MemoryDependenceAnalysis.cpp revision 7bce462c15356229e13d78d14560feaac30a0f5f
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 } else if (LI->getOrdering() == Monotonic) { 127 Loc = AA->getLocation(LI); 128 return AliasAnalysis::ModRef; 129 } 130 Loc = AliasAnalysis::Location(); 131 return AliasAnalysis::ModRef; 132 } 133 134 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 135 if (SI->isUnordered()) { 136 Loc = AA->getLocation(SI); 137 return AliasAnalysis::Mod; 138 } else if (SI->getOrdering() == Monotonic) { 139 Loc = AA->getLocation(SI); 140 return AliasAnalysis::ModRef; 141 } 142 Loc = AliasAnalysis::Location(); 143 return AliasAnalysis::ModRef; 144 } 145 146 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 147 Loc = AA->getLocation(V); 148 return AliasAnalysis::ModRef; 149 } 150 151 if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) { 152 // calls to free() deallocate the entire structure 153 Loc = AliasAnalysis::Location(CI->getArgOperand(0)); 154 return AliasAnalysis::Mod; 155 } 156 157 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) 158 switch (II->getIntrinsicID()) { 159 case Intrinsic::lifetime_start: 160 case Intrinsic::lifetime_end: 161 case Intrinsic::invariant_start: 162 Loc = AliasAnalysis::Location(II->getArgOperand(1), 163 cast<ConstantInt>(II->getArgOperand(0)) 164 ->getZExtValue(), 165 II->getMetadata(LLVMContext::MD_tbaa)); 166 // These intrinsics don't really modify the memory, but returning Mod 167 // will allow them to be handled conservatively. 168 return AliasAnalysis::Mod; 169 case Intrinsic::invariant_end: 170 Loc = AliasAnalysis::Location(II->getArgOperand(2), 171 cast<ConstantInt>(II->getArgOperand(1)) 172 ->getZExtValue(), 173 II->getMetadata(LLVMContext::MD_tbaa)); 174 // These intrinsics don't really modify the memory, but returning Mod 175 // will allow them to be handled conservatively. 176 return AliasAnalysis::Mod; 177 default: 178 break; 179 } 180 181 // Otherwise, just do the coarse-grained thing that always works. 182 if (Inst->mayWriteToMemory()) 183 return AliasAnalysis::ModRef; 184 if (Inst->mayReadFromMemory()) 185 return AliasAnalysis::Ref; 186 return AliasAnalysis::NoModRef; 187} 188 189/// getCallSiteDependencyFrom - Private helper for finding the local 190/// dependencies of a call site. 191MemDepResult MemoryDependenceAnalysis:: 192getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, 193 BasicBlock::iterator ScanIt, BasicBlock *BB) { 194 unsigned Limit = BlockScanLimit; 195 196 // Walk backwards through the block, looking for dependencies 197 while (ScanIt != BB->begin()) { 198 // Limit the amount of scanning we do so we don't end up with quadratic 199 // running time on extreme testcases. 200 --Limit; 201 if (!Limit) 202 return MemDepResult::getUnknown(); 203 204 Instruction *Inst = --ScanIt; 205 206 // If this inst is a memory op, get the pointer it accessed 207 AliasAnalysis::Location Loc; 208 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA); 209 if (Loc.Ptr) { 210 // A simple instruction. 211 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef) 212 return MemDepResult::getClobber(Inst); 213 continue; 214 } 215 216 if (CallSite InstCS = cast<Value>(Inst)) { 217 // Debug intrinsics don't cause dependences. 218 if (isa<DbgInfoIntrinsic>(Inst)) continue; 219 // If these two calls do not interfere, look past it. 220 switch (AA->getModRefInfo(CS, InstCS)) { 221 case AliasAnalysis::NoModRef: 222 // If the two calls are the same, return InstCS as a Def, so that 223 // CS can be found redundant and eliminated. 224 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) && 225 CS.getInstruction()->isIdenticalToWhenDefined(Inst)) 226 return MemDepResult::getDef(Inst); 227 228 // Otherwise if the two calls don't interact (e.g. InstCS is readnone) 229 // keep scanning. 230 continue; 231 default: 232 return MemDepResult::getClobber(Inst); 233 } 234 } 235 236 // If we could not obtain a pointer for the instruction and the instruction 237 // touches memory then assume that this is a dependency. 238 if (MR != AliasAnalysis::NoModRef) 239 return MemDepResult::getClobber(Inst); 240 } 241 242 // No dependence found. If this is the entry block of the function, it is 243 // unknown, otherwise it is non-local. 244 if (BB != &BB->getParent()->getEntryBlock()) 245 return MemDepResult::getNonLocal(); 246 return MemDepResult::getNonFuncLocal(); 247} 248 249/// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that 250/// would fully overlap MemLoc if done as a wider legal integer load. 251/// 252/// MemLocBase, MemLocOffset are lazily computed here the first time the 253/// base/offs of memloc is needed. 254static bool 255isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc, 256 const Value *&MemLocBase, 257 int64_t &MemLocOffs, 258 const LoadInst *LI, 259 const DataLayout *TD) { 260 // If we have no target data, we can't do this. 261 if (TD == 0) return false; 262 263 // If we haven't already computed the base/offset of MemLoc, do so now. 264 if (MemLocBase == 0) 265 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, TD); 266 267 unsigned Size = MemoryDependenceAnalysis:: 268 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size, 269 LI, *TD); 270 return Size != 0; 271} 272 273/// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that 274/// looks at a memory location for a load (specified by MemLocBase, Offs, 275/// and Size) and compares it against a load. If the specified load could 276/// be safely widened to a larger integer load that is 1) still efficient, 277/// 2) safe for the target, and 3) would provide the specified memory 278/// location value, then this function returns the size in bytes of the 279/// load width to use. If not, this returns zero. 280unsigned MemoryDependenceAnalysis:: 281getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs, 282 unsigned MemLocSize, const LoadInst *LI, 283 const DataLayout &TD) { 284 // We can only extend simple integer loads. 285 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0; 286 287 // Load widening is hostile to ThreadSanitizer: it may cause false positives 288 // or make the reports more cryptic (access sizes are wrong). 289 if (LI->getParent()->getParent()->getAttributes(). 290 hasAttribute(AttributeSet::FunctionIndex, Attribute::ThreadSafety)) 291 return 0; 292 293 // Get the base of this load. 294 int64_t LIOffs = 0; 295 const Value *LIBase = 296 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, &TD); 297 298 // If the two pointers are not based on the same pointer, we can't tell that 299 // they are related. 300 if (LIBase != MemLocBase) return 0; 301 302 // Okay, the two values are based on the same pointer, but returned as 303 // no-alias. This happens when we have things like two byte loads at "P+1" 304 // and "P+3". Check to see if increasing the size of the "LI" load up to its 305 // alignment (or the largest native integer type) will allow us to load all 306 // the bits required by MemLoc. 