(RemInst); 1466 DEBUG(verifyRemoved(RemInst)); 1467} 1468/// verifyRemoved - Verify that the specified instruction does not occur 1469/// in our internal data structures. 1470void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1471 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1472 E = LocalDeps.end(); I != E; ++I) { 1473 assert(I->first != D && "Inst occurs in data structures"); 1474 assert(I->second.getInst() != D && 1475 "Inst occurs in data structures"); 1476 } 1477 1478 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1479 E = NonLocalPointerDeps.end(); I != E; ++I) { 1480 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1481 const NonLocalDepInfo &Val = I->second.NonLocalDeps; 1482 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1483 II != E; ++II) 1484 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1485 } 1486 1487 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1488 E = NonLocalDeps.end(); I != E; ++I) { 1489 assert(I->first != D && "Inst occurs in data structures"); 1490 const PerInstNLInfo &INLD = I->second; 1491 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1492 EE = INLD.first.end(); II != EE; ++II) 1493 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1494 } 1495 1496 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1497 E = ReverseLocalDeps.end(); I != E; ++I) { 1498 assert(I->first != D && "Inst occurs in data structures"); 1499 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1500 EE = I->second.end(); II != EE; ++II) 1501 assert(*II != D && "Inst occurs in data structures"); 1502 } 1503 1504 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1505 E = ReverseNonLocalDeps.end(); 1506 I != E; ++I) { 1507 assert(I->first != D && "Inst occurs in data structures"); 1508 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1509 EE = I->second.end(); II != EE; ++II) 1510 assert(*II != D && "Inst occurs in data structures"); 1511 } 1512 1513 for (ReverseNonLocalPtrDepTy::const_iterator 1514 I = ReverseNonLocalPtrDeps.begin(), 1515 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1516 assert(I->first != D && "Inst occurs in rev NLPD map"); 1517 1518 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1519 E = I->second.end(); II != E; ++II) 1520 assert(*II != ValueIsLoadPair(D, false) && 1521 *II != ValueIsLoadPair(D, true) && 1522 "Inst occurs in ReverseNonLocalPtrDeps map"); 1523 } 1524 1525} 1526