MemoryDependenceAnalysis.cpp revision d04a8d4b33ff316ca4cf961e06c9e312eff8e64f
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/DataLayout.h" 28#include "llvm/Function.h" 29#include "llvm/Instructions.h" 30#include "llvm/IntrinsicInst.h" 31#include "llvm/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 // Get the base of this load. 288 int64_t LIOffs = 0; 289 const Value *LIBase = 290 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, TD); 291 292 // If the two pointers are not based on the same pointer, we can't tell that 293 // they are related. 294 if (LIBase != MemLocBase) return 0; 295 296 // Okay, the two values are based on the same pointer, but returned as 297 // no-alias. This happens when we have things like two byte loads at "P+1" 298 // and "P+3". Check to see if increasing the size of the "LI" load up to its 299 // alignment (or the largest native integer type) will allow us to load all 300 // the bits required by MemLoc. 301 302 // If MemLoc is before LI, then no widening of LI will help us out. 303 if (MemLocOffs < LIOffs) return 0; 304 305 // Get the alignment of the load in bytes. We assume that it is safe to load 306 // any legal integer up to this size without a problem. For example, if we're 307 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can 308 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it 309 // to i16. 310 unsigned LoadAlign = LI->getAlignment(); 311 312 int64_t MemLocEnd = MemLocOffs+MemLocSize; 313 314 // If no amount of rounding up will let MemLoc fit into LI, then bail out. 315 if (LIOffs+LoadAlign < MemLocEnd) return 0; 316 317 // This is the size of the load to try. Start with the next larger power of 318 // two. 319 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U; 320 NewLoadByteSize = NextPowerOf2(NewLoadByteSize); 321 322 while (1) { 323 // If this load size is bigger than our known alignment or would not fit 324 // into a native integer register, then we fail. 325 if (NewLoadByteSize > LoadAlign || 326 !TD.fitsInLegalInteger(NewLoadByteSize*8)) 327 return 0; 328 329 if (LIOffs+NewLoadByteSize > MemLocEnd && 330 LI->getParent()->getParent()->getFnAttributes(). 331 hasAttribute(Attributes::AddressSafety)) 332 // We will be reading past the location accessed by the original program. 333 // While this is safe in a regular build, Address Safety analysis tools 334 // may start reporting false warnings. So, don't do widening. 335 return 0; 336 337 // If a load of this width would include all of MemLoc, then we succeed. 338 if (LIOffs+NewLoadByteSize >= MemLocEnd) 339 return NewLoadByteSize; 340 341 NewLoadByteSize <<= 1; 342 } 343} 344 345/// getPointerDependencyFrom - Return the instruction on which a memory 346/// location depends. If isLoad is true, this routine ignores may-aliases with 347/// read-only operations. If isLoad is false, this routine ignores may-aliases 348/// with reads from read-only locations. 349MemDepResult MemoryDependenceAnalysis:: 350getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad, 351 BasicBlock::iterator ScanIt, BasicBlock *BB) { 352 353 const Value *MemLocBase = 0; 354 int64_t MemLocOffset = 0; 355 356 unsigned Limit = BlockScanLimit; 357 358 // Walk backwards through the basic block, looking for dependencies. 359 while (ScanIt != BB->begin()) { 360 // Limit the amount of scanning we do so we don't end up with quadratic 361 // running time on extreme testcases. 362 --Limit; 363 if (!Limit) 364 return MemDepResult::getUnknown(); 365 366 Instruction *Inst = --ScanIt; 367 368 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 369 // Debug intrinsics don't (and can't) cause dependences. 370 if (isa<DbgInfoIntrinsic>(II)) continue; 371 372 // If we reach a lifetime begin or end marker, then the query ends here 373 // because the value is undefined. 374 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 375 // FIXME: This only considers queries directly on the invariant-tagged 376 // pointer, not on query pointers that are indexed off of them. It'd 377 // be nice to handle that at some point (the right approach is to use 378 // GetPointerBaseWithConstantOffset). 379 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)), 380 MemLoc)) 381 return MemDepResult::getDef(II); 382 continue; 383 } 384 } 385 386 // Values depend on loads if the pointers are must aliased. This means that 387 // a load depends on another must aliased load from the same value. 388 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 389 // Atomic loads have complications involved. 390 // FIXME: This is overly conservative. 391 if (!LI->isUnordered()) 392 return MemDepResult::getClobber(LI); 393 394 AliasAnalysis::Location LoadLoc = AA->getLocation(LI); 395 396 // If we found a pointer, check if it could be the same as our pointer. 397 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc); 398 399 if (isLoad) { 400 if (R == AliasAnalysis::NoAlias) { 401 // If this is an over-aligned integer load (for example, 402 // "load i8* %P, align 4") see if it would obviously overlap with the 403 // queried location if widened to a larger load (e.g. if the queried 404 // location is 1 byte at P+1). If so, return it as a load/load 405 // clobber result, allowing the client to decide to widen the load if 406 // it wants to. 407 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) 408 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() && 409 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase, 410 MemLocOffset, LI, TD)) 411 return MemDepResult::getClobber(Inst); 412 413 continue; 414 } 415 416 // Must aliased loads are defs of each other. 417 if (R == AliasAnalysis::MustAlias) 418 return MemDepResult::getDef(Inst); 419 420#if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads 421 // in terms of clobbering loads, but since it does this by looking 422 // at the clobbering load directly, it doesn't know about any 423 // phi translation that may have happened along the way. 424 425 // If we have a partial alias, then return this as a clobber for the 426 // client to handle. 427 if (R == AliasAnalysis::PartialAlias) 428 return MemDepResult::getClobber(Inst); 429#endif 430 431 // Random may-alias loads don't depend on each other without a 432 // dependence. 