307 308 // If MemLoc is before LI, then no widening of LI will help us out. 309 if (MemLocOffs < LIOffs) return 0; 310 311 // Get the alignment of the load in bytes. We assume that it is safe to load 312 // any legal integer up to this size without a problem. For example, if we're 313 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can 314 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it 315 // to i16. 316 unsigned LoadAlign = LI->getAlignment(); 317 318 int64_t MemLocEnd = MemLocOffs+MemLocSize; 319 320 // If no amount of rounding up will let MemLoc fit into LI, then bail out. 321 if (LIOffs+LoadAlign < MemLocEnd) return 0; 322 323 // This is the size of the load to try. Start with the next larger power of 324 // two. 325 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U; 326 NewLoadByteSize = NextPowerOf2(NewLoadByteSize); 327 328 while (1) { 329 // If this load size is bigger than our known alignment or would not fit 330 // into a native integer register, then we fail. 331 if (NewLoadByteSize > LoadAlign || 332 !TD.fitsInLegalInteger(NewLoadByteSize*8)) 333 return 0; 334 335 if (LIOffs+NewLoadByteSize > MemLocEnd && 336 LI->getParent()->getParent()->getAttributes(). 337 hasAttribute(AttributeSet::FunctionIndex, Attribute::AddressSafety)) 338 // We will be reading past the location accessed by the original program. 339 // While this is safe in a regular build, Address Safety analysis tools 340 // may start reporting false warnings. So, don't do widening. 341 return 0; 342 343 // If a load of this width would include all of MemLoc, then we succeed. 344 if (LIOffs+NewLoadByteSize >= MemLocEnd) 345 return NewLoadByteSize; 346 347 NewLoadByteSize <<= 1; 348 } 349} 350 351/// getPointerDependencyFrom - Return the instruction on which a memory 352/// location depends. If isLoad is true, this routine ignores may-aliases with 353/// read-only operations. If isLoad is false, this routine ignores may-aliases 354/// with reads from read-only locations. 355MemDepResult MemoryDependenceAnalysis:: 356getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad, 357 BasicBlock::iterator ScanIt, BasicBlock *BB) { 358 359 const Value *MemLocBase = 0; 360 int64_t MemLocOffset = 0; 361 362 unsigned Limit = BlockScanLimit; 363 364 // Walk backwards through the basic block, looking for dependencies. 365 while (ScanIt != BB->begin()) { 366 // Limit the amount of scanning we do so we don't end up with quadratic 367 // running time on extreme testcases. 368 --Limit; 369 if (!Limit) 370 return MemDepResult::getUnknown(); 371 372 Instruction *Inst = --ScanIt; 373 374 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 375 // Debug intrinsics don't (and can't) cause dependences. 376 if (isa<DbgInfoIntrinsic>(II)) continue; 377 378 // If we reach a lifetime begin or end marker, then the query ends here 379 // because the value is undefined. 380 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 381 // FIXME: This only considers queries directly on the invariant-tagged 382 // pointer, not on query pointers that are indexed off of them. It'd 383 // be nice to handle that at some point (the right approach is to use 384 // GetPointerBaseWithConstantOffset). 385 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)), 386 MemLoc)) 387 return MemDepResult::getDef(II); 388 continue; 389 } 390 } 391 392 // Values depend on loads if the pointers are must aliased. This means that 393 // a load depends on another must aliased load from the same value. 394 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 395 // Atomic loads have complications involved. 396 // FIXME: This is overly conservative. 397 if (!LI->isUnordered()) 398 return MemDepResult::getClobber(LI); 399 400 AliasAnalysis::Location LoadLoc = AA->getLocation(LI); 401 402 // If we found a pointer, check if it could be the same as our pointer. 403 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc); 404 405 if (isLoad) { 406 if (R == AliasAnalysis::NoAlias) { 407 // If this is an over-aligned integer load (for example, 408 // "load i8* %P, align 4") see if it would obviously overlap with the 409 // queried location if widened to a larger load (e.g. if the queried 410 // location is 1 byte at P+1). If so, return it as a load/load 411 // clobber result, allowing the client to decide to widen the load if 412 // it wants to. 413 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) 414 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() && 415 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase, 416 MemLocOffset, LI, TD)) 417 return MemDepResult::getClobber(Inst); 418 419 continue; 420 } 421 422 // Must aliased loads are defs of each other. 423 if (R == AliasAnalysis::MustAlias) 424 return MemDepResult::getDef(Inst); 425 426#if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads 427 // in terms of clobbering loads, but since it does this by looking 428 // at the clobbering load directly, it doesn't know about any 429 // phi translation that may have happened along the way. 430 431 // If we have a partial alias, then return this as a clobber for the 432 // client to handle. 433 if (R == AliasAnalysis::PartialAlias) 434 return MemDepResult::getClobber(Inst); 435#endif 436 437 // Random may-alias loads don't depend on each other without a 438 // dependence. 439 continue; 440 } 441 442 // Stores don't depend on other no-aliased accesses. 443 if (R == AliasAnalysis::NoAlias) 444 continue; 445 446 // Stores don't alias loads from read-only memory. 447 if (AA->pointsToConstantMemory(LoadLoc)) 448 continue; 449 450 // Stores depend on may/must aliased loads. 451 return MemDepResult::getDef(Inst); 452 } 453 454 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 455 // Atomic stores have complications involved. 456 // FIXME: This is overly conservative. 457 if (!SI->isUnordered()) 458 return MemDepResult::getClobber(SI); 459 460 // If alias analysis can tell that this store is guaranteed to not modify 461 // the query pointer, ignore it. Use getModRefInfo to handle cases where 462 // the query pointer points to constant memory etc. 463 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef) 464 continue; 465 466 // Ok, this store might clobber the query pointer. Check to see if it is 467 // a must alias: in this case, we want to return this as a def. 468 AliasAnalysis::Location StoreLoc = AA->getLocation(SI); 469 470 // If we found a pointer, check if it could be the same as our pointer. 471 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc); 472 473 if (R == AliasAnalysis::NoAlias) 474 continue; 475 if (R == AliasAnalysis::MustAlias) 476 return MemDepResult::getDef(Inst); 477 return MemDepResult::getClobber(Inst); 478 } 479 480 // If this is an allocation, and if we know that the accessed pointer is to 481 // the allocation, return Def. This means that there is no dependence and 482 // the access can be optimized based on that. For example, a load could 483 // turn into undef. 