1 //===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation -------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements an analysis that determines, for a given memory 10 // operation, what preceding memory operations it depends on. It builds on 11 // alias analysis information, and tries to provide a lazy, caching interface to 12 // a common kind of alias information query. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "llvm/Analysis/MemoryDependenceAnalysis.h" 17 #include "llvm/ADT/DenseMap.h" 18 #include "llvm/ADT/STLExtras.h" 19 #include "llvm/ADT/SmallPtrSet.h" 20 #include "llvm/ADT/SmallVector.h" 21 #include "llvm/ADT/Statistic.h" 22 #include "llvm/Analysis/AliasAnalysis.h" 23 #include "llvm/Analysis/AssumptionCache.h" 24 #include "llvm/Analysis/MemoryBuiltins.h" 25 #include "llvm/Analysis/MemoryLocation.h" 26 #include "llvm/Analysis/PHITransAddr.h" 27 #include "llvm/Analysis/PhiValues.h" 28 #include "llvm/Analysis/TargetLibraryInfo.h" 29 #include "llvm/Analysis/ValueTracking.h" 30 #include "llvm/IR/BasicBlock.h" 31 #include "llvm/IR/Dominators.h" 32 #include "llvm/IR/Function.h" 33 #include "llvm/IR/InstrTypes.h" 34 #include "llvm/IR/Instruction.h" 35 #include "llvm/IR/Instructions.h" 36 #include "llvm/IR/IntrinsicInst.h" 37 #include "llvm/IR/LLVMContext.h" 38 #include "llvm/IR/Metadata.h" 39 #include "llvm/IR/Module.h" 40 #include "llvm/IR/PredIteratorCache.h" 41 #include "llvm/IR/Type.h" 42 #include "llvm/IR/Use.h" 43 #include "llvm/IR/Value.h" 44 #include "llvm/InitializePasses.h" 45 #include "llvm/Pass.h" 46 #include "llvm/Support/AtomicOrdering.h" 47 #include "llvm/Support/Casting.h" 48 #include "llvm/Support/CommandLine.h" 49 #include "llvm/Support/Compiler.h" 50 #include "llvm/Support/Debug.h" 51 #include <algorithm> 52 #include <cassert> 53 #include <iterator> 54 #include <utility> 55 56 using namespace llvm; 57 58 #define DEBUG_TYPE "memdep" 59 60 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); 61 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); 62 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); 63 64 STATISTIC(NumCacheNonLocalPtr, 65 "Number of fully cached non-local ptr responses"); 66 STATISTIC(NumCacheDirtyNonLocalPtr, 67 "Number of cached, but dirty, non-local ptr responses"); 68 STATISTIC(NumUncacheNonLocalPtr, "Number of uncached non-local ptr responses"); 69 STATISTIC(NumCacheCompleteNonLocalPtr, 70 "Number of block queries that were completely cached"); 71 72 // Limit for the number of instructions to scan in a block. 73 74 static cl::opt<unsigned> BlockScanLimit( 75 "memdep-block-scan-limit", cl::Hidden, cl::init(100), 76 cl::desc("The number of instructions to scan in a block in memory " 77 "dependency analysis (default = 100)")); 78 79 static cl::opt<unsigned> 80 BlockNumberLimit("memdep-block-number-limit", cl::Hidden, cl::init(1000), 81 cl::desc("The number of blocks to scan during memory " 82 "dependency analysis (default = 1000)")); 83 84 // Limit on the number of memdep results to process. 85 static const unsigned int NumResultsLimit = 100; 86 87 /// This is a helper function that removes Val from 'Inst's set in ReverseMap. 88 /// 89 /// If the set becomes empty, remove Inst's entry. 90 template <typename KeyTy> 91 static void 92 RemoveFromReverseMap(DenseMap<Instruction *, SmallPtrSet<KeyTy, 4>> &ReverseMap, 93 Instruction *Inst, KeyTy Val) { 94 typename DenseMap<Instruction *, SmallPtrSet<KeyTy, 4>>::iterator InstIt = 95 ReverseMap.find(Inst); 96 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); 97 bool Found = InstIt->second.erase(Val); 98 assert(Found && "Invalid reverse map!"); 99 (void)Found; 100 if (InstIt->second.empty()) 101 ReverseMap.erase(InstIt); 102 } 103 104 /// If the given instruction references a specific memory location, fill in Loc 105 /// with the details, otherwise set Loc.Ptr to null. 106 /// 107 /// Returns a ModRefInfo value describing the general behavior of the 108 /// instruction. 109 static ModRefInfo GetLocation(const Instruction *Inst, MemoryLocation &Loc, 110 const TargetLibraryInfo &TLI) { 111 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 112 if (LI->isUnordered()) { 113 Loc = MemoryLocation::get(LI); 114 return ModRefInfo::Ref; 115 } 116 if (LI->getOrdering() == AtomicOrdering::Monotonic) { 117 Loc = MemoryLocation::get(LI); 118 return ModRefInfo::ModRef; 119 } 120 Loc = MemoryLocation(); 121 return ModRefInfo::ModRef; 122 } 123 124 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 125 if (SI->isUnordered()) { 126 Loc = MemoryLocation::get(SI); 127 return ModRefInfo::Mod; 128 } 129 if (SI->getOrdering() == AtomicOrdering::Monotonic) { 130 Loc = MemoryLocation::get(SI); 131 return ModRefInfo::ModRef; 132 } 133 Loc = MemoryLocation(); 134 return ModRefInfo::ModRef; 135 } 136 137 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { 138 Loc = MemoryLocation::get(V); 139 return ModRefInfo::ModRef; 140 } 141 142 if (const CallBase *CB = dyn_cast<CallBase>(Inst)) { 143 if (Value *FreedOp = getFreedOperand(CB, &TLI)) { 144 // calls to free() deallocate the entire structure 145 Loc = MemoryLocation::getAfter(FreedOp); 146 return ModRefInfo::Mod; 147 } 148 } 149 150 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 151 switch (II->getIntrinsicID()) { 152 case Intrinsic::lifetime_start: 153 case Intrinsic::lifetime_end: 154 case Intrinsic::invariant_start: 155 Loc = MemoryLocation::getForArgument(II, 1, TLI); 156 // These intrinsics don't really modify the memory, but returning Mod 157 // will allow them to be handled conservatively. 158 return ModRefInfo::Mod; 159 case Intrinsic::invariant_end: 160 Loc = MemoryLocation::getForArgument(II, 2, TLI); 161 // These intrinsics don't really modify the memory, but returning Mod 162 // will allow them to be handled conservatively. 163 return ModRefInfo::Mod; 164 case Intrinsic::masked_load: 165 Loc = MemoryLocation::getForArgument(II, 0, TLI); 166 return ModRefInfo::Ref; 167 case Intrinsic::masked_store: 168 Loc = MemoryLocation::getForArgument(II, 1, TLI); 169 return ModRefInfo::Mod; 170 default: 171 break; 172 } 173 } 174 175 // Otherwise, just do the coarse-grained thing that always works. 176 if (Inst->mayWriteToMemory()) 177 return ModRefInfo::ModRef; 178 if (Inst->mayReadFromMemory()) 179 return ModRefInfo::Ref; 180 return ModRefInfo::NoModRef; 181 } 182 183 /// Private helper for finding the local dependencies of a call site. 184 MemDepResult MemoryDependenceResults::getCallDependencyFrom( 185 CallBase *Call, bool isReadOnlyCall, BasicBlock::iterator ScanIt, 186 BasicBlock *BB) { 187 unsigned Limit = getDefaultBlockScanLimit(); 188 189 // Walk backwards through the block, looking for dependencies. 190 while (ScanIt != BB->begin()) { 191 Instruction *Inst = &*--ScanIt; 192 // Debug intrinsics don't cause dependences and should not affect Limit 193 if (isa<DbgInfoIntrinsic>(Inst)) 194 continue; 195 196 // Limit the amount of scanning we do so we don't end up with quadratic 197 // running time on extreme testcases. 198 --Limit; 199 if (!Limit) 200 return MemDepResult::getUnknown(); 201 202 // If this inst is a memory op, get the pointer it accessed 203 MemoryLocation Loc; 204 ModRefInfo MR = GetLocation(Inst, Loc, TLI); 205 if (Loc.Ptr) { 206 // A simple instruction. 207 if (isModOrRefSet(AA.getModRefInfo(Call, Loc))) 208 return MemDepResult::getClobber(Inst); 209 continue; 210 } 211 212 if (auto *CallB = dyn_cast<CallBase>(Inst)) { 213 // If these two calls do not interfere, look past it. 214 if (isNoModRef(AA.getModRefInfo(Call, CallB))) { 215 // If the two calls are the same, return Inst as a Def, so that 216 // Call can be found redundant and eliminated. 217 if (isReadOnlyCall && !isModSet(MR) && 218 Call->isIdenticalToWhenDefined(CallB)) 219 return MemDepResult::getDef(Inst); 220 221 // Otherwise if the two calls don't interact (e.g. CallB is readnone) 222 // keep scanning. 223 continue; 224 } else 225 return MemDepResult::getClobber(Inst); 226 } 227 228 // If we could not obtain a pointer for the instruction and the instruction 229 // touches memory then assume that this is a dependency. 230 if (isModOrRefSet(MR)) 231 return MemDepResult::getClobber(Inst); 232 } 233 234 // No dependence found. If this is the entry block of the function, it is 235 // unknown, otherwise it is non-local. 