433 continue; 434 } 435 436 // Stores don't depend on other no-aliased accesses. 437 if (R == AliasAnalysis::NoAlias) 438 continue; 439 440 // Stores don't alias loads from read-only memory. 441 if (AA->pointsToConstantMemory(LoadLoc)) 442 continue; 443 444 // Stores depend on may/must aliased loads. 445 return MemDepResult::getDef(Inst); 446 } 447 448 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 449 // Atomic stores have complications involved. 450 // FIXME: This is overly conservative. 451 if (!SI->isUnordered()) 452 return MemDepResult::getClobber(SI); 453 454 // If alias analysis can tell that this store is guaranteed to not modify 455 // the query pointer, ignore it. Use getModRefInfo to handle cases where 456 // the query pointer points to constant memory etc. 457 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef) 458 continue; 459 460 // Ok, this store might clobber the query pointer. Check to see if it is 461 // a must alias: in this case, we want to return this as a def. 462 AliasAnalysis::Location StoreLoc = AA->getLocation(SI); 463 464 // If we found a pointer, check if it could be the same as our pointer. 465 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc); 466 467 if (R == AliasAnalysis::NoAlias) 468 continue; 469 if (R == AliasAnalysis::MustAlias) 470 return MemDepResult::getDef(Inst); 471 return MemDepResult::getClobber(Inst); 472 } 473 474 // If this is an allocation, and if we know that the accessed pointer is to 475 // the allocation, return Def. This means that there is no dependence and 476 // the access can be optimized based on that. For example, a load could 477 // turn into undef. 478 // Note: Only determine this to be a malloc if Inst is the malloc call, not 479 // a subsequent bitcast of the malloc call result. There can be stores to 480 // the malloced memory between the malloc call and its bitcast uses, and we 481 // need to continue scanning until the malloc call. 482 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo(); 483 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) { 484 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD); 485 486 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr)) 487 return MemDepResult::getDef(Inst); 488 // Be conservative if the accessed pointer may alias the allocation. 489 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias) 490 return MemDepResult::getClobber(Inst); 491 // If the allocation is not aliased and does not read memory (like 492 // strdup), it is safe to ignore. 493 if (isa<AllocaInst>(Inst) || 494 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI)) 495 continue; 496 } 497 498 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 499 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc); 500 // If necessary, perform additional analysis. 501 if (MR == AliasAnalysis::ModRef) 502 MR = AA->callCapturesBefore(Inst, MemLoc, DT); 503 switch (MR) { 504 case AliasAnalysis::NoModRef: 505 // If the call has no effect on the queried pointer, just ignore it. 506 continue; 507 case AliasAnalysis::Mod: 508 return MemDepResult::getClobber(Inst); 509 case AliasAnalysis::Ref: 510 // If the call is known to never store to the pointer, and if this is a 511 // load query, we can safely ignore it (scan past it). 512 if (isLoad) 513 continue; 514 default: 515 // Otherwise, there is a potential dependence. Return a clobber. 516 return MemDepResult::getClobber(Inst); 517 } 518 } 519 520 // No dependence found. If this is the entry block of the function, it is 521 // unknown, otherwise it is non-local. 522 if (BB != &BB->getParent()->getEntryBlock()) 523 return MemDepResult::getNonLocal(); 524 return MemDepResult::getNonFuncLocal(); 525} 526 527/// getDependency - Return the instruction on which a memory operation 528/// depends. 529MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { 530 Instruction *ScanPos = QueryInst; 531 532 // Check for a cached result 533 MemDepResult &LocalCache = LocalDeps[QueryInst]; 534 535 // If the cached entry is non-dirty, just return it. Note that this depends 536 // on MemDepResult's default constructing to 'dirty'. 537 if (!LocalCache.isDirty()) 538 return LocalCache; 539 540 // Otherwise, if we have a dirty entry, we know we can start the scan at that 541 // instruction, which may save us some work. 542 if (Instruction *Inst = LocalCache.getInst()) { 543 ScanPos = Inst; 544 545 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 546 } 547 548 BasicBlock *QueryParent = QueryInst->getParent(); 549 550 // Do the scan. 551 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 552 // No dependence found. If this is the entry block of the function, it is 553 // unknown, otherwise it is non-local. 554 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 555 LocalCache = MemDepResult::getNonLocal(); 556 else 557 LocalCache = MemDepResult::getNonFuncLocal(); 558 } else { 559 AliasAnalysis::Location MemLoc; 560 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA); 561 if (MemLoc.Ptr) { 562 // If we can do a pointer scan, make it happen. 563 bool isLoad = !(MR & AliasAnalysis::Mod); 564 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst)) 565 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start; 566 567 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos, 568 QueryParent); 569 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { 570 CallSite QueryCS(QueryInst); 571 bool isReadOnly = AA->onlyReadsMemory(QueryCS); 572 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, 573 QueryParent); 574 } else 575 // Non-memory instruction. 576 LocalCache = MemDepResult::getUnknown(); 577 } 578 579 // Remember the result! 580 if (Instruction *I = LocalCache.getInst()) 581 ReverseLocalDeps[I].insert(QueryInst); 582 583 return LocalCache; 584} 585 586#ifndef NDEBUG 587/// AssertSorted - This method is used when -debug is specified to verify that 588/// cache arrays are properly kept sorted. 