484 // Note: Only determine this to be a malloc if Inst is the malloc call, not 485 // a subsequent bitcast of the malloc call result. There can be stores to 486 // the malloced memory between the malloc call and its bitcast uses, and we 487 // need to continue scanning until the malloc call. 488 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo(); 489 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) { 490 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD); 491 492 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr)) 493 return MemDepResult::getDef(Inst); 494 // Be conservative if the accessed pointer may alias the allocation. 495 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias) 496 return MemDepResult::getClobber(Inst); 497 // If the allocation is not aliased and does not read memory (like 498 // strdup), it is safe to ignore. 499 if (isa<AllocaInst>(Inst) || 500 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI)) 501 continue; 502 } 503 504 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 505 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc); 506 // If necessary, perform additional analysis. 507 if (MR == AliasAnalysis::ModRef) 508 MR = AA->callCapturesBefore(Inst, MemLoc, DT); 509 switch (MR) { 510 case AliasAnalysis::NoModRef: 511 // If the call has no effect on the queried pointer, just ignore it. 512 continue; 513 case AliasAnalysis::Mod: 514 return MemDepResult::getClobber(Inst); 515 case AliasAnalysis::Ref: 516 // If the call is known to never store to the pointer, and if this is a 517 // load query, we can safely ignore it (scan past it). 518 if (isLoad) 519 continue; 520 default: 521 // Otherwise, there is a potential dependence. Return a clobber. 522 return MemDepResult::getClobber(Inst); 523 } 524 } 525 526 // No dependence found. If this is the entry block of the function, it is 527 // unknown, otherwise it is non-local. 528 if (BB != &BB->getParent()->getEntryBlock()) 529 return MemDepResult::getNonLocal(); 530 return MemDepResult::getNonFuncLocal(); 531} 532 533/// getDependency - Return the instruction on which a memory operation 534/// depends. 535MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 536 Instruction *ScanPos = QueryInst; 537 538 // Check for a cached result 539 MemDepResult &LocalCache = LocalDeps[QueryInst]; 540 541 // If the cached entry is non-dirty, just return it. Note that this depends 542 // on MemDepResult's default constructing to 'dirty'. 543 if (!LocalCache.isDirty()) 544 return LocalCache; 545 546 // Otherwise, if we have a dirty entry, we know we can start the scan at that 547 // instruction, which may save us some work. 548 if (Instruction *Inst = LocalCache.getInst()) { 549 ScanPos = Inst; 550 551 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 552 } 553 554 BasicBlock *QueryParent = QueryInst->getParent(); 555 556 // Do the scan. 557 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 558 // No dependence found. If this is the entry block of the function, it is 559 // unknown, otherwise it is non-local. 560 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 561 LocalCache = MemDepResult::getNonLocal(); 562 else 563 LocalCache = MemDepResult::getNonFuncLocal(); 564 } else { 565 AliasAnalysis::Location MemLoc; 566 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA); 567 if (MemLoc.Ptr) { 568 // If we can do a pointer scan, make it happen. 569 bool isLoad = !(MR & AliasAnalysis::Mod); 570 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst)) 571 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start; 572 573 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos, 574 QueryParent); 575 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 576 CallSite QueryCS(QueryInst); 577 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 578 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 579 QueryParent); 580 } else 581 // Non-memory instruction. 582 LocalCache = MemDepResult::getUnknown(); 583 } 584 585 // Remember the result! 586 if (Instruction *I = LocalCache.getInst()) 587 ReverseLocalDeps[I].insert(QueryInst); 588 589 return LocalCache; 590} 591 592#ifndef NDEBUG 593/// AssertSorted - This method is used when -debug is specified to verify that 594/// cache arrays are properly kept sorted. 595static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 596 int Count = -1) { 597 if (Count == -1) Count = Cache.size(); 598 if (Count == 0) return; 599 600 for (unsigned i = 1; i != unsigned(Count); ++i) 601 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 602} 603#endif 604 605/// getNonLocalCallDependency - Perform a full dependency query for the 606/// specified call, returning the set of blocks that the value is 607/// potentially live across. The returned set of results will include a 608/// "NonLocal" result for all blocks where the value is live across. 609/// 610/// This method assumes the instruction returns a "NonLocal" dependency 611/// within its own block. 612/// 613/// This returns a reference to an internal data structure that may be 614/// invalidated on the next non-local query or when an instruction is 615/// removed. Clients must copy this data if they want it around longer than 616/// that. 617const MemoryDependenceAnalysis::NonLocalDepInfo & 618MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 619 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 620 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 621 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 622 NonLocalDepInfo &Cache = CacheP.first; 623 624 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 625 /// the cached case, this can happen due to instructions being deleted etc. In 626 /// the uncached case, this starts out as the set of predecessors we care 627 /// about. 628 SmallVector<BasicBlock*, 32> DirtyBlocks; 629 630 if (!Cache.empty()) { 631 // Okay, we have a cache entry. If we know it is not dirty, just return it 632 // with no computation. 633 if (!CacheP.second) { 634 ++NumCacheNonLocal; 635 return Cache; 636 } 637 638 // If we already have a partially computed set of results, scan them to 639 // determine what is dirty, seeding our initial DirtyBlocks worklist. 640 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 641 I != E; ++I) 642 if (I->getResult().isDirty()) 643 DirtyBlocks.push_back(I->getBB()); 644 645 // Sort the cache so that we can do fast binary search lookups below. 646 std::sort(Cache.begin(), Cache.end()); 647 648 ++NumCacheDirtyNonLocal; 649 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 650 // << Cache.size() << " cached: " << *QueryInst; 651 } else { 652 // Seed DirtyBlocks with each of the preds of QueryInst's block. 