236 if (BB != &BB->getParent()->getEntryBlock()) 237 return MemDepResult::getNonLocal(); 238 return MemDepResult::getNonFuncLocal(); 239 } 240 241 MemDepResult MemoryDependenceResults::getPointerDependencyFrom( 242 const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt, 243 BasicBlock *BB, Instruction *QueryInst, unsigned *Limit, 244 BatchAAResults &BatchAA) { 245 MemDepResult InvariantGroupDependency = MemDepResult::getUnknown(); 246 if (QueryInst != nullptr) { 247 if (auto *LI = dyn_cast<LoadInst>(QueryInst)) { 248 InvariantGroupDependency = getInvariantGroupPointerDependency(LI, BB); 249 250 if (InvariantGroupDependency.isDef()) 251 return InvariantGroupDependency; 252 } 253 } 254 MemDepResult SimpleDep = getSimplePointerDependencyFrom( 255 MemLoc, isLoad, ScanIt, BB, QueryInst, Limit, BatchAA); 256 if (SimpleDep.isDef()) 257 return SimpleDep; 258 // Non-local invariant group dependency indicates there is non local Def 259 // (it only returns nonLocal if it finds nonLocal def), which is better than 260 // local clobber and everything else. 261 if (InvariantGroupDependency.isNonLocal()) 262 return InvariantGroupDependency; 263 264 assert(InvariantGroupDependency.isUnknown() && 265 "InvariantGroupDependency should be only unknown at this point"); 266 return SimpleDep; 267 } 268 269 MemDepResult MemoryDependenceResults::getPointerDependencyFrom( 270 const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt, 271 BasicBlock *BB, Instruction *QueryInst, unsigned *Limit) { 272 BatchAAResults BatchAA(AA); 273 return getPointerDependencyFrom(MemLoc, isLoad, ScanIt, BB, QueryInst, Limit, 274 BatchAA); 275 } 276 277 MemDepResult 278 MemoryDependenceResults::getInvariantGroupPointerDependency(LoadInst *LI, 279 BasicBlock *BB) { 280 281 if (!LI->hasMetadata(LLVMContext::MD_invariant_group)) 282 return MemDepResult::getUnknown(); 283 284 // Take the ptr operand after all casts and geps 0. This way we can search 285 // cast graph down only. 286 Value *LoadOperand = LI->getPointerOperand()->stripPointerCasts(); 287 288 // It's is not safe to walk the use list of global value, because function 289 // passes aren't allowed to look outside their functions. 290 // FIXME: this could be fixed by filtering instructions from outside 291 // of current function. 292 if (isa<GlobalValue>(LoadOperand)) 293 return MemDepResult::getUnknown(); 294 295 // Queue to process all pointers that are equivalent to load operand. 296 SmallVector<const Value *, 8> LoadOperandsQueue; 297 LoadOperandsQueue.push_back(LoadOperand); 298 299 Instruction *ClosestDependency = nullptr; 300 // Order of instructions in uses list is unpredictible. In order to always 301 // get the same result, we will look for the closest dominance. 302 auto GetClosestDependency = [this](Instruction *Best, Instruction *Other) { 303 assert(Other && "Must call it with not null instruction"); 304 if (Best == nullptr || DT.dominates(Best, Other)) 305 return Other; 306 return Best; 307 }; 308 309 // FIXME: This loop is O(N^2) because dominates can be O(n) and in worst case 310 // we will see all the instructions. This should be fixed in MSSA. 311 while (!LoadOperandsQueue.empty()) { 312 const Value *Ptr = LoadOperandsQueue.pop_back_val(); 313 assert(Ptr && !isa<GlobalValue>(Ptr) && 314 "Null or GlobalValue should not be inserted"); 315 316 for (const Use &Us : Ptr->uses()) { 317 auto *U = dyn_cast<Instruction>(Us.getUser()); 318 if (!U || U == LI || !DT.dominates(U, LI)) 319 continue; 320 321 // Bitcast or gep with zeros are using Ptr. Add to queue to check it's 322 // users. U = bitcast Ptr 323 if (isa<BitCastInst>(U)) { 324 LoadOperandsQueue.push_back(U); 325 continue; 326 } 327 // Gep with zeros is equivalent to bitcast. 328 // FIXME: we are not sure if some bitcast should be canonicalized to gep 0 329 // or gep 0 to bitcast because of SROA, so there are 2 forms. When 330 // typeless pointers will be ready then both cases will be gone 331 // (and this BFS also won't be needed). 332 if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) 333 if (GEP->hasAllZeroIndices()) { 334 LoadOperandsQueue.push_back(U); 335 continue; 336 } 337 338 // If we hit load/store with the same invariant.group metadata (and the 339 // same pointer operand) we can assume that value pointed by pointer 340 // operand didn't change. 341 if ((isa<LoadInst>(U) || 342 (isa<StoreInst>(U) && 343 cast<StoreInst>(U)->getPointerOperand() == Ptr)) && 344 U->hasMetadata(LLVMContext::MD_invariant_group)) 345 ClosestDependency = GetClosestDependency(ClosestDependency, U); 346 } 347 } 348 349 if (!ClosestDependency) 350 return MemDepResult::getUnknown(); 351 if (ClosestDependency->getParent() == BB) 352 return MemDepResult::getDef(ClosestDependency); 353 // Def(U) can't be returned here because it is non-local. If local 354 // dependency won't be found then return nonLocal counting that the 355 // user will call getNonLocalPointerDependency, which will return cached 356 // result. 357 NonLocalDefsCache.try_emplace( 358 LI, NonLocalDepResult(ClosestDependency->getParent(), 359 MemDepResult::getDef(ClosestDependency), nullptr)); 360 ReverseNonLocalDefsCache[ClosestDependency].insert(LI); 361 return MemDepResult::getNonLocal(); 362 } 363 364 MemDepResult MemoryDependenceResults::getSimplePointerDependencyFrom( 365 const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt, 366 BasicBlock *BB, Instruction *QueryInst, unsigned *Limit, 367 BatchAAResults &BatchAA) { 368 bool isInvariantLoad = false; 369 370 unsigned DefaultLimit = getDefaultBlockScanLimit(); 371 if (!Limit) 372 Limit = &DefaultLimit; 373 374 // We must be careful with atomic accesses, as they may allow another thread 375 // to touch this location, clobbering it. We are conservative: if the 376 // QueryInst is not a simple (non-atomic) memory access, we automatically 377 // return getClobber. 378 // If it is simple, we know based on the results of 379 // "Compiler testing via a theory of sound optimisations in the C11/C++11 380 // memory model" in PLDI 2013, that a non-atomic location can only be 381 // clobbered between a pair of a release and an acquire action, with no 382 // access to the location in between. 383 // Here is an example for giving the general intuition behind this rule. 384 // In the following code: 385 // store x 0; 386 // release action; [1] 387 // acquire action; [4] 388 // %val = load x; 389 // It is unsafe to replace %val by 0 because another thread may be running: 390 // acquire action; [2] 391 // store x 42; 392 // release action; [3] 393 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val 394 // being 42. A key property of this program however is that if either 395 // 1 or 4 were missing, there would be a race between the store of 42 396 // either the store of 0 or the load (making the whole program racy). 397 // The paper mentioned above shows that the same property is respected 398 // by every program that can detect any optimization of that kind: either 399 // it is racy (undefined) or there is a release followed by an acquire 400 // between the pair of accesses under consideration. 401 402 // If the load is invariant, we "know" that it doesn't alias *any* write. We 403 // do want to respect mustalias results since defs are useful for value 404 // forwarding, but any mayalias write can be assumed to be noalias. 405 // Arguably, this logic should be pushed inside AliasAnalysis itself. 406 if (isLoad && QueryInst) { 407 LoadInst *LI = dyn_cast<LoadInst>(QueryInst); 408 if (LI && LI->hasMetadata(LLVMContext::MD_invariant_load)) 409 isInvariantLoad = true; 410 } 411 412 // True for volatile instruction. 413 // For Load/Store return true if atomic ordering is stronger than AO, 414 // for other instruction just true if it can read or write to memory. 415 auto isComplexForReordering = [](Instruction * I, AtomicOrdering AO)->bool { 416 if (I->isVolatile()) 417 return true; 418 if (auto *LI = dyn_cast<LoadInst>(I)) 419 return isStrongerThan(LI->getOrdering(), AO); 420 if (auto *SI = dyn_cast<StoreInst>(I)) 421 return isStrongerThan(SI->getOrdering(), AO); 422 return I->mayReadOrWriteMemory(); 423 }; 424 425 // Walk backwards through the basic block, looking for dependencies. 426 while (ScanIt != BB->begin()) { 427 Instruction *Inst = &*--ScanIt; 428 429 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) 430 // Debug intrinsics don't (and can't) cause dependencies. 431 if (isa<DbgInfoIntrinsic>(II)) 432 continue; 433 434 // Limit the amount of scanning we do so we don't end up with quadratic 435 // running time on extreme testcases. 436 --*Limit; 437 if (!*Limit) 438 return MemDepResult::getUnknown(); 439 440 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 441 // If we reach a lifetime begin or end marker, then the query ends here 442 // because the value is undefined. 443 Intrinsic::ID ID = II->getIntrinsicID(); 444 switch (ID) { 445 case Intrinsic::lifetime_start: { 446 // FIXME: This only considers queries directly on the invariant-tagged 447 // pointer, not on query pointers that are indexed off of them. It'd 448 // be nice to handle that at some point (the right approach is to use 449 // GetPointerBaseWithConstantOffset). 