589static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 590 int Count = -1) { 591 if (Count == -1) Count = Cache.size(); 592 if (Count == 0) return; 593 594 for (unsigned i = 1; i != unsigned(Count); ++i) 595 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!"); 596} 597#endif 598 599/// getNonLocalCallDependency - Perform a full dependency query for the 600/// specified call, returning the set of blocks that the value is 601/// potentially live across. The returned set of results will include a 602/// "NonLocal" result for all blocks where the value is live across. 603/// 604/// This method assumes the instruction returns a "NonLocal" dependency 605/// within its own block. 606/// 607/// This returns a reference to an internal data structure that may be 608/// invalidated on the next non-local query or when an instruction is 609/// removed. Clients must copy this data if they want it around longer than 610/// that. 611const MemoryDependenceAnalysis::NonLocalDepInfo & 612MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 613 assert(getDependency(QueryCS.getInstruction()).isNonLocal() && 614 "getNonLocalCallDependency should only be used on calls with non-local deps!"); 615 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; 616 NonLocalDepInfo &Cache = CacheP.first; 617 618 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In 619 /// the cached case, this can happen due to instructions being deleted etc. In 620 /// the uncached case, this starts out as the set of predecessors we care 621 /// about. 622 SmallVector<BasicBlock*, 32> DirtyBlocks; 623 624 if (!Cache.empty()) { 625 // Okay, we have a cache entry. If we know it is not dirty, just return it 626 // with no computation. 627 if (!CacheP.second) { 628 ++NumCacheNonLocal; 629 return Cache; 630 } 631 632 // If we already have a partially computed set of results, scan them to 633 // determine what is dirty, seeding our initial DirtyBlocks worklist. 634 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); 635 I != E; ++I) 636 if (I->getResult().isDirty()) 637 DirtyBlocks.push_back(I->getBB()); 638 639 // Sort the cache so that we can do fast binary search lookups below. 640 std::sort(Cache.begin(), Cache.end()); 641 642 ++NumCacheDirtyNonLocal; 643 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " 644 // << Cache.size() << " cached: " << *QueryInst; 645 } else { 646 // Seed DirtyBlocks with each of the preds of QueryInst's block. 647 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); 648 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) 649 DirtyBlocks.push_back(*PI); 650 ++NumUncacheNonLocal; 651 } 652 653 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 654 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); 655 656 SmallPtrSet<BasicBlock*, 64> Visited; 657 658 unsigned NumSortedEntries = Cache.size(); 659 DEBUG(AssertSorted(Cache)); 660 661 // Iterate while we still have blocks to update. 662 while (!DirtyBlocks.empty()) { 663 BasicBlock *DirtyBB = DirtyBlocks.back(); 664 DirtyBlocks.pop_back(); 665 666 // Already processed this block? 667 if (!Visited.insert(DirtyBB)) 668 continue; 669 670 // Do a binary search to see if we already have an entry for this block in 671 // the cache set. If so, find it. 672 DEBUG(AssertSorted(Cache, NumSortedEntries)); 673 NonLocalDepInfo::iterator Entry = 674 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, 675 NonLocalDepEntry(DirtyBB)); 676 if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB) 677 --Entry; 678 679 NonLocalDepEntry *ExistingResult = 0; 680 if (Entry != Cache.begin()+NumSortedEntries && 681 Entry->getBB() == DirtyBB) { 682 // If we already have an entry, and if it isn't already dirty, the block 683 // is done. 684 if (!Entry->getResult().isDirty()) 685 continue; 686 687 // Otherwise, remember this slot so we can update the value. 688 ExistingResult = &*Entry; 689 } 690 691 // If the dirty entry has a pointer, start scanning from it so we don't have 692 // to rescan the entire block. 693 BasicBlock::iterator ScanPos = DirtyBB->end(); 694 if (ExistingResult) { 695 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 696 ScanPos = Inst; 697 // We're removing QueryInst's use of Inst. 698 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 699 QueryCS.getInstruction()); 700 } 701 } 702 703 // Find out if this block has a local dependency for QueryInst. 704 MemDepResult Dep; 705 706 if (ScanPos != DirtyBB->begin()) { 707 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); 708 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 709 // No dependence found. If this is the entry block of the function, it is 710 // a clobber, otherwise it is unknown. 711 Dep = MemDepResult::getNonLocal(); 712 } else { 713 Dep = MemDepResult::getNonFuncLocal(); 714 } 715 716 // If we had a dirty entry for the block, update it. Otherwise, just add 717 // a new entry. 718 if (ExistingResult) 719 ExistingResult->setResult(Dep); 720 else 721 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 722 723 // If the block has a dependency (i.e. it isn't completely transparent to 724 // the value), remember the association! 725 if (!Dep.isNonLocal()) { 726 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 727 // update this when we remove instructions. 728 if (Instruction *Inst = Dep.getInst()) 729 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); 730 } else { 731 732 // If the block *is* completely transparent to the load, we need to check 733 // the predecessors of this block. Add them to our worklist. 734 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) 735 DirtyBlocks.push_back(*PI); 736 } 737 } 738 739 return Cache; 740} 741 742/// getNonLocalPointerDependency - Perform a full dependency query for an 743/// access to the specified (non-volatile) memory location, returning the 744/// set of instructions that either define or clobber the value. 745/// 746/// This method assumes the pointer has a "NonLocal" dependency within its 747/// own block. 748/// 749void MemoryDependenceAnalysis:: 750getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad, 751 BasicBlock *FromBB, 752 SmallVectorImpl<NonLocalDepResult> &Result) { 753 assert(Loc.Ptr->getType()->isPointerTy() && 754 "Can't get pointer deps of a non-pointer!"); 755 Result.clear(); 756 757 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD); 758 759 // This is the set of blocks we've inspected, and the pointer we consider in 760 // each block. Because of critical edges, we currently bail out if querying 761 // a block with multiple different pointers. This can happen during PHI 762 // translation. 