653 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 654 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 655 DirtyBlocks.push_back(*PI); 656 ++NumUncacheNonLocal; 657 } 658 659 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 660 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 661 662 SmallPtrSet<BasicBlock*, 64> Visited; 663 664 unsigned NumSortedEntries = Cache.size(); 665 DEBUG(AssertSorted(Cache)); 666 667 // Iterate while we still have blocks to update. 668 while (!DirtyBlocks.empty()) { 669 BasicBlock *DirtyBB = DirtyBlocks.back(); 670 DirtyBlocks.pop_back(); 671 672 // Already processed this block? 673 if (!Visited.insert(DirtyBB)) 674 continue; 675 676 // Do a binary search to see if we already have an entry for this block in 677 // the cache set. If so, find it. 678 DEBUG(AssertSorted(Cache, NumSortedEntries)); 679 NonLocalDepInfo::iterator Entry = 680 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 681 NonLocalDepEntry(DirtyBB)); 682 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB) 683 --Entry; 684 685 NonLocalDepEntry *ExistingResult = 0; 686 if (Entry != Cache.begin()+NumSortedEntries && 687 Entry->getBB() == DirtyBB) { 688 // If we already have an entry, and if it isn't already dirty, the block 689 // is done. 690 if (!Entry->getResult().isDirty()) 691 continue; 692 693 // Otherwise, remember this slot so we can update the value. 694 ExistingResult = &*Entry; 695 } 696 697 // If the dirty entry has a pointer, start scanning from it so we don't have 698 // to rescan the entire block. 699 BasicBlock::iterator ScanPos = DirtyBB->end(); 700 if (ExistingResult) { 701 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 702 ScanPos = Inst; 703 // We're removing QueryInst's use of Inst. 704 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 705 QueryCS.getInstruction()); 706 } 707 } 708 709 // Find out if this block has a local dependency for QueryInst. 710 MemDepResult Dep; 711 712 if (ScanPos != DirtyBB->begin()) { 713 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 714 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 715 // No dependence found. If this is the entry block of the function, it is 716 // a clobber, otherwise it is unknown. 717 Dep = MemDepResult::getNonLocal(); 718 } else { 719 Dep = MemDepResult::getNonFuncLocal(); 720 } 721 722 // If we had a dirty entry for the block, update it. Otherwise, just add 723 // a new entry. 724 if (ExistingResult) 725 ExistingResult->setResult(Dep); 726 else 727 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 728 729 // If the block has a dependency (i.e. it isn't completely transparent to 730 // the value), remember the association! 731 if (!Dep.isNonLocal()) { 732 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 733 // update this when we remove instructions. 734 if (Instruction *Inst = Dep.getInst()) 735 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 736 } else { 737 738 // If the block *is* completely transparent to the load, we need to check 739 // the predecessors of this block. Add them to our worklist. 740 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 741 DirtyBlocks.push_back(*PI); 742 } 743 } 744 745 return Cache; 746} 747 748/// getNonLocalPointerDependency - Perform a full dependency query for an 749/// access to the specified (non-volatile) memory location, returning the 750/// set of instructions that either define or clobber the value. 751/// 752/// This method assumes the pointer has a "NonLocal" dependency within its 753/// own block. 754/// 755void MemoryDependenceAnalysis:: 756getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad, 757 BasicBlock *FromBB, 758 SmallVectorImpl<NonLocalDepResult> &Result) { 759 assert(Loc.Ptr->getType()->isPointerTy() && 760 "Can't get pointer deps of a non-pointer!"); 761 Result.clear(); 762 763 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD); 764 765 // This is the set of blocks we've inspected, and the pointer we consider in 766 // each block. Because of critical edges, we currently bail out if querying 767 // a block with multiple different pointers. This can happen during PHI 768 // translation. 769 DenseMap<BasicBlock*, Value*> Visited; 770 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB, 771 Result, Visited, true)) 772 return; 773 Result.clear(); 774 Result.push_back(NonLocalDepResult(FromBB, 775 MemDepResult::getUnknown(), 776 const_cast<Value *>(Loc.Ptr))); 777} 778 779/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 780/// Pointer/PointeeSize using either cached information in Cache or by doing a 781/// lookup (which may use dirty cache info if available). If we do a lookup, 782/// add the result to the cache. 783MemDepResult MemoryDependenceAnalysis:: 784GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc, 785 bool isLoad, BasicBlock *BB, 786 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 787 788 // Do a binary search to see if we already have an entry for this block in 789 // the cache set. If so, find it. 790 NonLocalDepInfo::iterator Entry = 791 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 792 NonLocalDepEntry(BB)); 793 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 794 --Entry; 795 796 NonLocalDepEntry *ExistingResult = 0; 797 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 798 ExistingResult = &*Entry; 799 800 // If we have a cached entry, and it is non-dirty, use it as the value for 801 // this dependency. 802 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 803 ++NumCacheNonLocalPtr; 804 return ExistingResult->getResult(); 805 } 806 807 // Otherwise, we have to scan for the value. If we have a dirty cache 808 // entry, start scanning from its position, otherwise we scan from the end 809 // of the block. 810 BasicBlock::iterator ScanPos = BB->end(); 811 if (ExistingResult && ExistingResult->getResult().getInst()) { 812 assert(ExistingResult->getResult().getInst()->getParent() == BB && 813 "Instruction invalidated?"); 814 ++NumCacheDirtyNonLocalPtr; 815 ScanPos = ExistingResult->getResult().getInst(); 816 817 // Eliminating the dirty entry from 'Cache', so update the reverse info. 818 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 819 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 820 } else { 821 ++NumUncacheNonLocalPtr; 822 } 823 824 // Scan the block for the dependency. 825 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB); 826 827 // If we had a dirty entry for the block, update it. Otherwise, just add 828 // a new entry. 829 if (ExistingResult) 830 ExistingResult->setResult(Dep); 831 else 832 Cache->push_back(NonLocalDepEntry(BB, Dep)); 833 834 // If the block has a dependency (i.e. it isn't completely transparent to 835 // the value), remember the reverse association because we just added it 836 // to Cache! 837 if (!Dep.isDef() && !Dep.isClobber()) 838 return Dep; 839 840 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 841 // update MemDep when we remove instructions. 