450 MemoryLocation ArgLoc = MemoryLocation::getAfter(II->getArgOperand(1)); 451 if (BatchAA.isMustAlias(ArgLoc, MemLoc)) 452 return MemDepResult::getDef(II); 453 continue; 454 } 455 case Intrinsic::masked_load: 456 case Intrinsic::masked_store: { 457 MemoryLocation Loc; 458 /*ModRefInfo MR =*/ GetLocation(II, Loc, TLI); 459 AliasResult R = BatchAA.alias(Loc, MemLoc); 460 if (R == AliasResult::NoAlias) 461 continue; 462 if (R == AliasResult::MustAlias) 463 return MemDepResult::getDef(II); 464 if (ID == Intrinsic::masked_load) 465 continue; 466 return MemDepResult::getClobber(II); 467 } 468 } 469 } 470 471 // Values depend on loads if the pointers are must aliased. This means 472 // that a load depends on another must aliased load from the same value. 473 // One exception is atomic loads: a value can depend on an atomic load that 474 // it does not alias with when this atomic load indicates that another 475 // thread may be accessing the location. 476 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 477 // While volatile access cannot be eliminated, they do not have to clobber 478 // non-aliasing locations, as normal accesses, for example, can be safely 479 // reordered with volatile accesses. 480 if (LI->isVolatile()) { 481 if (!QueryInst) 482 // Original QueryInst *may* be volatile 483 return MemDepResult::getClobber(LI); 484 if (QueryInst->isVolatile()) 485 // Ordering required if QueryInst is itself volatile 486 return MemDepResult::getClobber(LI); 487 // Otherwise, volatile doesn't imply any special ordering 488 } 489 490 // Atomic loads have complications involved. 491 // A Monotonic (or higher) load is OK if the query inst is itself not 492 // atomic. 493 // FIXME: This is overly conservative. 494 if (LI->isAtomic() && isStrongerThanUnordered(LI->getOrdering())) { 495 if (!QueryInst || 496 isComplexForReordering(QueryInst, AtomicOrdering::NotAtomic)) 497 return MemDepResult::getClobber(LI); 498 if (LI->getOrdering() != AtomicOrdering::Monotonic) 499 return MemDepResult::getClobber(LI); 500 } 501 502 MemoryLocation LoadLoc = MemoryLocation::get(LI); 503 504 // If we found a pointer, check if it could be the same as our pointer. 505 AliasResult R = BatchAA.alias(LoadLoc, MemLoc); 506 507 if (R == AliasResult::NoAlias) 508 continue; 509 510 if (isLoad) { 511 // Must aliased loads are defs of each other. 512 if (R == AliasResult::MustAlias) 513 return MemDepResult::getDef(Inst); 514 515 // If we have a partial alias, then return this as a clobber for the 516 // client to handle. 517 if (R == AliasResult::PartialAlias && R.hasOffset()) { 518 ClobberOffsets[LI] = R.getOffset(); 519 return MemDepResult::getClobber(Inst); 520 } 521 522 // Random may-alias loads don't depend on each other without a 523 // dependence. 524 continue; 525 } 526 527 // Stores don't alias loads from read-only memory. 528 if (BatchAA.pointsToConstantMemory(LoadLoc)) 529 continue; 530 531 // Stores depend on may/must aliased loads. 532 return MemDepResult::getDef(Inst); 533 } 534 535 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 536 // Atomic stores have complications involved. 537 // A Monotonic store is OK if the query inst is itself not atomic. 538 // FIXME: This is overly conservative. 539 if (!SI->isUnordered() && SI->isAtomic()) { 540 if (!QueryInst || 541 isComplexForReordering(QueryInst, AtomicOrdering::Unordered)) 542 return MemDepResult::getClobber(SI); 543 // Ok, if we are here the guard above guarantee us that 544 // QueryInst is a non-atomic or unordered load/store. 545 // SI is atomic with monotonic or release semantic (seq_cst for store 546 // is actually a release semantic plus total order over other seq_cst 547 // instructions, as soon as QueryInst is not seq_cst we can consider it 548 // as simple release semantic). 549 // Monotonic and Release semantic allows re-ordering before store 550 // so we are safe to go further and check the aliasing. It will prohibit 551 // re-ordering in case locations are may or must alias. 552 } 553 554 // While volatile access cannot be eliminated, they do not have to clobber 555 // non-aliasing locations, as normal accesses can for example be reordered 556 // with volatile accesses. 557 if (SI->isVolatile()) 558 if (!QueryInst || QueryInst->isVolatile()) 559 return MemDepResult::getClobber(SI); 560 561 // If alias analysis can tell that this store is guaranteed to not modify 562 // the query pointer, ignore it. Use getModRefInfo to handle cases where 563 // the query pointer points to constant memory etc. 564 if (!isModOrRefSet(BatchAA.getModRefInfo(SI, MemLoc))) 565 continue; 566 567 // Ok, this store might clobber the query pointer. Check to see if it is 568 // a must alias: in this case, we want to return this as a def. 569 // FIXME: Use ModRefInfo::Must bit from getModRefInfo call above. 570 MemoryLocation StoreLoc = MemoryLocation::get(SI); 571 572 // If we found a pointer, check if it could be the same as our pointer. 573 AliasResult R = BatchAA.alias(StoreLoc, MemLoc); 574 575 if (R == AliasResult::NoAlias) 576 continue; 577 if (R == AliasResult::MustAlias) 578 return MemDepResult::getDef(Inst); 579 if (isInvariantLoad) 580 continue; 581 return MemDepResult::getClobber(Inst); 582 } 583 584 // If this is an allocation, and if we know that the accessed pointer is to 585 // the allocation, return Def. This means that there is no dependence and 586 // the access can be optimized based on that. For example, a load could 587 // turn into undef. Note that we can bypass the allocation itself when 588 // looking for a clobber in many cases; that's an alias property and is 589 // handled by BasicAA. 590 if (isa<AllocaInst>(Inst) || isNoAliasCall(Inst)) { 591 const Value *AccessPtr = getUnderlyingObject(MemLoc.Ptr); 592 if (AccessPtr == Inst || BatchAA.isMustAlias(Inst, AccessPtr)) 593 return MemDepResult::getDef(Inst); 594 } 595 596 if (isInvariantLoad) 597 continue; 598 599 // A release fence requires that all stores complete before it, but does 600 // not prevent the reordering of following loads or stores 'before' the 601 // fence. As a result, we look past it when finding a dependency for 602 // loads. DSE uses this to find preceding stores to delete and thus we 603 // can't bypass the fence if the query instruction is a store. 604 if (FenceInst *FI = dyn_cast<FenceInst>(Inst)) 605 if (isLoad && FI->getOrdering() == AtomicOrdering::Release) 606 continue; 607 608 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. 609 ModRefInfo MR = BatchAA.getModRefInfo(Inst, MemLoc); 610 // If necessary, perform additional analysis. 611 if (isModAndRefSet(MR)) 612 MR = BatchAA.callCapturesBefore(Inst, MemLoc, &DT); 613 switch (clearMust(MR)) { 614 case ModRefInfo::NoModRef: 615 // If the call has no effect on the queried pointer, just ignore it. 616 continue; 617 case ModRefInfo::Mod: 618 return MemDepResult::getClobber(Inst); 619 case ModRefInfo::Ref: 620 // If the call is known to never store to the pointer, and if this is a 621 // load query, we can safely ignore it (scan past it). 622 if (isLoad) 623 continue; 624 LLVM_FALLTHROUGH; 625 default: 626 // Otherwise, there is a potential dependence. Return a clobber. 627 return MemDepResult::getClobber(Inst); 628 } 629 } 630 631 // No dependence found. If this is the entry block of the function, it is 632 // unknown, otherwise it is non-local. 633 if (BB != &BB->getParent()->getEntryBlock()) 634 return MemDepResult::getNonLocal(); 635 return MemDepResult::getNonFuncLocal(); 636 } 637 638 MemDepResult MemoryDependenceResults::getDependency(Instruction *QueryInst) { 639 ClobberOffsets.clear(); 640 Instruction *ScanPos = QueryInst; 641 642 // Check for a cached result 643 MemDepResult &LocalCache = LocalDeps[QueryInst]; 644 645 // If the cached entry is non-dirty, just return it. Note that this depends 646 // on MemDepResult's default constructing to 'dirty'. 647 if (!LocalCache.isDirty()) 648 return LocalCache; 649 650 // Otherwise, if we have a dirty entry, we know we can start the scan at that 651 // instruction, which may save us some work. 652 if (Instruction *Inst = LocalCache.getInst()) { 653 ScanPos = Inst; 654 655 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); 656 } 657 658 BasicBlock *QueryParent = QueryInst->getParent(); 659 660 // Do the scan. 661 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { 662 // No dependence found. If this is the entry block of the function, it is 663 // unknown, otherwise it is non-local. 664 if (QueryParent != &QueryParent->getParent()->getEntryBlock()) 665 LocalCache = MemDepResult::getNonLocal(); 666 else 667 LocalCache = MemDepResult::getNonFuncLocal(); 668 } else { 669 MemoryLocation MemLoc; 670 ModRefInfo MR = GetLocation(QueryInst, MemLoc, TLI); 671 if (MemLoc.Ptr) { 672 // If we can do a pointer scan, make it happen. 