763 DenseMap<BasicBlock*, Value*> Visited; 764 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB, 765 Result, Visited, true)) 766 return; 767 Result.clear(); 768 Result.push_back(NonLocalDepResult(FromBB, 769 MemDepResult::getUnknown(), 770 const_cast<Value *>(Loc.Ptr))); 771} 772 773/// GetNonLocalInfoForBlock - Compute the memdep value for BB with 774/// Pointer/PointeeSize using either cached information in Cache or by doing a 775/// lookup (which may use dirty cache info if available). If we do a lookup, 776/// add the result to the cache. 777MemDepResult MemoryDependenceAnalysis:: 778GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc, 779 bool isLoad, BasicBlock *BB, 780 NonLocalDepInfo *Cache, unsigned NumSortedEntries) { 781 782 // Do a binary search to see if we already have an entry for this block in 783 // the cache set. If so, find it. 784 NonLocalDepInfo::iterator Entry = 785 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, 786 NonLocalDepEntry(BB)); 787 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB) 788 --Entry; 789 790 NonLocalDepEntry *ExistingResult = 0; 791 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB) 792 ExistingResult = &*Entry; 793 794 // If we have a cached entry, and it is non-dirty, use it as the value for 795 // this dependency. 796 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 797 ++NumCacheNonLocalPtr; 798 return ExistingResult->getResult(); 799 } 800 801 // Otherwise, we have to scan for the value. If we have a dirty cache 802 // entry, start scanning from its position, otherwise we scan from the end 803 // of the block. 804 BasicBlock::iterator ScanPos = BB->end(); 805 if (ExistingResult && ExistingResult->getResult().getInst()) { 806 assert(ExistingResult->getResult().getInst()->getParent() == BB && 807 "Instruction invalidated?"); 808 ++NumCacheDirtyNonLocalPtr; 809 ScanPos = ExistingResult->getResult().getInst(); 810 811 // Eliminating the dirty entry from 'Cache', so update the reverse info. 812 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 813 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); 814 } else { 815 ++NumUncacheNonLocalPtr; 816 } 817 818 // Scan the block for the dependency. 819 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB); 820 821 // If we had a dirty entry for the block, update it. Otherwise, just add 822 // a new entry. 823 if (ExistingResult) 824 ExistingResult->setResult(Dep); 825 else 826 Cache->push_back(NonLocalDepEntry(BB, Dep)); 827 828 // If the block has a dependency (i.e. it isn't completely transparent to 829 // the value), remember the reverse association because we just added it 830 // to Cache! 831 if (!Dep.isDef() && !Dep.isClobber()) 832 return Dep; 833 834 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 835 // update MemDep when we remove instructions. 836 Instruction *Inst = Dep.getInst(); 837 assert(Inst && "Didn't depend on anything?"); 838 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 839 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 840 return Dep; 841} 842 843/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain 844/// number of elements in the array that are already properly ordered. This is 845/// optimized for the case when only a few entries are added. 846static void 847SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, 848 unsigned NumSortedEntries) { 849 switch (Cache.size() - NumSortedEntries) { 850 case 0: 851 // done, no new entries. 852 break; 853 case 2: { 854 // Two new entries, insert the last one into place. 855 NonLocalDepEntry Val = Cache.back(); 856 Cache.pop_back(); 857 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 858 std::upper_bound(Cache.begin(), Cache.end()-1, Val); 859 Cache.insert(Entry, Val); 860 // FALL THROUGH. 861 } 862 case 1: 863 // One new entry, Just insert the new value at the appropriate position. 864 if (Cache.size() != 1) { 865 NonLocalDepEntry Val = Cache.back(); 866 Cache.pop_back(); 867 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = 868 std::upper_bound(Cache.begin(), Cache.end(), Val); 869 Cache.insert(Entry, Val); 870 } 871 break; 872 default: 873 // Added many values, do a full scale sort. 874 std::sort(Cache.begin(), Cache.end()); 875 break; 876 } 877} 878 879/// getNonLocalPointerDepFromBB - Perform a dependency query based on 880/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def 881/// results to the results vector and keep track of which blocks are visited in 882/// 'Visited'. 883/// 884/// This has special behavior for the first block queries (when SkipFirstBlock 885/// is true). In this special case, it ignores the contents of the specified 886/// block and starts returning dependence info for its predecessors. 887/// 888/// This function returns false on success, or true to indicate that it could 889/// not compute dependence information for some reason. This should be treated 890/// as a clobber dependence on the first instruction in the predecessor block. 891bool MemoryDependenceAnalysis:: 892getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, 893 const AliasAnalysis::Location &Loc, 894 bool isLoad, BasicBlock *StartBB, 895 SmallVectorImpl<NonLocalDepResult> &Result, 896 DenseMap<BasicBlock*, Value*> &Visited, 897 bool SkipFirstBlock) { 898 899 // Look up the cached info for Pointer. 900 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 901 902 // Set up a temporary NLPI value. If the map doesn't yet have an entry for 903 // CacheKey, this value will be inserted as the associated value. Otherwise, 904 // it'll be ignored, and we'll have to check to see if the cached size and 905 // tbaa tag are consistent with the current query. 906 NonLocalPointerInfo InitialNLPI; 907 InitialNLPI.Size = Loc.Size; 908 InitialNLPI.TBAATag = Loc.TBAATag; 909 910 // Get the NLPI for CacheKey, inserting one into the map if it doesn't 911 // already have one. 912 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair = 913 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI)); 914 NonLocalPointerInfo *CacheInfo = &Pair.first->second; 915 916 // If we already have a cache entry for this CacheKey, we may need to do some 917 // work to reconcile the cache entry and the current query. 