842 Instruction *Inst = Dep.getInst(); 843 assert(Inst && "Didn't depend on anything?"); 844 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 845 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 846 return Dep; 847} 848 849/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 850/// number of elements in the array that are already properly ordered. This is 851/// optimized for the case when only a few entries are added. 852static void 853SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 854 unsigned NumSortedEntries) { 855 switch (Cache.size() - NumSortedEntries) { 856 case 0: 857 // done, no new entries. 858 break; 859 case 2: { 860 // Two new entries, insert the last one into place. 861 NonLocalDepEntry Val = Cache.back(); 862 Cache.pop_back(); 863 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 864 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 865 Cache.insert(Entry, Val); 866 // FALL THROUGH. 867 } 868 case 1: 869 // One new entry, Just insert the new value at the appropriate position. 870 if (Cache.size() != 1) { 871 NonLocalDepEntry Val = Cache.back(); 872 Cache.pop_back(); 873 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 874 std::upper_bound(Cache.begin(), Cache.end(), Val); 875 Cache.insert(Entry, Val); 876 } 877 break; 878 default: 879 // Added many values, do a full scale sort. 880 std::sort(Cache.begin(), Cache.end()); 881 break; 882 } 883} 884 885/// getNonLocalPointerDepFromBB - Perform a dependency query based on 886/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 887/// results to the results vector and keep track of which blocks are visited in 888/// 'Visited'. 889/// 890/// This has special behavior for the first block queries (when SkipFirstBlock 891/// is true). In this special case, it ignores the contents of the specified 892/// block and starts returning dependence info for its predecessors. 893/// 894/// This function returns false on success, or true to indicate that it could 895/// not compute dependence information for some reason. This should be treated 896/// as a clobber dependence on the first instruction in the predecessor block. 897bool MemoryDependenceAnalysis:: 898getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, 899 const AliasAnalysis::Location &Loc, 900 bool isLoad, BasicBlock *StartBB, 901 SmallVectorImpl<NonLocalDepResult> &Result, 902 DenseMap<BasicBlock*, Value*> &Visited, 903 bool SkipFirstBlock) { 904 905 // Look up the cached info for Pointer. 906 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 907 908 // Set up a temporary NLPI value. If the map doesn't yet have an entry for 909 // CacheKey, this value will be inserted as the associated value. Otherwise, 910 // it'll be ignored, and we'll have to check to see if the cached size and 911 // tbaa tag are consistent with the current query. 912 NonLocalPointerInfo InitialNLPI; 913 InitialNLPI.Size = Loc.Size; 914 InitialNLPI.TBAATag = Loc.TBAATag; 915 916 // Get the NLPI for CacheKey, inserting one into the map if it doesn't 917 // already have one. 918 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair = 919 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI)); 920 NonLocalPointerInfo *CacheInfo = &Pair.first->second; 921 922 // If we already have a cache entry for this CacheKey, we may need to do some 923 // work to reconcile the cache entry and the current query. 924 if (!Pair.second) { 925 if (CacheInfo->Size < Loc.Size) { 926 // The query's Size is greater than the cached one. Throw out the 927 // cached data and proceed with the query at the greater size. 928 CacheInfo->Pair = BBSkipFirstBlockPair(); 929 CacheInfo->Size = Loc.Size; 930 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 931 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 932 if (Instruction *Inst = DI->getResult().getInst()) 933 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 934 CacheInfo->NonLocalDeps.clear(); 935 } else if (CacheInfo->Size > Loc.Size) { 936 // This query's Size is less than the cached one. Conservatively restart 937 // the query using the greater size. 938 return getNonLocalPointerDepFromBB(Pointer, 939 Loc.getWithNewSize(CacheInfo->Size), 940 isLoad, StartBB, Result, Visited, 941 SkipFirstBlock); 942 } 943 944 // If the query's TBAATag is inconsistent with the cached one, 945 // conservatively throw out the cached data and restart the query with 946 // no tag if needed. 947 if (CacheInfo->TBAATag != Loc.TBAATag) { 948 if (CacheInfo->TBAATag) { 949 CacheInfo->Pair = BBSkipFirstBlockPair(); 950 CacheInfo->TBAATag = 0; 951 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 952 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 953 if (Instruction *Inst = DI->getResult().getInst()) 954 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 955 CacheInfo->NonLocalDeps.clear(); 956 } 957 if (Loc.TBAATag) 958 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(), 959 isLoad, StartBB, Result, Visited, 960 SkipFirstBlock); 961 } 962 } 963 964 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps; 965 966 // If we have valid cached information for exactly the block we are 967 // investigating, just return it with no recomputation. 968 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 969 // We have a fully cached result for this query then we can just return the 970 // cached results and populate the visited set. However, we have to verify 971 // that we don't already have conflicting results for these blocks. Check 972 // to ensure that if a block in the results set is in the visited set that 973 // it was for the same pointer query. 974 if (!Visited.empty()) { 975 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 976 I != E; ++I) { 977 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 978 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 979 continue; 980 981 // We have a pointer mismatch in a block. Just return clobber, saying 982 // that something was clobbered in this result. We could also do a 983 // non-fully cached query, but there is little point in doing this. 984 return true; 985 } 986 } 987 988 Value *Addr = Pointer.getAddr(); 989 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 990 I != E; ++I) { 991 Visited.insert(std::make_pair(I->getBB(), Addr)); 992 if (!I->getResult().isNonLocal() && DT->isReachableFromEntry(I->getBB())) 993 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 994 } 995 ++NumCacheCompleteNonLocalPtr; 996 return false; 997 } 998 999 // Otherwise, either this is a new block, a block with an invalid cache 1000 // pointer or one that we're about to invalidate by putting more info into it 1001 // than its valid cache info. If empty, the result will be valid cache info, 1002 // otherwise it isn't. 1003 if (Cache->empty()) 1004 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 1005 else 1006 CacheInfo->Pair = BBSkipFirstBlockPair(); 1007 1008 SmallVector<BasicBlock*, 32> Worklist; 1009 Worklist.push_back(StartBB); 1010 1011 // PredList used inside loop. 1012 