673 bool isLoad = !isModSet(MR); 674 if (auto *II = dyn_cast<IntrinsicInst>(QueryInst)) 675 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start; 676 677 LocalCache = 678 getPointerDependencyFrom(MemLoc, isLoad, ScanPos->getIterator(), 679 QueryParent, QueryInst, nullptr); 680 } else if (auto *QueryCall = dyn_cast<CallBase>(QueryInst)) { 681 bool isReadOnly = AA.onlyReadsMemory(QueryCall); 682 LocalCache = getCallDependencyFrom(QueryCall, isReadOnly, 683 ScanPos->getIterator(), QueryParent); 684 } else 685 // Non-memory instruction. 686 LocalCache = MemDepResult::getUnknown(); 687 } 688 689 // Remember the result! 690 if (Instruction *I = LocalCache.getInst()) 691 ReverseLocalDeps[I].insert(QueryInst); 692 693 return LocalCache; 694 } 695 696 #ifndef NDEBUG 697 /// This method is used when -debug is specified to verify that cache arrays 698 /// are properly kept sorted. 699 static void AssertSorted(MemoryDependenceResults::NonLocalDepInfo &Cache, 700 int Count = -1) { 701 if (Count == -1) 702 Count = Cache.size(); 703 assert(std::is_sorted(Cache.begin(), Cache.begin() + Count) && 704 "Cache isn't sorted!"); 705 } 706 #endif 707 708 const MemoryDependenceResults::NonLocalDepInfo & 709 MemoryDependenceResults::getNonLocalCallDependency(CallBase *QueryCall) { 710 assert(getDependency(QueryCall).isNonLocal() && 711 "getNonLocalCallDependency should only be used on calls with " 712 "non-local deps!"); 713 PerInstNLInfo &CacheP = NonLocalDepsMap[QueryCall]; 714 NonLocalDepInfo &Cache = CacheP.first; 715 716 // This is the set of blocks that need to be recomputed. In the cached case, 717 // this can happen due to instructions being deleted etc. In the uncached 718 // case, this starts out as the set of predecessors we care about. 719 SmallVector<BasicBlock *, 32> DirtyBlocks; 720 721 if (!Cache.empty()) { 722 // Okay, we have a cache entry. If we know it is not dirty, just return it 723 // with no computation. 724 if (!CacheP.second) { 725 ++NumCacheNonLocal; 726 return Cache; 727 } 728 729 // If we already have a partially computed set of results, scan them to 730 // determine what is dirty, seeding our initial DirtyBlocks worklist. 731 for (auto &Entry : Cache) 732 if (Entry.getResult().isDirty()) 733 DirtyBlocks.push_back(Entry.getBB()); 734 735 // Sort the cache so that we can do fast binary search lookups below. 736 llvm::sort(Cache); 737 738 ++NumCacheDirtyNonLocal; 739 } else { 740 // Seed DirtyBlocks with each of the preds of QueryInst's block. 741 BasicBlock *QueryBB = QueryCall->getParent(); 742 append_range(DirtyBlocks, PredCache.get(QueryBB)); 743 ++NumUncacheNonLocal; 744 } 745 746 // isReadonlyCall - If this is a read-only call, we can be more aggressive. 747 bool isReadonlyCall = AA.onlyReadsMemory(QueryCall); 748 749 SmallPtrSet<BasicBlock *, 32> Visited; 750 751 unsigned NumSortedEntries = Cache.size(); 752 LLVM_DEBUG(AssertSorted(Cache)); 753 754 // Iterate while we still have blocks to update. 755 while (!DirtyBlocks.empty()) { 756 BasicBlock *DirtyBB = DirtyBlocks.pop_back_val(); 757 758 // Already processed this block? 759 if (!Visited.insert(DirtyBB).second) 760 continue; 761 762 // Do a binary search to see if we already have an entry for this block in 763 // the cache set. If so, find it. 764 LLVM_DEBUG(AssertSorted(Cache, NumSortedEntries)); 765 NonLocalDepInfo::iterator Entry = 766 std::upper_bound(Cache.begin(), Cache.begin() + NumSortedEntries, 767 NonLocalDepEntry(DirtyBB)); 768 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB) 769 --Entry; 770 771 NonLocalDepEntry *ExistingResult = nullptr; 772 if (Entry != Cache.begin() + NumSortedEntries && 773 Entry->getBB() == DirtyBB) { 774 // If we already have an entry, and if it isn't already dirty, the block 775 // is done. 776 if (!Entry->getResult().isDirty()) 777 continue; 778 779 // Otherwise, remember this slot so we can update the value. 780 ExistingResult = &*Entry; 781 } 782 783 // If the dirty entry has a pointer, start scanning from it so we don't have 784 // to rescan the entire block. 785 BasicBlock::iterator ScanPos = DirtyBB->end(); 786 if (ExistingResult) { 787 if (Instruction *Inst = ExistingResult->getResult().getInst()) { 788 ScanPos = Inst->getIterator(); 789 // We're removing QueryInst's use of Inst. 790 RemoveFromReverseMap<Instruction *>(ReverseNonLocalDeps, Inst, 791 QueryCall); 792 } 793 } 794 795 // Find out if this block has a local dependency for QueryInst. 796 MemDepResult Dep; 797 798 if (ScanPos != DirtyBB->begin()) { 799 Dep = getCallDependencyFrom(QueryCall, isReadonlyCall, ScanPos, DirtyBB); 800 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { 801 // No dependence found. If this is the entry block of the function, it is 802 // a clobber, otherwise it is unknown. 803 Dep = MemDepResult::getNonLocal(); 804 } else { 805 Dep = MemDepResult::getNonFuncLocal(); 806 } 807 808 // If we had a dirty entry for the block, update it. Otherwise, just add 809 // a new entry. 810 if (ExistingResult) 811 ExistingResult->setResult(Dep); 812 else 813 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep)); 814 815 // If the block has a dependency (i.e. it isn't completely transparent to 816 // the value), remember the association! 817 if (!Dep.isNonLocal()) { 818 // Keep the ReverseNonLocalDeps map up to date so we can efficiently 819 // update this when we remove instructions. 820 if (Instruction *Inst = Dep.getInst()) 821 ReverseNonLocalDeps[Inst].insert(QueryCall); 822 } else { 823 824 // If the block *is* completely transparent to the load, we need to check 825 // the predecessors of this block. Add them to our worklist. 826 append_range(DirtyBlocks, PredCache.get(DirtyBB)); 827 } 828 } 829 830 return Cache; 831 } 832 833 void MemoryDependenceResults::getNonLocalPointerDependency( 834 Instruction *QueryInst, SmallVectorImpl<NonLocalDepResult> &Result) { 835 const MemoryLocation Loc = MemoryLocation::get(QueryInst); 836 bool isLoad = isa<LoadInst>(QueryInst); 837 BasicBlock *FromBB = QueryInst->getParent(); 838 assert(FromBB); 839 840 assert(Loc.Ptr->getType()->isPointerTy() && 841 "Can't get pointer deps of a non-pointer!"); 842 Result.clear(); 843 { 844 // Check if there is cached Def with invariant.group. 845 auto NonLocalDefIt = NonLocalDefsCache.find(QueryInst); 846 if (NonLocalDefIt != NonLocalDefsCache.end()) { 847 Result.push_back(NonLocalDefIt->second); 848 ReverseNonLocalDefsCache[NonLocalDefIt->second.getResult().getInst()] 849 .erase(QueryInst); 850 NonLocalDefsCache.erase(NonLocalDefIt); 851 return; 852 } 853 } 854 // This routine does not expect to deal with volatile instructions. 855 // Doing so would require piping through the QueryInst all the way through. 856 // TODO: volatiles can't be elided, but they can be reordered with other 857 // non-volatile accesses. 858 859 // We currently give up on any instruction which is ordered, but we do handle 860 // atomic instructions which are unordered. 861 // TODO: Handle ordered instructions 862 auto isOrdered = [](Instruction *Inst) { 863 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 864 return !LI->isUnordered(); 865 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 866 return !SI->isUnordered(); 867 } 868 return false; 869 }; 870 if (QueryInst->isVolatile() || isOrdered(QueryInst)) { 871 Result.push_back(NonLocalDepResult(FromBB, MemDepResult::getUnknown(), 872 const_cast<Value *>(Loc.Ptr))); 873 return; 874 } 875 const DataLayout &DL = FromBB->getModule()->getDataLayout(); 876 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, &AC); 877 878 // This is the set of blocks we've inspected, and the pointer we consider in 879 // each block. Because of critical edges, we currently bail out if querying 880 // a block with multiple different pointers. This can happen during PHI 881 // translation. 882 DenseMap<BasicBlock *, Value *> Visited; 883 if (getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB, 884 Result, Visited, true)) 885 return; 886 Result.clear(); 887 Result.push_back(NonLocalDepResult(FromBB, MemDepResult::getUnknown(), 888 const_cast<Value *>(Loc.Ptr))); 889 } 890 891 /// Compute the memdep value for BB with Pointer/PointeeSize using either 892 /// cached information in Cache or by doing a lookup (which may use dirty cache 893 /// info if available). 894 /// 895 /// If we do a lookup, add the result to the cache. 