918 if (!Pair.second) { 919 if (CacheInfo->Size < Loc.Size) { 920 // The query's Size is greater than the cached one. Throw out the 921 // cached data and proceed with the query at the greater size. 922 CacheInfo->Pair = BBSkipFirstBlockPair(); 923 CacheInfo->Size = Loc.Size; 924 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 925 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 926 if (Instruction *Inst = DI->getResult().getInst()) 927 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 928 CacheInfo->NonLocalDeps.clear(); 929 } else if (CacheInfo->Size > Loc.Size) { 930 // This query's Size is less than the cached one. Conservatively restart 931 // the query using the greater size. 932 return getNonLocalPointerDepFromBB(Pointer, 933 Loc.getWithNewSize(CacheInfo->Size), 934 isLoad, StartBB, Result, Visited, 935 SkipFirstBlock); 936 } 937 938 // If the query's TBAATag is inconsistent with the cached one, 939 // conservatively throw out the cached data and restart the query with 940 // no tag if needed. 941 if (CacheInfo->TBAATag != Loc.TBAATag) { 942 if (CacheInfo->TBAATag) { 943 CacheInfo->Pair = BBSkipFirstBlockPair(); 944 CacheInfo->TBAATag = 0; 945 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), 946 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) 947 if (Instruction *Inst = DI->getResult().getInst()) 948 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 949 CacheInfo->NonLocalDeps.clear(); 950 } 951 if (Loc.TBAATag) 952 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(), 953 isLoad, StartBB, Result, Visited, 954 SkipFirstBlock); 955 } 956 } 957 958 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps; 959 960 // If we have valid cached information for exactly the block we are 961 // investigating, just return it with no recomputation. 962 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 963 // We have a fully cached result for this query then we can just return the 964 // cached results and populate the visited set. However, we have to verify 965 // that we don't already have conflicting results for these blocks. Check 966 // to ensure that if a block in the results set is in the visited set that 967 // it was for the same pointer query. 968 if (!Visited.empty()) { 969 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 970 I != E; ++I) { 971 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB()); 972 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 973 continue; 974 975 // We have a pointer mismatch in a block. Just return clobber, saying 976 // that something was clobbered in this result. We could also do a 977 // non-fully cached query, but there is little point in doing this. 978 return true; 979 } 980 } 981 982 Value *Addr = Pointer.getAddr(); 983 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); 984 I != E; ++I) { 985 Visited.insert(std::make_pair(I->getBB(), Addr)); 986 if (!I->getResult().isNonLocal() && DT->isReachableFromEntry(I->getBB())) 987 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); 988 } 989 ++NumCacheCompleteNonLocalPtr; 990 return false; 991 } 992 993 // Otherwise, either this is a new block, a block with an invalid cache 994 // pointer or one that we're about to invalidate by putting more info into it 995 // than its valid cache info. If empty, the result will be valid cache info, 996 // otherwise it isn't. 997 if (Cache->empty()) 998 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 999 else 1000 CacheInfo->Pair = BBSkipFirstBlockPair(); 1001 1002 SmallVector<BasicBlock*, 32> Worklist; 1003 Worklist.push_back(StartBB); 1004 1005 // PredList used inside loop. 1006 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList; 1007 1008 // Keep track of the entries that we know are sorted. Previously cached 1009 // entries will all be sorted. The entries we add we only sort on demand (we 1010 // don't insert every element into its sorted position). We know that we 1011 // won't get any reuse from currently inserted values, because we don't 1012 // revisit blocks after we insert info for them. 1013 unsigned NumSortedEntries = Cache->size(); 1014 DEBUG(AssertSorted(*Cache)); 1015 1016 while (!Worklist.empty()) { 1017 BasicBlock *BB = Worklist.pop_back_val(); 1018 1019 // Skip the first block if we have it. 1020 if (!SkipFirstBlock) { 1021 // Analyze the dependency of *Pointer in FromBB. See if we already have 1022 // been here. 1023 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 1024 1025 // Get the dependency info for Pointer in BB. If we have cached 1026 // information, we will use it, otherwise we compute it. 1027 DEBUG(AssertSorted(*Cache, NumSortedEntries)); 1028 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache, 1029 NumSortedEntries); 1030 1031 // If we got a Def or Clobber, add this to the list of results. 1032 if (!Dep.isNonLocal() && DT->isReachableFromEntry(BB)) { 1033 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 1034 continue; 1035 } 1036 } 1037 1038 // If 'Pointer' is an instruction defined in this block, then we need to do 1039 // phi translation to change it into a value live in the predecessor block. 1040 // If not, we just add the predecessors to the worklist and scan them with 1041 // the same Pointer. 1042 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 1043 SkipFirstBlock = false; 1044 SmallVector<BasicBlock*, 16> NewBlocks; 1045 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1046 // Verify that we haven't looked at this block yet. 1047 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1048 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr())); 1049 if (InsertRes.second) { 1050 // First time we've looked at *PI. 1051 NewBlocks.push_back(*PI); 1052 continue; 1053 } 1054 1055 // If we have seen this block before, but it was with a different 1056 // pointer then we have a phi translation failure and we have to treat 1057 // this as a clobber. 