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList; 1013 1014 // Keep track of the entries that we know are sorted. Previously cached 1015 // entries will all be sorted. The entries we add we only sort on demand (we 1016 // don't insert every element into its sorted position). We know that we 1017 // won't get any reuse from currently inserted values, because we don't 1018 // revisit blocks after we insert info for them. 1019 unsigned NumSortedEntries = Cache->size(); 1020 DEBUG(AssertSorted(*Cache)); 1021 1022 while (!Worklist.empty()) { 1023 BasicBlock *BB = Worklist.pop_back_val(); 1024 1025 // Skip the first block if we have it. 1026 if (!SkipFirstBlock) { 1027 // Analyze the dependency of *Pointer in FromBB. See if we already have 1028 // been here. 1029 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 1030 1031 // Get the dependency info for Pointer in BB. If we have cached 1032 // information, we will use it, otherwise we compute it. 1033 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 1034 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache, 1035 NumSortedEntries); 1036 1037 // If we got a Def or Clobber, add this to the list of results. 1038 if (!Dep.isNonLocal() && DT->isReachableFromEntry(BB)) { 1039 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 1040 continue; 1041 } 1042 } 1043 1044 // If 'Pointer' is an instruction defined in this block, then we need to do 1045 // phi translation to change it into a value live in the predecessor block. 1046 // If not, we just add the predecessors to the worklist and scan them with 1047 // the same Pointer. 1048 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 1049 SkipFirstBlock = false; 1050 SmallVector<BasicBlock*, 16> NewBlocks; 1051 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1052 // Verify that we haven't looked at this block yet. 1053 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1054 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 1055 if (InsertRes.second) { 1056 // First time we've looked at *PI. 1057 NewBlocks.push_back(*PI); 1058 continue; 1059 } 1060 1061 // If we have seen this block before, but it was with a different 1062 // pointer then we have a phi translation failure and we have to treat 1063 // this as a clobber. 1064 if (InsertRes.first->second != Pointer.getAddr()) { 1065 // Make sure to clean up the Visited map before continuing on to 1066 // PredTranslationFailure. 1067 for (unsigned i = 0; i < NewBlocks.size(); i++) 1068 Visited.erase(NewBlocks[i]); 1069 goto PredTranslationFailure; 1070 } 1071 } 1072 Worklist.append(NewBlocks.begin(), NewBlocks.end()); 1073 continue; 1074 } 1075 1076 // We do need to do phi translation, if we know ahead of time we can't phi 1077 // translate this value, don't even try. 1078 if (!Pointer.IsPotentiallyPHITranslatable()) 1079 goto PredTranslationFailure; 1080 1081 // We may have added values to the cache list before this PHI translation. 1082 // If so, we haven't done anything to ensure that the cache remains sorted. 1083 // Sort it now (if needed) so that recursive invocations of 1084 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 1085 // value will only see properly sorted cache arrays. 1086 if (Cache && NumSortedEntries != Cache->size()) { 1087 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1088 NumSortedEntries = Cache->size(); 1089 } 1090 Cache = 0; 1091 1092 PredList.clear(); 1093 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1094 BasicBlock *Pred = *PI; 1095 PredList.push_back(std::make_pair(Pred, Pointer)); 1096 1097 // Get the PHI translated pointer in this predecessor. This can fail if 1098 // not translatable, in which case the getAddr() returns null. 1099 PHITransAddr &PredPointer = PredList.back().second; 1100 PredPointer.PHITranslateValue(BB, Pred, 0); 1101 1102 Value *PredPtrVal = PredPointer.getAddr(); 1103 1104 // Check to see if we have already visited this pred block with another 1105 // pointer. If so, we can't do this lookup. This failure can occur 1106 // with PHI translation when a critical edge exists and the PHI node in 1107 // the successor translates to a pointer value different than the 1108 // pointer the block was first analyzed with. 1109 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1110 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 1111 1112 if (!InsertRes.second) { 1113 // We found the pred; take it off the list of preds to visit. 1114 PredList.pop_back(); 1115 1116 // If the predecessor was visited with PredPtr, then we already did 1117 // the analysis and can ignore it. 1118 if (InsertRes.first->second == PredPtrVal) 1119 continue; 1120 1121 // Otherwise, the block was previously analyzed with a different 1122 // pointer. We can't represent the result of this case, so we just 1123 // treat this as a phi translation failure. 1124 1125 // Make sure to clean up the Visited map before continuing on to 1126 // PredTranslationFailure. 1127 for (unsigned i = 0; i < PredList.size(); i++) 1128 Visited.erase(PredList[i].first); 1129 1130 goto PredTranslationFailure; 1131 } 1132 } 1133 1134 // Actually process results here; this need to be a separate loop to avoid 1135 // calling getNonLocalPointerDepFromBB for blocks we don't want to return 1136 // any results for. (getNonLocalPointerDepFromBB will modify our 1137 // datastructures in ways the code after the PredTranslationFailure label 1138 // doesn't expect.) 1139 for (unsigned i = 0; i < PredList.size(); i++) { 1140 BasicBlock *Pred = PredList[i].first; 1141 PHITransAddr &PredPointer = PredList[i].second; 1142 Value *PredPtrVal = PredPointer.getAddr(); 1143 1144 bool CanTranslate = true; 1145 // If PHI translation was unable to find an available pointer in this 1146 // predecessor, then we have to assume that the pointer is clobbered in 1147 // that predecessor. We can still do PRE of the load, which would insert 1148 // a computation of the pointer in this predecessor. 1149 if (PredPtrVal == 0) 1150 CanTranslate = false; 1151 1152 // FIXME: it is entirely possible that PHI translating will end up with 1153 // the same value. Consider PHI translating something like: 1154 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 1155 // to recurse here, pedantically speaking. 1156 1157 // If getNonLocalPointerDepFromBB fails here, that means the cached 1158 // result conflicted with the Visited list; we have to conservatively 1159 // assume it is unknown, but this also does not block PRE of the load. 1160 if (!CanTranslate || 1161 getNonLocalPointerDepFromBB(PredPointer, 1162 Loc.getWithNewPtr(PredPtrVal), 1163 isLoad, Pred, 1164 Result, Visited)) { 1165 // Add the entry to the Result list. 1166 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal); 1167 Result.push_back(Entry); 1168 1169 // Since we had a phi translation failure, the cache for CacheKey won't 1170 // include all of the entries that we need to immediately satisfy future 1171 // queries. Mark this in NonLocalPointerDeps by setting the 1172 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 1173 // cached value to do more work but not miss the phi trans failure. 