896 MemDepResult MemoryDependenceResults::getNonLocalInfoForBlock( 897 Instruction *QueryInst, const MemoryLocation &Loc, bool isLoad, 898 BasicBlock *BB, NonLocalDepInfo *Cache, unsigned NumSortedEntries, 899 BatchAAResults &BatchAA) { 900 901 bool isInvariantLoad = false; 902 903 if (LoadInst *LI = dyn_cast_or_null<LoadInst>(QueryInst)) 904 isInvariantLoad = LI->getMetadata(LLVMContext::MD_invariant_load); 905 906 // Do a binary search to see if we already have an entry for this block in 907 // the cache set. If so, find it. 908 NonLocalDepInfo::iterator Entry = std::upper_bound( 909 Cache->begin(), Cache->begin() + NumSortedEntries, NonLocalDepEntry(BB)); 910 if (Entry != Cache->begin() && (Entry - 1)->getBB() == BB) 911 --Entry; 912 913 NonLocalDepEntry *ExistingResult = nullptr; 914 if (Entry != Cache->begin() + NumSortedEntries && Entry->getBB() == BB) 915 ExistingResult = &*Entry; 916 917 // Use cached result for invariant load only if there is no dependency for non 918 // invariant load. In this case invariant load can not have any dependency as 919 // well. 920 if (ExistingResult && isInvariantLoad && 921 !ExistingResult->getResult().isNonFuncLocal()) 922 ExistingResult = nullptr; 923 924 // If we have a cached entry, and it is non-dirty, use it as the value for 925 // this dependency. 926 if (ExistingResult && !ExistingResult->getResult().isDirty()) { 927 ++NumCacheNonLocalPtr; 928 return ExistingResult->getResult(); 929 } 930 931 // Otherwise, we have to scan for the value. If we have a dirty cache 932 // entry, start scanning from its position, otherwise we scan from the end 933 // of the block. 934 BasicBlock::iterator ScanPos = BB->end(); 935 if (ExistingResult && ExistingResult->getResult().getInst()) { 936 assert(ExistingResult->getResult().getInst()->getParent() == BB && 937 "Instruction invalidated?"); 938 ++NumCacheDirtyNonLocalPtr; 939 ScanPos = ExistingResult->getResult().getInst()->getIterator(); 940 941 // Eliminating the dirty entry from 'Cache', so update the reverse info. 942 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 943 RemoveFromReverseMap(ReverseNonLocalPtrDeps, &*ScanPos, CacheKey); 944 } else { 945 ++NumUncacheNonLocalPtr; 946 } 947 948 // Scan the block for the dependency. 949 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB, 950 QueryInst, nullptr, BatchAA); 951 952 // Don't cache results for invariant load. 953 if (isInvariantLoad) 954 return Dep; 955 956 // If we had a dirty entry for the block, update it. Otherwise, just add 957 // a new entry. 958 if (ExistingResult) 959 ExistingResult->setResult(Dep); 960 else 961 Cache->push_back(NonLocalDepEntry(BB, Dep)); 962 963 // If the block has a dependency (i.e. it isn't completely transparent to 964 // the value), remember the reverse association because we just added it 965 // to Cache! 966 if (!Dep.isDef() && !Dep.isClobber()) 967 return Dep; 968 969 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently 970 // update MemDep when we remove instructions. 971 Instruction *Inst = Dep.getInst(); 972 assert(Inst && "Didn't depend on anything?"); 973 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); 974 ReverseNonLocalPtrDeps[Inst].insert(CacheKey); 975 return Dep; 976 } 977 978 /// Sort the NonLocalDepInfo cache, given a certain number of elements in the 979 /// array that are already properly ordered. 980 /// 981 /// This is optimized for the case when only a few entries are added. 982 static void 983 SortNonLocalDepInfoCache(MemoryDependenceResults::NonLocalDepInfo &Cache, 984 unsigned NumSortedEntries) { 985 switch (Cache.size() - NumSortedEntries) { 986 case 0: 987 // done, no new entries. 988 break; 989 case 2: { 990 // Two new entries, insert the last one into place. 991 NonLocalDepEntry Val = Cache.back(); 992 Cache.pop_back(); 993 MemoryDependenceResults::NonLocalDepInfo::iterator Entry = 994 std::upper_bound(Cache.begin(), Cache.end() - 1, Val); 995 Cache.insert(Entry, Val); 996 LLVM_FALLTHROUGH; 997 } 998 case 1: 999 // One new entry, Just insert the new value at the appropriate position. 1000 if (Cache.size() != 1) { 1001 NonLocalDepEntry Val = Cache.back(); 1002 Cache.pop_back(); 1003 MemoryDependenceResults::NonLocalDepInfo::iterator Entry = 1004 llvm::upper_bound(Cache, Val); 1005 Cache.insert(Entry, Val); 1006 } 1007 break; 1008 default: 1009 // Added many values, do a full scale sort. 1010 llvm::sort(Cache); 1011 break; 1012 } 1013 } 1014 1015 /// Perform a dependency query based on pointer/pointeesize starting at the end 1016 /// of StartBB. 1017 /// 1018 /// Add any clobber/def results to the results vector and keep track of which 1019 /// blocks are visited in 'Visited'. 1020 /// 1021 /// This has special behavior for the first block queries (when SkipFirstBlock 1022 /// is true). In this special case, it ignores the contents of the specified 1023 /// block and starts returning dependence info for its predecessors. 1024 /// 1025 /// This function returns true on success, or false to indicate that it could 1026 /// not compute dependence information for some reason. This should be treated 1027 /// as a clobber dependence on the first instruction in the predecessor block. 1028 bool MemoryDependenceResults::getNonLocalPointerDepFromBB( 1029 Instruction *QueryInst, const PHITransAddr &Pointer, 1030 const MemoryLocation &Loc, bool isLoad, BasicBlock *StartBB, 1031 SmallVectorImpl<NonLocalDepResult> &Result, 1032 DenseMap<BasicBlock *, Value *> &Visited, bool SkipFirstBlock, 1033 bool IsIncomplete) { 1034 // Look up the cached info for Pointer. 1035 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); 1036 1037 // Set up a temporary NLPI value. If the map doesn't yet have an entry for 1038 // CacheKey, this value will be inserted as the associated value. Otherwise, 1039 // it'll be ignored, and we'll have to check to see if the cached size and 1040 // aa tags are consistent with the current query. 1041 NonLocalPointerInfo InitialNLPI; 1042 InitialNLPI.Size = Loc.Size; 1043 InitialNLPI.AATags = Loc.AATags; 1044 1045 bool isInvariantLoad = false; 1046 if (LoadInst *LI = dyn_cast_or_null<LoadInst>(QueryInst)) 1047 isInvariantLoad = LI->getMetadata(LLVMContext::MD_invariant_load); 1048 1049 // Get the NLPI for CacheKey, inserting one into the map if it doesn't 1050 // already have one. 1051 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair = 1052 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI)); 1053 NonLocalPointerInfo *CacheInfo = &Pair.first->second; 1054 1055 // If we already have a cache entry for this CacheKey, we may need to do some 1056 // work to reconcile the cache entry and the current query. 1057 // Invariant loads don't participate in caching. Thus no need to reconcile. 1058 if (!isInvariantLoad && !Pair.second) { 1059 if (CacheInfo->Size != Loc.Size) { 1060 bool ThrowOutEverything; 1061 if (CacheInfo->Size.hasValue() && Loc.Size.hasValue()) { 1062 // FIXME: We may be able to do better in the face of results with mixed 1063 // precision. We don't appear to get them in practice, though, so just 1064 // be conservative. 1065 ThrowOutEverything = 1066 CacheInfo->Size.isPrecise() != Loc.Size.isPrecise() || 1067 CacheInfo->Size.getValue() < Loc.Size.getValue(); 1068 } else { 1069 // For our purposes, unknown size > all others. 1070 ThrowOutEverything = !Loc.Size.hasValue(); 1071 } 1072 1073 if (ThrowOutEverything) { 1074 // The query's Size is greater than the cached one. Throw out the 1075 // cached data and proceed with the query at the greater size. 1076 CacheInfo->Pair = BBSkipFirstBlockPair(); 1077 CacheInfo->Size = Loc.Size; 1078 for (auto &Entry : CacheInfo->NonLocalDeps) 1079 if (Instruction *Inst = Entry.getResult().getInst()) 1080 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 1081 CacheInfo->NonLocalDeps.clear(); 1082 // The cache is cleared (in the above line) so we will have lost 1083 // information about blocks we have already visited. We therefore must 1084 // assume that the cache information is incomplete. 1085 IsIncomplete = true; 1086 } else { 1087 // This query's Size is less than the cached one. Conservatively restart 1088 // the query using the greater size. 1089 return getNonLocalPointerDepFromBB( 1090 QueryInst, Pointer, Loc.getWithNewSize(CacheInfo->Size), isLoad, 1091 StartBB, Result, Visited, SkipFirstBlock, IsIncomplete); 1092 } 1093 } 1094 1095 // If the query's AATags are inconsistent with the cached one, 1096 // conservatively throw out the cached data and restart the query with 1097 // no tag if needed. 1098 if (CacheInfo->AATags != Loc.AATags) { 1099 if (CacheInfo->AATags) { 1100 CacheInfo->Pair = BBSkipFirstBlockPair(); 1101 CacheInfo->AATags = AAMDNodes(); 1102 for (auto &Entry : CacheInfo->NonLocalDeps) 1103 if (Instruction *Inst = Entry.getResult().getInst()) 1104 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); 1105 CacheInfo->NonLocalDeps.clear(); 1106 // The cache is cleared (in the above line) so we will have lost 1107 // information about blocks we have already visited. We therefore must 1108 // assume that the cache information is incomplete. 1109 IsIncomplete = true; 1110 } 1111 if (Loc.AATags) 1112 return getNonLocalPointerDepFromBB( 1113 QueryInst, Pointer, Loc.getWithoutAATags(), isLoad, StartBB, Result, 1114 Visited, SkipFirstBlock, IsIncomplete); 1115 } 1116 } 1117 1118 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps; 1119 1120 // If we have valid cached information for exactly the block we are 1121 // investigating, just return it with no recomputation. 