1058 if (InsertRes.first->second != Pointer.getAddr()) { 1059 // Make sure to clean up the Visited map before continuing on to 1060 // PredTranslationFailure. 1061 for (unsigned i = 0; i < NewBlocks.size(); i++) 1062 Visited.erase(NewBlocks[i]); 1063 goto PredTranslationFailure; 1064 } 1065 } 1066 Worklist.append(NewBlocks.begin(), NewBlocks.end()); 1067 continue; 1068 } 1069 1070 // We do need to do phi translation, if we know ahead of time we can't phi 1071 // translate this value, don't even try. 1072 if (!Pointer.IsPotentiallyPHITranslatable()) 1073 goto PredTranslationFailure; 1074 1075 // We may have added values to the cache list before this PHI translation. 1076 // If so, we haven't done anything to ensure that the cache remains sorted. 1077 // Sort it now (if needed) so that recursive invocations of 1078 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 1079 // value will only see properly sorted cache arrays. 1080 if (Cache && NumSortedEntries != Cache->size()) { 1081 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1082 NumSortedEntries = Cache->size(); 1083 } 1084 Cache = 0; 1085 1086 PredList.clear(); 1087 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { 1088 BasicBlock *Pred = *PI; 1089 PredList.push_back(std::make_pair(Pred, Pointer)); 1090 1091 // Get the PHI translated pointer in this predecessor. This can fail if 1092 // not translatable, in which case the getAddr() returns null. 1093 PHITransAddr &PredPointer = PredList.back().second; 1094 PredPointer.PHITranslateValue(BB, Pred, 0); 1095 1096 Value *PredPtrVal = PredPointer.getAddr(); 1097 1098 // Check to see if we have already visited this pred block with another 1099 // pointer. If so, we can't do this lookup. This failure can occur 1100 // with PHI translation when a critical edge exists and the PHI node in 1101 // the successor translates to a pointer value different than the 1102 // pointer the block was first analyzed with. 1103 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool> 1104 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); 1105 1106 if (!InsertRes.second) { 1107 // We found the pred; take it off the list of preds to visit. 1108 PredList.pop_back(); 1109 1110 // If the predecessor was visited with PredPtr, then we already did 1111 // the analysis and can ignore it. 1112 if (InsertRes.first->second == PredPtrVal) 1113 continue; 1114 1115 // Otherwise, the block was previously analyzed with a different 1116 // pointer. We can't represent the result of this case, so we just 1117 // treat this as a phi translation failure. 1118 1119 // Make sure to clean up the Visited map before continuing on to 1120 // PredTranslationFailure. 1121 for (unsigned i = 0; i < PredList.size(); i++) 1122 Visited.erase(PredList[i].first); 1123 1124 goto PredTranslationFailure; 1125 } 1126 } 1127 1128 // Actually process results here; this need to be a separate loop to avoid 1129 // calling getNonLocalPointerDepFromBB for blocks we don't want to return 1130 // any results for. (getNonLocalPointerDepFromBB will modify our 1131 // datastructures in ways the code after the PredTranslationFailure label 1132 // doesn't expect.) 1133 for (unsigned i = 0; i < PredList.size(); i++) { 1134 BasicBlock *Pred = PredList[i].first; 1135 PHITransAddr &PredPointer = PredList[i].second; 1136 Value *PredPtrVal = PredPointer.getAddr(); 1137 1138 bool CanTranslate = true; 1139 // If PHI translation was unable to find an available pointer in this 1140 // predecessor, then we have to assume that the pointer is clobbered in 1141 // that predecessor. We can still do PRE of the load, which would insert 1142 // a computation of the pointer in this predecessor. 1143 if (PredPtrVal == 0) 1144 CanTranslate = false; 1145 1146 // FIXME: it is entirely possible that PHI translating will end up with 1147 // the same value. Consider PHI translating something like: 1148 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 1149 // to recurse here, pedantically speaking. 1150 1151 // If getNonLocalPointerDepFromBB fails here, that means the cached 1152 // result conflicted with the Visited list; we have to conservatively 1153 // assume it is unknown, but this also does not block PRE of the load. 1154 if (!CanTranslate || 1155 getNonLocalPointerDepFromBB(PredPointer, 1156 Loc.getWithNewPtr(PredPtrVal), 1157 isLoad, Pred, 1158 Result, Visited)) { 1159 // Add the entry to the Result list. 1160 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal); 1161 Result.push_back(Entry); 1162 1163 // Since we had a phi translation failure, the cache for CacheKey won't 1164 // include all of the entries that we need to immediately satisfy future 1165 // queries. Mark this in NonLocalPointerDeps by setting the 1166 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 1167 // cached value to do more work but not miss the phi trans failure. 1168 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey]; 1169 NLPI.Pair = BBSkipFirstBlockPair(); 1170 continue; 1171 } 1172 } 1173 1174 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 1175 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1176 Cache = &CacheInfo->NonLocalDeps; 1177 NumSortedEntries = Cache->size(); 1178 1179 // Since we did phi translation, the "Cache" set won't contain all of the 1180 // results for the query. This is ok (we can still use it to accelerate 1181 // specific block queries) but we can't do the fastpath "return all 1182 // results from the set" Clear out the indicator for this. 1183 CacheInfo->Pair = BBSkipFirstBlockPair(); 1184 SkipFirstBlock = false; 1185 continue; 1186 1187 PredTranslationFailure: 1188 // The following code is "failure"; we can't produce a sane translation 1189 // for the given block. It assumes that we haven't modified any of 1190 // our datastructures while processing the current block. 1191 1192 if (Cache == 0) { 1193 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1194 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1195 Cache = &CacheInfo->NonLocalDeps; 1196 NumSortedEntries = Cache->size(); 1197 } 1198 1199 // Since we failed phi translation, the "Cache" set won't contain all of the 1200 // results for the query. This is ok (we can still use it to accelerate 1201 // specific block queries) but we can't do the fastpath "return all 1202 // results from the set". Clear out the indicator for this. 1203 CacheInfo->Pair = BBSkipFirstBlockPair(); 1204 1205 // If *nothing* works, mark the pointer as unknown. 