1174 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey]; 1175 NLPI.Pair = BBSkipFirstBlockPair(); 1176 continue; 1177 } 1178 } 1179 1180 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 1181 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1182 Cache = &CacheInfo->NonLocalDeps; 1183 NumSortedEntries = Cache->size(); 1184 1185 // Since we did phi translation, the "Cache" set won't contain all of the 1186 // results for the query. This is ok (we can still use it to accelerate 1187 // specific block queries) but we can't do the fastpath "return all 1188 // results from the set" Clear out the indicator for this. 1189 CacheInfo->Pair = BBSkipFirstBlockPair(); 1190 SkipFirstBlock = false; 1191 continue; 1192 1193 PredTranslationFailure: 1194 // The following code is "failure"; we can't produce a sane translation 1195 // for the given block. It assumes that we haven't modified any of 1196 // our datastructures while processing the current block. 1197 1198 if (Cache == 0) { 1199 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1200 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1201 Cache = &CacheInfo->NonLocalDeps; 1202 NumSortedEntries = Cache->size(); 1203 } 1204 1205 // Since we failed phi translation, the "Cache" set won't contain all of the 1206 // results for the query. This is ok (we can still use it to accelerate 1207 // specific block queries) but we can't do the fastpath "return all 1208 // results from the set". Clear out the indicator for this. 1209 CacheInfo->Pair = BBSkipFirstBlockPair(); 1210 1211 // If *nothing* works, mark the pointer as unknown. 1212 // 1213 // If this is the magic first block, return this as a clobber of the whole 1214 // incoming value. Since we can't phi translate to one of the predecessors, 1215 // we have to bail out. 1216 if (SkipFirstBlock) 1217 return true; 1218 1219 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 1220 assert(I != Cache->rend() && "Didn't find current block??"); 1221 if (I->getBB() != BB) 1222 continue; 1223 1224 assert(I->getResult().isNonLocal() && 1225 "Should only be here with transparent block"); 1226 I->setResult(MemDepResult::getUnknown()); 1227 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 1228 Pointer.getAddr())); 1229 break; 1230 } 1231 } 1232 1233 // Okay, we're done now. If we added new values to the cache, re-sort it. 1234 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1235 DEBUG(AssertSorted(*Cache)); 1236 return false; 1237} 1238 1239/// RemoveCachedNonLocalPointerDependencies - If P exists in 1240/// CachedNonLocalPointerInfo, remove it. 1241void MemoryDependenceAnalysis:: 1242RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 1243 CachedNonLocalPointerInfo::iterator It = 1244 NonLocalPointerDeps.find(P); 1245 if (It == NonLocalPointerDeps.end()) return; 1246 1247 // Remove all of the entries in the BB->val map. This involves removing 1248 // instructions from the reverse map. 1249 NonLocalDepInfo &PInfo = It->second.NonLocalDeps; 1250 1251 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 1252 Instruction *Target = PInfo[i].getResult().getInst(); 1253 if (Target == 0) continue; // Ignore non-local dep results. 1254 assert(Target->getParent() == PInfo[i].getBB()); 1255 1256 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1257 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1258 } 1259 1260 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1261 NonLocalPointerDeps.erase(It); 1262} 1263 1264 1265/// invalidateCachedPointerInfo - This method is used to invalidate cached 1266/// information about the specified pointer, because it may be too 1267/// conservative in memdep. This is an optional call that can be used when 1268/// the client detects an equivalence between the pointer and some other 1269/// value and replaces the other value with ptr. This can make Ptr available 1270/// in more places that cached info does not necessarily keep. 1271void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1272 // If Ptr isn't really a pointer, just ignore it. 1273 if (!Ptr->getType()->isPointerTy()) return; 1274 // Flush store info for the pointer. 1275 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1276 // Flush load info for the pointer. 1277 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1278} 1279 1280/// invalidateCachedPredecessors - Clear the PredIteratorCache info. 1281/// This needs to be done when the CFG changes, e.g., due to splitting 1282/// critical edges. 1283void MemoryDependenceAnalysis::invalidateCachedPredecessors() { 1284 PredCache->clear(); 1285} 1286 1287/// removeInstruction - Remove an instruction from the dependence analysis, 1288/// updating the dependence of instructions that previously depended on it. 1289/// This method attempts to keep the cache coherent using the reverse map. 1290void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1291 // Walk through the Non-local dependencies, removing this one as the value 1292 // for any cached queries. 1293 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1294 if (NLDI != NonLocalDeps.end()) { 1295 NonLocalDepInfo &BlockMap = NLDI->second.first; 1296 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1297 DI != DE; ++DI) 1298 if (Instruction *Inst = DI->getResult().getInst()) 1299 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1300 NonLocalDeps.erase(NLDI); 1301 } 1302 1303 // If we have a cached local dependence query for this instruction, remove it. 1304 // 1305 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1306 if (LocalDepEntry != LocalDeps.end()) { 1307 // Remove us from DepInst's reverse set now that the local dep info is gone. 1308 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1309 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1310 1311 // Remove this local dependency info. 1312 LocalDeps.erase(LocalDepEntry); 1313 } 1314 1315 // If we have any cached pointer dependencies on this instruction, remove 1316 // them. If the instruction has non-pointer type, then it can't be a pointer 1317 // base. 1318 1319 // Remove it from both the load info and the store info. The instruction 1320 // can't be in either of these maps if it is non-pointer. 1321 if (RemInst->getType()->isPointerTy()) { 1322 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1323 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1324 } 1325 1326 // Loop over all of the things that depend on the instruction we're removing. 1327 // 1328 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1329 1330 // If we find RemInst as a clobber or Def in any of the maps for other values, 1331 // we need to replace its entry with a dirty version of the instruction after 1332 // it. If RemInst is a terminator, we use a null dirty value. 