1122 // Don't use cached information for invariant loads since it is valid for 1123 // non-invariant loads only. 1124 if (!IsIncomplete && !isInvariantLoad && 1125 CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { 1126 // We have a fully cached result for this query then we can just return the 1127 // cached results and populate the visited set. However, we have to verify 1128 // that we don't already have conflicting results for these blocks. Check 1129 // to ensure that if a block in the results set is in the visited set that 1130 // it was for the same pointer query. 1131 if (!Visited.empty()) { 1132 for (auto &Entry : *Cache) { 1133 DenseMap<BasicBlock *, Value *>::iterator VI = 1134 Visited.find(Entry.getBB()); 1135 if (VI == Visited.end() || VI->second == Pointer.getAddr()) 1136 continue; 1137 1138 // We have a pointer mismatch in a block. Just return false, saying 1139 // that something was clobbered in this result. We could also do a 1140 // non-fully cached query, but there is little point in doing this. 1141 return false; 1142 } 1143 } 1144 1145 Value *Addr = Pointer.getAddr(); 1146 for (auto &Entry : *Cache) { 1147 Visited.insert(std::make_pair(Entry.getBB(), Addr)); 1148 if (Entry.getResult().isNonLocal()) { 1149 continue; 1150 } 1151 1152 if (DT.isReachableFromEntry(Entry.getBB())) { 1153 Result.push_back( 1154 NonLocalDepResult(Entry.getBB(), Entry.getResult(), Addr)); 1155 } 1156 } 1157 ++NumCacheCompleteNonLocalPtr; 1158 return true; 1159 } 1160 1161 // Otherwise, either this is a new block, a block with an invalid cache 1162 // pointer or one that we're about to invalidate by putting more info into 1163 // it than its valid cache info. If empty and not explicitly indicated as 1164 // incomplete, the result will be valid cache info, otherwise it isn't. 1165 // 1166 // Invariant loads don't affect cache in any way thus no need to update 1167 // CacheInfo as well. 1168 if (!isInvariantLoad) { 1169 if (!IsIncomplete && Cache->empty()) 1170 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); 1171 else 1172 CacheInfo->Pair = BBSkipFirstBlockPair(); 1173 } 1174 1175 SmallVector<BasicBlock *, 32> Worklist; 1176 Worklist.push_back(StartBB); 1177 1178 // PredList used inside loop. 1179 SmallVector<std::pair<BasicBlock *, PHITransAddr>, 16> PredList; 1180 1181 // Keep track of the entries that we know are sorted. Previously cached 1182 // entries will all be sorted. The entries we add we only sort on demand (we 1183 // don't insert every element into its sorted position). We know that we 1184 // won't get any reuse from currently inserted values, because we don't 1185 // revisit blocks after we insert info for them. 1186 unsigned NumSortedEntries = Cache->size(); 1187 unsigned WorklistEntries = BlockNumberLimit; 1188 bool GotWorklistLimit = false; 1189 LLVM_DEBUG(AssertSorted(*Cache)); 1190 1191 BatchAAResults BatchAA(AA); 1192 while (!Worklist.empty()) { 1193 BasicBlock *BB = Worklist.pop_back_val(); 1194 1195 // If we do process a large number of blocks it becomes very expensive and 1196 // likely it isn't worth worrying about 1197 if (Result.size() > NumResultsLimit) { 1198 // Sort it now (if needed) so that recursive invocations of 1199 // getNonLocalPointerDepFromBB and other routines that could reuse the 1200 // cache value will only see properly sorted cache arrays. 1201 if (Cache && NumSortedEntries != Cache->size()) { 1202 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1203 } 1204 // Since we bail out, the "Cache" set won't contain all of the 1205 // results for the query. This is ok (we can still use it to accelerate 1206 // specific block queries) but we can't do the fastpath "return all 1207 // results from the set". Clear out the indicator for this. 1208 CacheInfo->Pair = BBSkipFirstBlockPair(); 1209 return false; 1210 } 1211 1212 // Skip the first block if we have it. 1213 if (!SkipFirstBlock) { 1214 // Analyze the dependency of *Pointer in FromBB. See if we already have 1215 // been here. 1216 assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); 1217 1218 // Get the dependency info for Pointer in BB. If we have cached 1219 // information, we will use it, otherwise we compute it. 1220 LLVM_DEBUG(AssertSorted(*Cache, NumSortedEntries)); 1221 MemDepResult Dep = getNonLocalInfoForBlock( 1222 QueryInst, Loc, isLoad, BB, Cache, NumSortedEntries, BatchAA); 1223 1224 // If we got a Def or Clobber, add this to the list of results. 1225 if (!Dep.isNonLocal()) { 1226 if (DT.isReachableFromEntry(BB)) { 1227 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); 1228 continue; 1229 } 1230 } 1231 } 1232 1233 // If 'Pointer' is an instruction defined in this block, then we need to do 1234 // phi translation to change it into a value live in the predecessor block. 1235 // If not, we just add the predecessors to the worklist and scan them with 1236 // the same Pointer. 1237 if (!Pointer.NeedsPHITranslationFromBlock(BB)) { 1238 SkipFirstBlock = false; 1239 SmallVector<BasicBlock *, 16> NewBlocks; 1240 for (BasicBlock *Pred : PredCache.get(BB)) { 1241 // Verify that we haven't looked at this block yet. 1242 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> InsertRes = 1243 Visited.insert(std::make_pair(Pred, Pointer.getAddr())); 1244 if (InsertRes.second) { 1245 // First time we've looked at *PI. 1246 NewBlocks.push_back(Pred); 1247 continue; 1248 } 1249 1250 // If we have seen this block before, but it was with a different 1251 // pointer then we have a phi translation failure and we have to treat 1252 // this as a clobber. 1253 if (InsertRes.first->second != Pointer.getAddr()) { 1254 // Make sure to clean up the Visited map before continuing on to 1255 // PredTranslationFailure. 1256 for (unsigned i = 0; i < NewBlocks.size(); i++) 1257 Visited.erase(NewBlocks[i]); 1258 goto PredTranslationFailure; 1259 } 1260 } 1261 if (NewBlocks.size() > WorklistEntries) { 1262 // Make sure to clean up the Visited map before continuing on to 1263 // PredTranslationFailure. 1264 for (unsigned i = 0; i < NewBlocks.size(); i++) 1265 Visited.erase(NewBlocks[i]); 1266 GotWorklistLimit = true; 1267 goto PredTranslationFailure; 1268 } 1269 WorklistEntries -= NewBlocks.size(); 1270 Worklist.append(NewBlocks.begin(), NewBlocks.end()); 1271 continue; 1272 } 1273 1274 // We do need to do phi translation, if we know ahead of time we can't phi 1275 // translate this value, don't even try. 1276 if (!Pointer.IsPotentiallyPHITranslatable()) 1277 goto PredTranslationFailure; 1278 1279 // We may have added values to the cache list before this PHI translation. 1280 // If so, we haven't done anything to ensure that the cache remains sorted. 1281 // Sort it now (if needed) so that recursive invocations of 1282 // getNonLocalPointerDepFromBB and other routines that could reuse the cache 1283 // value will only see properly sorted cache arrays. 1284 if (Cache && NumSortedEntries != Cache->size()) { 1285 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1286 NumSortedEntries = Cache->size(); 1287 } 1288 Cache = nullptr; 1289 1290 PredList.clear(); 1291 for (BasicBlock *Pred : PredCache.get(BB)) { 1292 PredList.push_back(std::make_pair(Pred, Pointer)); 1293 1294 // Get the PHI translated pointer in this predecessor. This can fail if 1295 // not translatable, in which case the getAddr() returns null. 1296 PHITransAddr &PredPointer = PredList.back().second; 1297 PredPointer.PHITranslateValue(BB, Pred, &DT, /*MustDominate=*/false); 1298 Value *PredPtrVal = PredPointer.getAddr(); 1299 1300 // Check to see if we have already visited this pred block with another 1301 // pointer. If so, we can't do this lookup. This failure can occur 1302 // with PHI translation when a critical edge exists and the PHI node in 1303 // the successor translates to a pointer value different than the 1304 // pointer the block was first analyzed with. 1305 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> InsertRes = 1306 Visited.insert(std::make_pair(Pred, PredPtrVal)); 1307 1308 if (!InsertRes.second) { 1309 // We found the pred; take it off the list of preds to visit. 1310 PredList.pop_back(); 1311 1312 // If the predecessor was visited with PredPtr, then we already did 1313 // the analysis and can ignore it. 1314 if (InsertRes.first->second == PredPtrVal) 1315 continue; 1316 1317 // Otherwise, the block was previously analyzed with a different 1318 // pointer. We can't represent the result of this case, so we just 1319 // treat this as a phi translation failure. 1320 1321 // Make sure to clean up the Visited map before continuing on to 1322 // PredTranslationFailure. 