1206 // 1207 // If this is the magic first block, return this as a clobber of the whole 1208 // incoming value. Since we can't phi translate to one of the predecessors, 1209 // we have to bail out. 1210 if (SkipFirstBlock) 1211 return true; 1212 1213 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { 1214 assert(I != Cache->rend() && "Didn't find current block??"); 1215 if (I->getBB() != BB) 1216 continue; 1217 1218 assert(I->getResult().isNonLocal() && 1219 "Should only be here with transparent block"); 1220 I->setResult(MemDepResult::getUnknown()); 1221 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), 1222 Pointer.getAddr())); 1223 break; 1224 } 1225 } 1226 1227 // Okay, we're done now. If we added new values to the cache, re-sort it. 1228 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1229 DEBUG(AssertSorted(*Cache)); 1230 return false; 1231} 1232 1233/// RemoveCachedNonLocalPointerDependencies - If P exists in 1234/// CachedNonLocalPointerInfo, remove it. 1235void MemoryDependenceAnalysis:: 1236RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { 1237 CachedNonLocalPointerInfo::iterator It = 1238 NonLocalPointerDeps.find(P); 1239 if (It == NonLocalPointerDeps.end()) return; 1240 1241 // Remove all of the entries in the BB->val map. This involves removing 1242 // instructions from the reverse map. 1243 NonLocalDepInfo &PInfo = It->second.NonLocalDeps; 1244 1245 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { 1246 Instruction *Target = PInfo[i].getResult().getInst(); 1247 if (Target == 0) continue; // Ignore non-local dep results. 1248 assert(Target->getParent() == PInfo[i].getBB()); 1249 1250 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1251 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1252 } 1253 1254 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1255 NonLocalPointerDeps.erase(It); 1256} 1257 1258 1259/// invalidateCachedPointerInfo - This method is used to invalidate cached 1260/// information about the specified pointer, because it may be too 1261/// conservative in memdep. This is an optional call that can be used when 1262/// the client detects an equivalence between the pointer and some other 1263/// value and replaces the other value with ptr. This can make Ptr available 1264/// in more places that cached info does not necessarily keep. 1265void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { 1266 // If Ptr isn't really a pointer, just ignore it. 1267 if (!Ptr->getType()->isPointerTy()) return; 1268 // Flush store info for the pointer. 1269 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1270 // Flush load info for the pointer. 1271 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1272} 1273 1274/// invalidateCachedPredecessors - Clear the PredIteratorCache info. 1275/// This needs to be done when the CFG changes, e.g., due to splitting 1276/// critical edges. 1277void MemoryDependenceAnalysis::invalidateCachedPredecessors() { 1278 PredCache->clear(); 1279} 1280 1281/// removeInstruction - Remove an instruction from the dependence analysis, 1282/// updating the dependence of instructions that previously depended on it. 1283/// This method attempts to keep the cache coherent using the reverse map. 1284void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { 1285 // Walk through the Non-local dependencies, removing this one as the value 1286 // for any cached queries. 1287 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); 1288 if (NLDI != NonLocalDeps.end()) { 1289 NonLocalDepInfo &BlockMap = NLDI->second.first; 1290 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); 1291 DI != DE; ++DI) 1292 if (Instruction *Inst = DI->getResult().getInst()) 1293 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1294 NonLocalDeps.erase(NLDI); 1295 } 1296 1297 // If we have a cached local dependence query for this instruction, remove it. 1298 // 1299 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1300 if (LocalDepEntry != LocalDeps.end()) { 1301 // Remove us from DepInst's reverse set now that the local dep info is gone. 1302 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1303 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1304 1305 // Remove this local dependency info. 1306 LocalDeps.erase(LocalDepEntry); 1307 } 1308 1309 // If we have any cached pointer dependencies on this instruction, remove 1310 // them. If the instruction has non-pointer type, then it can't be a pointer 1311 // base. 1312 1313 // Remove it from both the load info and the store info. The instruction 1314 // can't be in either of these maps if it is non-pointer. 1315 if (RemInst->getType()->isPointerTy()) { 1316 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1317 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1318 } 1319 1320 // Loop over all of the things that depend on the instruction we're removing. 1321 // 1322 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd; 1323 1324 // If we find RemInst as a clobber or Def in any of the maps for other values, 1325 // we need to replace its entry with a dirty version of the instruction after 1326 // it. If RemInst is a terminator, we use a null dirty value. 1327 // 1328 // Using a dirty version of the instruction after RemInst saves having to scan 1329 // the entire block to get to this point. 1330 MemDepResult NewDirtyVal; 1331 if (!RemInst->isTerminator()) 1332 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); 1333 1334 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1335 if (ReverseDepIt != ReverseLocalDeps.end()) { 1336 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second; 1337 // RemInst can't be the terminator if it has local stuff depending on it. 1338 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) && 1339 "Nothing can locally depend on a terminator"); 1340 1341 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(), 1342 E = ReverseDeps.end(); I != E; ++I) { 1343 Instruction *InstDependingOnRemInst = *I; 1344 assert(InstDependingOnRemInst != RemInst && 1345 "Already removed our local dep info"); 1346 1347 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1348 1349 // Make sure to remember that new things depend on NewDepInst. 