1333 // 1334 // Using a dirty version of the instruction after RemInst saves having to scan 1335 // the entire block to get to this point. 1336 MemDepResult NewDirtyVal; 1337 if (!RemInst->isTerminator()) 1338 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1339 1340 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1341 if (ReverseDepIt != ReverseLocalDeps.end()) { 1342 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1343 // RemInst can't be the terminator if it has local stuff depending on it. 1344 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1345 "Nothing can locally depend on a terminator"); 1346 1347 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1348 E = ReverseDeps.end(); I != E; ++I) { 1349 Instruction *InstDependingOnRemInst = *I; 1350 assert(InstDependingOnRemInst != RemInst && 1351 "Already removed our local dep info"); 1352 1353 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1354 1355 // Make sure to remember that new things depend on NewDepInst. 1356 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1357 "a local dep on this if it is a terminator!"); 1358 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1359 InstDependingOnRemInst)); 1360 } 1361 1362 ReverseLocalDeps.erase(ReverseDepIt); 1363 1364 // Add new reverse deps after scanning the set, to avoid invalidating the 1365 // 'ReverseDeps' reference. 1366 while (!ReverseDepsToAdd.empty()) { 1367 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1368 .insert(ReverseDepsToAdd.back().second); 1369 ReverseDepsToAdd.pop_back(); 1370 } 1371 } 1372 1373 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1374 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1375 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1376 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1377 I != E; ++I) { 1378 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1379 1380 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1381 // The information is now dirty! 1382 INLD.second = true; 1383 1384 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1385 DE = INLD.first.end(); DI != DE; ++DI) { 1386 if (DI->getResult().getInst() != RemInst) continue; 1387 1388 // Convert to a dirty entry for the subsequent instruction. 1389 DI->setResult(NewDirtyVal); 1390 1391 if (Instruction *NextI = NewDirtyVal.getInst()) 1392 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1393 } 1394 } 1395 1396 ReverseNonLocalDeps.erase(ReverseDepIt); 1397 1398 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1399 while (!ReverseDepsToAdd.empty()) { 1400 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1401 .insert(ReverseDepsToAdd.back().second); 1402 ReverseDepsToAdd.pop_back(); 1403 } 1404 } 1405 1406 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1407 // value in the NonLocalPointerDeps info. 1408 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1409 ReverseNonLocalPtrDeps.find(RemInst); 1410 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1411 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1412 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1413 1414 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1415 E = Set.end(); I != E; ++I) { 1416 ValueIsLoadPair P = *I; 1417 assert(P.getPointer() != RemInst && 1418 "Already removed NonLocalPointerDeps info for RemInst"); 1419 1420 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps; 1421 1422 // The cache is not valid for any specific block anymore. 1423 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair(); 1424 1425 // Update any entries for RemInst to use the instruction after it. 1426 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1427 DI != DE; ++DI) { 1428 if (DI->getResult().getInst() != RemInst) continue; 1429 1430 // Convert to a dirty entry for the subsequent instruction. 1431 DI->setResult(NewDirtyVal); 1432 1433 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1434 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1435 } 1436 1437 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1438 // subsequent value may invalidate the sortedness. 1439 std::sort(NLPDI.begin(), NLPDI.end()); 1440 } 1441 1442 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1443 1444 while (!ReversePtrDepsToAdd.empty()) { 1445 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1446 .insert(ReversePtrDepsToAdd.back().second); 1447 ReversePtrDepsToAdd.pop_back(); 1448 } 1449 } 1450 1451 1452 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1453 AA->deleteValue(RemInst); 1454 DEBUG(verifyRemoved(RemInst)); 1455} 1456/// verifyRemoved - Verify that the specified instruction does not occur 1457/// in our internal data structures. 1458void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1459 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1460 E = LocalDeps.end(); I != E; ++I) { 1461 assert(I->first != D && "Inst occurs in data structures"); 1462 assert(I->second.getInst() != D && 1463 "Inst occurs in data structures"); 1464 } 1465 1466 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1467 E = NonLocalPointerDeps.end(); I != E; ++I) { 1468 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1469 const NonLocalDepInfo &Val = I->second.NonLocalDeps; 1470 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1471 II != E; ++II) 1472 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1473 } 1474 1475 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1476 E = NonLocalDeps.end(); I != E; ++I) { 1477 assert(I->first != D && "Inst occurs in data structures"); 1478 const PerInstNLInfo &INLD = I->second; 1479 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1480 EE = INLD.first.end(); II != EE; ++II) 1481 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1482 } 1483 1484 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1485 E = ReverseLocalDeps.end(); I != E; ++I) { 1486 assert(I->first != D && "Inst occurs in data structures"); 1487 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1488 EE = I->second.end(); II != EE; ++II) 1489 assert(*II != D && "Inst occurs in data structures"); 1490 } 1491 1492 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1493 E = ReverseNonLocalDeps.end(); 1494 I != E; ++I) { 1495 assert(I->first != D && "Inst occurs in data structures"); 1496 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1497 EE = I->second.end(); II != EE; ++II) 1498 assert(*II != D && "Inst occurs in data structures"); 1499 } 1500 1501 for (ReverseNonLocalPtrDepTy::const_iterator 1502 I = ReverseNonLocalPtrDeps.begin(), 1503 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1504 assert(I->first != D && "Inst occurs in rev NLPD map"); 1505 1506 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1507 E = I->second.end(); II != E; ++II) 1508 assert(*II != ValueIsLoadPair(D, false) && 1509 *II != ValueIsLoadPair(D, true) && 1510 "Inst occurs in ReverseNonLocalPtrDeps map"); 1511 } 1512 1513} 1514