1323 for (unsigned i = 0, n = PredList.size(); i < n; ++i) 1324 Visited.erase(PredList[i].first); 1325 1326 goto PredTranslationFailure; 1327 } 1328 } 1329 1330 // Actually process results here; this need to be a separate loop to avoid 1331 // calling getNonLocalPointerDepFromBB for blocks we don't want to return 1332 // any results for. (getNonLocalPointerDepFromBB will modify our 1333 // datastructures in ways the code after the PredTranslationFailure label 1334 // doesn't expect.) 1335 for (unsigned i = 0, n = PredList.size(); i < n; ++i) { 1336 BasicBlock *Pred = PredList[i].first; 1337 PHITransAddr &PredPointer = PredList[i].second; 1338 Value *PredPtrVal = PredPointer.getAddr(); 1339 1340 bool CanTranslate = true; 1341 // If PHI translation was unable to find an available pointer in this 1342 // predecessor, then we have to assume that the pointer is clobbered in 1343 // that predecessor. We can still do PRE of the load, which would insert 1344 // a computation of the pointer in this predecessor. 1345 if (!PredPtrVal) 1346 CanTranslate = false; 1347 1348 // FIXME: it is entirely possible that PHI translating will end up with 1349 // the same value. Consider PHI translating something like: 1350 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* 1351 // to recurse here, pedantically speaking. 1352 1353 // If getNonLocalPointerDepFromBB fails here, that means the cached 1354 // result conflicted with the Visited list; we have to conservatively 1355 // assume it is unknown, but this also does not block PRE of the load. 1356 if (!CanTranslate || 1357 !getNonLocalPointerDepFromBB(QueryInst, PredPointer, 1358 Loc.getWithNewPtr(PredPtrVal), isLoad, 1359 Pred, Result, Visited)) { 1360 // Add the entry to the Result list. 1361 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal); 1362 Result.push_back(Entry); 1363 1364 // Since we had a phi translation failure, the cache for CacheKey won't 1365 // include all of the entries that we need to immediately satisfy future 1366 // queries. Mark this in NonLocalPointerDeps by setting the 1367 // BBSkipFirstBlockPair pointer to null. This requires reuse of the 1368 // cached value to do more work but not miss the phi trans failure. 1369 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey]; 1370 NLPI.Pair = BBSkipFirstBlockPair(); 1371 continue; 1372 } 1373 } 1374 1375 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. 1376 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1377 Cache = &CacheInfo->NonLocalDeps; 1378 NumSortedEntries = Cache->size(); 1379 1380 // Since we did phi translation, the "Cache" set won't contain all of the 1381 // results for the query. This is ok (we can still use it to accelerate 1382 // specific block queries) but we can't do the fastpath "return all 1383 // results from the set" Clear out the indicator for this. 1384 CacheInfo->Pair = BBSkipFirstBlockPair(); 1385 SkipFirstBlock = false; 1386 continue; 1387 1388 PredTranslationFailure: 1389 // The following code is "failure"; we can't produce a sane translation 1390 // for the given block. It assumes that we haven't modified any of 1391 // our datastructures while processing the current block. 1392 1393 if (!Cache) { 1394 // Refresh the CacheInfo/Cache pointer if it got invalidated. 1395 CacheInfo = &NonLocalPointerDeps[CacheKey]; 1396 Cache = &CacheInfo->NonLocalDeps; 1397 NumSortedEntries = Cache->size(); 1398 } 1399 1400 // Since we failed phi translation, the "Cache" set won't contain all of the 1401 // results for the query. This is ok (we can still use it to accelerate 1402 // specific block queries) but we can't do the fastpath "return all 1403 // results from the set". Clear out the indicator for this. 1404 CacheInfo->Pair = BBSkipFirstBlockPair(); 1405 1406 // If *nothing* works, mark the pointer as unknown. 1407 // 1408 // If this is the magic first block, return this as a clobber of the whole 1409 // incoming value. Since we can't phi translate to one of the predecessors, 1410 // we have to bail out. 1411 if (SkipFirstBlock) 1412 return false; 1413 1414 // Results of invariant loads are not cached thus no need to update cached 1415 // information. 1416 if (!isInvariantLoad) { 1417 for (NonLocalDepEntry &I : llvm::reverse(*Cache)) { 1418 if (I.getBB() != BB) 1419 continue; 1420 1421 assert((GotWorklistLimit || I.getResult().isNonLocal() || 1422 !DT.isReachableFromEntry(BB)) && 1423 "Should only be here with transparent block"); 1424 1425 I.setResult(MemDepResult::getUnknown()); 1426 1427 1428 break; 1429 } 1430 } 1431 (void)GotWorklistLimit; 1432 // Go ahead and report unknown dependence. 1433 Result.push_back( 1434 NonLocalDepResult(BB, MemDepResult::getUnknown(), Pointer.getAddr())); 1435 } 1436 1437 // Okay, we're done now. If we added new values to the cache, re-sort it. 1438 SortNonLocalDepInfoCache(*Cache, NumSortedEntries); 1439 LLVM_DEBUG(AssertSorted(*Cache)); 1440 return true; 1441 } 1442 1443 /// If P exists in CachedNonLocalPointerInfo or NonLocalDefsCache, remove it. 1444 void MemoryDependenceResults::removeCachedNonLocalPointerDependencies( 1445 ValueIsLoadPair P) { 1446 1447 // Most of the time this cache is empty. 1448 if (!NonLocalDefsCache.empty()) { 1449 auto it = NonLocalDefsCache.find(P.getPointer()); 1450 if (it != NonLocalDefsCache.end()) { 1451 RemoveFromReverseMap(ReverseNonLocalDefsCache, 1452 it->second.getResult().getInst(), P.getPointer()); 1453 NonLocalDefsCache.erase(it); 1454 } 1455 1456 if (auto *I = dyn_cast<Instruction>(P.getPointer())) { 1457 auto toRemoveIt = ReverseNonLocalDefsCache.find(I); 1458 if (toRemoveIt != ReverseNonLocalDefsCache.end()) { 1459 for (const auto *entry : toRemoveIt->second) 1460 NonLocalDefsCache.erase(entry); 1461 ReverseNonLocalDefsCache.erase(toRemoveIt); 1462 } 1463 } 1464 } 1465 1466 CachedNonLocalPointerInfo::iterator It = NonLocalPointerDeps.find(P); 1467 if (It == NonLocalPointerDeps.end()) 1468 return; 1469 1470 // Remove all of the entries in the BB->val map. This involves removing 1471 // instructions from the reverse map. 1472 NonLocalDepInfo &PInfo = It->second.NonLocalDeps; 1473 1474 for (const NonLocalDepEntry &DE : PInfo) { 1475 Instruction *Target = DE.getResult().getInst(); 1476 if (!Target) 1477 continue; // Ignore non-local dep results. 1478 assert(Target->getParent() == DE.getBB()); 1479 1480 // Eliminating the dirty entry from 'Cache', so update the reverse info. 1481 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); 1482 } 1483 1484 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). 1485 NonLocalPointerDeps.erase(It); 1486 } 1487 1488 void MemoryDependenceResults::invalidateCachedPointerInfo(Value *Ptr) { 1489 // If Ptr isn't really a pointer, just ignore it. 1490 if (!Ptr->getType()->isPointerTy()) 1491 return; 1492 // Flush store info for the pointer. 1493 removeCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); 1494 // Flush load info for the pointer. 1495 removeCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); 1496 // Invalidate phis that use the pointer. 1497 PV.invalidateValue(Ptr); 1498 } 1499 1500 void MemoryDependenceResults::invalidateCachedPredecessors() { 1501 PredCache.clear(); 1502 } 1503 1504 void MemoryDependenceResults::removeInstruction(Instruction *RemInst) { 1505 // Walk through the Non-local dependencies, removing this one as the value 1506 // for any cached queries. 1507 NonLocalDepMapType::iterator NLDI = NonLocalDepsMap.find(RemInst); 1508 if (NLDI != NonLocalDepsMap.end()) { 1509 NonLocalDepInfo &BlockMap = NLDI->second.first; 1510 for (auto &Entry : BlockMap) 1511 if (Instruction *Inst = Entry.getResult().getInst()) 1512 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); 1513 NonLocalDepsMap.erase(NLDI); 1514 } 1515 1516 // If we have a cached local dependence query for this instruction, remove it. 1517 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); 1518 if (LocalDepEntry != LocalDeps.end()) { 1519 // Remove us from DepInst's reverse set now that the local dep info is gone. 1520 if (Instruction *Inst = LocalDepEntry->second.getInst()) 1521 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); 1522 1523 // Remove this local dependency info. 1524 LocalDeps.erase(LocalDepEntry); 1525 } 1526 1527 // If we have any cached dependencies on this instruction, remove 1528 // them. 1529 1530 // If the instruction is a pointer, remove it from both the load info and the 1531 // store info. 1532 if (RemInst->getType()->isPointerTy()) { 1533 removeCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); 1534 removeCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); 1535 } else { 1536 // Otherwise, if the instructions is in the map directly, it must be a load. 1537 // Remove it. 1538 auto toRemoveIt = NonLocalDefsCache.find(RemInst); 1539 if (toRemoveIt != NonLocalDefsCache.end()) { 1540 assert(isa<LoadInst>(RemInst) && 1541 "only load instructions should be added directly"); 1542 const Instruction *DepV = toRemoveIt->second.getResult().getInst(); 1543 ReverseNonLocalDefsCache.find(DepV)->second.erase(RemInst); 1544 NonLocalDefsCache.erase(toRemoveIt); 1545 } 1546 } 1547 1548 // Loop over all of the things that depend on the instruction we're removing. 