1350 assert(NewDirtyVal.getInst() && "There is no way something else can have " 1351 "a local dep on this if it is a terminator!"); 1352 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), 1353 InstDependingOnRemInst)); 1354 } 1355 1356 ReverseLocalDeps.erase(ReverseDepIt); 1357 1358 // Add new reverse deps after scanning the set, to avoid invalidating the 1359 // 'ReverseDeps' reference. 1360 while (!ReverseDepsToAdd.empty()) { 1361 ReverseLocalDeps[ReverseDepsToAdd.back().first] 1362 .insert(ReverseDepsToAdd.back().second); 1363 ReverseDepsToAdd.pop_back(); 1364 } 1365 } 1366 1367 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1368 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1369 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second; 1370 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end(); 1371 I != E; ++I) { 1372 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); 1373 1374 PerInstNLInfo &INLD = NonLocalDeps[*I]; 1375 // The information is now dirty! 1376 INLD.second = true; 1377 1378 for (NonLocalDepInfo::iterator DI = INLD.first.begin(), 1379 DE = INLD.first.end(); DI != DE; ++DI) { 1380 if (DI->getResult().getInst() != RemInst) continue; 1381 1382 // Convert to a dirty entry for the subsequent instruction. 1383 DI->setResult(NewDirtyVal); 1384 1385 if (Instruction *NextI = NewDirtyVal.getInst()) 1386 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); 1387 } 1388 } 1389 1390 ReverseNonLocalDeps.erase(ReverseDepIt); 1391 1392 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1393 while (!ReverseDepsToAdd.empty()) { 1394 ReverseNonLocalDeps[ReverseDepsToAdd.back().first] 1395 .insert(ReverseDepsToAdd.back().second); 1396 ReverseDepsToAdd.pop_back(); 1397 } 1398 } 1399 1400 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1401 // value in the NonLocalPointerDeps info. 1402 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1403 ReverseNonLocalPtrDeps.find(RemInst); 1404 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1405 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second; 1406 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd; 1407 1408 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(), 1409 E = Set.end(); I != E; ++I) { 1410 ValueIsLoadPair P = *I; 1411 assert(P.getPointer() != RemInst && 1412 "Already removed NonLocalPointerDeps info for RemInst"); 1413 1414 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps; 1415 1416 // The cache is not valid for any specific block anymore. 1417 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair(); 1418 1419 // Update any entries for RemInst to use the instruction after it. 1420 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); 1421 DI != DE; ++DI) { 1422 if (DI->getResult().getInst() != RemInst) continue; 1423 1424 // Convert to a dirty entry for the subsequent instruction. 1425 DI->setResult(NewDirtyVal); 1426 1427 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1428 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1429 } 1430 1431 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1432 // subsequent value may invalidate the sortedness. 1433 std::sort(NLPDI.begin(), NLPDI.end()); 1434 } 1435 1436 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1437 1438 while (!ReversePtrDepsToAdd.empty()) { 1439 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] 1440 .insert(ReversePtrDepsToAdd.back().second); 1441 ReversePtrDepsToAdd.pop_back(); 1442 } 1443 } 1444 1445 1446 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); 1447 AA->deleteValue(RemInst); 1448 DEBUG(verifyRemoved(RemInst)); 1449} 1450/// verifyRemoved - Verify that the specified instruction does not occur 1451/// in our internal data structures. 1452void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { 1453 for (LocalDepMapType::const_iterator I = LocalDeps.begin(), 1454 E = LocalDeps.end(); I != E; ++I) { 1455 assert(I->first != D && "Inst occurs in data structures"); 1456 assert(I->second.getInst() != D && 1457 "Inst occurs in data structures"); 1458 } 1459 1460 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), 1461 E = NonLocalPointerDeps.end(); I != E; ++I) { 1462 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 1463 const NonLocalDepInfo &Val = I->second.NonLocalDeps; 1464 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); 1465 II != E; ++II) 1466 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); 1467 } 1468 1469 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), 1470 E = NonLocalDeps.end(); I != E; ++I) { 1471 assert(I->first != D && "Inst occurs in data structures"); 1472 const PerInstNLInfo &INLD = I->second; 1473 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), 1474 EE = INLD.first.end(); II != EE; ++II) 1475 assert(II->getResult().getInst() != D && "Inst occurs in data structures"); 1476 } 1477 1478 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), 1479 E = ReverseLocalDeps.end(); I != E; ++I) { 1480 assert(I->first != D && "Inst occurs in data structures"); 1481 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1482 EE = I->second.end(); II != EE; ++II) 1483 assert(*II != D && "Inst occurs in data structures"); 1484 } 1485 1486 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), 1487 E = ReverseNonLocalDeps.end(); 1488 I != E; ++I) { 1489 assert(I->first != D && "Inst occurs in data structures"); 1490 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(), 1491 EE = I->second.end(); II != EE; ++II) 1492 assert(*II != D && "Inst occurs in data structures"); 1493 } 1494 1495 for (ReverseNonLocalPtrDepTy::const_iterator 1496 I = ReverseNonLocalPtrDeps.begin(), 1497 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { 1498 assert(I->first != D && "Inst occurs in rev NLPD map"); 1499 1500 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(), 1501 E = I->second.end(); II != E; ++II) 1502 assert(*II != ValueIsLoadPair(D, false) && 1503 *II != ValueIsLoadPair(D, true) && 1504 "Inst occurs in ReverseNonLocalPtrDeps map"); 1505 } 1506 1507} 1508