1549 SmallVector<std::pair<Instruction *, Instruction *>, 8> ReverseDepsToAdd; 1550 1551 // If we find RemInst as a clobber or Def in any of the maps for other values, 1552 // we need to replace its entry with a dirty version of the instruction after 1553 // it. If RemInst is a terminator, we use a null dirty value. 1554 // 1555 // Using a dirty version of the instruction after RemInst saves having to scan 1556 // the entire block to get to this point. 1557 MemDepResult NewDirtyVal; 1558 if (!RemInst->isTerminator()) 1559 NewDirtyVal = MemDepResult::getDirty(&*++RemInst->getIterator()); 1560 1561 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); 1562 if (ReverseDepIt != ReverseLocalDeps.end()) { 1563 // RemInst can't be the terminator if it has local stuff depending on it. 1564 assert(!ReverseDepIt->second.empty() && !RemInst->isTerminator() && 1565 "Nothing can locally depend on a terminator"); 1566 1567 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) { 1568 assert(InstDependingOnRemInst != RemInst && 1569 "Already removed our local dep info"); 1570 1571 LocalDeps[InstDependingOnRemInst] = NewDirtyVal; 1572 1573 // Make sure to remember that new things depend on NewDepInst. 1574 assert(NewDirtyVal.getInst() && 1575 "There is no way something else can have " 1576 "a local dep on this if it is a terminator!"); 1577 ReverseDepsToAdd.push_back( 1578 std::make_pair(NewDirtyVal.getInst(), InstDependingOnRemInst)); 1579 } 1580 1581 ReverseLocalDeps.erase(ReverseDepIt); 1582 1583 // Add new reverse deps after scanning the set, to avoid invalidating the 1584 // 'ReverseDeps' reference. 1585 while (!ReverseDepsToAdd.empty()) { 1586 ReverseLocalDeps[ReverseDepsToAdd.back().first].insert( 1587 ReverseDepsToAdd.back().second); 1588 ReverseDepsToAdd.pop_back(); 1589 } 1590 } 1591 1592 ReverseDepIt = ReverseNonLocalDeps.find(RemInst); 1593 if (ReverseDepIt != ReverseNonLocalDeps.end()) { 1594 for (Instruction *I : ReverseDepIt->second) { 1595 assert(I != RemInst && "Already removed NonLocalDep info for RemInst"); 1596 1597 PerInstNLInfo &INLD = NonLocalDepsMap[I]; 1598 // The information is now dirty! 1599 INLD.second = true; 1600 1601 for (auto &Entry : INLD.first) { 1602 if (Entry.getResult().getInst() != RemInst) 1603 continue; 1604 1605 // Convert to a dirty entry for the subsequent instruction. 1606 Entry.setResult(NewDirtyVal); 1607 1608 if (Instruction *NextI = NewDirtyVal.getInst()) 1609 ReverseDepsToAdd.push_back(std::make_pair(NextI, I)); 1610 } 1611 } 1612 1613 ReverseNonLocalDeps.erase(ReverseDepIt); 1614 1615 // Add new reverse deps after scanning the set, to avoid invalidating 'Set' 1616 while (!ReverseDepsToAdd.empty()) { 1617 ReverseNonLocalDeps[ReverseDepsToAdd.back().first].insert( 1618 ReverseDepsToAdd.back().second); 1619 ReverseDepsToAdd.pop_back(); 1620 } 1621 } 1622 1623 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a 1624 // value in the NonLocalPointerDeps info. 1625 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = 1626 ReverseNonLocalPtrDeps.find(RemInst); 1627 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { 1628 SmallVector<std::pair<Instruction *, ValueIsLoadPair>, 8> 1629 ReversePtrDepsToAdd; 1630 1631 for (ValueIsLoadPair P : ReversePtrDepIt->second) { 1632 assert(P.getPointer() != RemInst && 1633 "Already removed NonLocalPointerDeps info for RemInst"); 1634 1635 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps; 1636 1637 // The cache is not valid for any specific block anymore. 1638 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair(); 1639 1640 // Update any entries for RemInst to use the instruction after it. 1641 for (auto &Entry : NLPDI) { 1642 if (Entry.getResult().getInst() != RemInst) 1643 continue; 1644 1645 // Convert to a dirty entry for the subsequent instruction. 1646 Entry.setResult(NewDirtyVal); 1647 1648 if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) 1649 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); 1650 } 1651 1652 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its 1653 // subsequent value may invalidate the sortedness. 1654 llvm::sort(NLPDI); 1655 } 1656 1657 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); 1658 1659 while (!ReversePtrDepsToAdd.empty()) { 1660 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first].insert( 1661 ReversePtrDepsToAdd.back().second); 1662 ReversePtrDepsToAdd.pop_back(); 1663 } 1664 } 1665 1666 // Invalidate phis that use the removed instruction. 1667 PV.invalidateValue(RemInst); 1668 1669 assert(!NonLocalDepsMap.count(RemInst) && "RemInst got reinserted?"); 1670 LLVM_DEBUG(verifyRemoved(RemInst)); 1671 } 1672 1673 /// Verify that the specified instruction does not occur in our internal data 1674 /// structures. 1675 /// 1676 /// This function verifies by asserting in debug builds. 1677 void MemoryDependenceResults::verifyRemoved(Instruction *D) const { 1678 #ifndef NDEBUG 1679 for (const auto &DepKV : LocalDeps) { 1680 assert(DepKV.first != D && "Inst occurs in data structures"); 1681 assert(DepKV.second.getInst() != D && "Inst occurs in data structures"); 1682 } 1683 1684 for (const auto &DepKV : NonLocalPointerDeps) { 1685 assert(DepKV.first.getPointer() != D && "Inst occurs in NLPD map key"); 1686 for (const auto &Entry : DepKV.second.NonLocalDeps) 1687 assert(Entry.getResult().getInst() != D && "Inst occurs as NLPD value"); 1688 } 1689 1690 for (const auto &DepKV : NonLocalDepsMap) { 1691 assert(DepKV.first != D && "Inst occurs in data structures"); 1692 const PerInstNLInfo &INLD = DepKV.second; 1693 for (const auto &Entry : INLD.first) 1694 assert(Entry.getResult().getInst() != D && 1695 "Inst occurs in data structures"); 1696 } 1697 1698 for (const auto &DepKV : ReverseLocalDeps) { 1699 assert(DepKV.first != D && "Inst occurs in data structures"); 1700 for (Instruction *Inst : DepKV.second) 1701 assert(Inst != D && "Inst occurs in data structures"); 1702 } 1703 1704 for (const auto &DepKV : ReverseNonLocalDeps) { 1705 assert(DepKV.first != D && "Inst occurs in data structures"); 1706 for (Instruction *Inst : DepKV.second) 1707 assert(Inst != D && "Inst occurs in data structures"); 1708 } 1709 1710 for (const auto &DepKV : ReverseNonLocalPtrDeps) { 1711 assert(DepKV.first != D && "Inst occurs in rev NLPD map"); 1712 1713 for (ValueIsLoadPair P : DepKV.second) 1714 assert(P != ValueIsLoadPair(D, false) && P != ValueIsLoadPair(D, true) && 1715 "Inst occurs in ReverseNonLocalPtrDeps map"); 1716 } 1717 #endif 1718 } 1719 1720 AnalysisKey MemoryDependenceAnalysis::Key; 1721 1722 MemoryDependenceAnalysis::MemoryDependenceAnalysis() 1723 : DefaultBlockScanLimit(BlockScanLimit) {} 1724 1725 MemoryDependenceResults 1726 MemoryDependenceAnalysis::run(Function &F, FunctionAnalysisManager &AM) { 1727 auto &AA = AM.getResult<AAManager>(F); 1728 auto &AC = AM.getResult<AssumptionAnalysis>(F); 1729 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1730 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 1731 auto &PV = AM.getResult<PhiValuesAnalysis>(F); 1732 return MemoryDependenceResults(AA, AC, TLI, DT, PV, DefaultBlockScanLimit); 1733 } 1734 1735 char MemoryDependenceWrapperPass::ID = 0; 1736 1737 INITIALIZE_PASS_BEGIN(MemoryDependenceWrapperPass, "memdep", 1738 "Memory Dependence Analysis", false, true) 1739 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 1740 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 1741 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 1742 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1743 INITIALIZE_PASS_DEPENDENCY(PhiValuesWrapperPass) 1744 INITIALIZE_PASS_END(MemoryDependenceWrapperPass, "memdep", 1745 "Memory Dependence Analysis", false, true) 1746 1747 MemoryDependenceWrapperPass::MemoryDependenceWrapperPass() : FunctionPass(ID) { 1748 initializeMemoryDependenceWrapperPassPass(*PassRegistry::getPassRegistry()); 1749 } 1750 1751 MemoryDependenceWrapperPass::~MemoryDependenceWrapperPass() = default; 1752 1753 void MemoryDependenceWrapperPass::releaseMemory() { 1754 MemDep.reset(); 1755 } 1756 1757 void MemoryDependenceWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 1758 AU.setPreservesAll(); 1759 AU.addRequired<AssumptionCacheTracker>(); 1760 AU.addRequired<DominatorTreeWrapperPass>(); 1761 AU.addRequired<PhiValuesWrapperPass>(); 1762 AU.addRequiredTransitive<AAResultsWrapperPass>(); 1763 AU.addRequiredTransitive<TargetLibraryInfoWrapperPass>(); 1764 } 1765 1766 bool MemoryDependenceResults::invalidate(Function &F, const PreservedAnalyses &PA, 1767 FunctionAnalysisManager::Invalidator &Inv) { 1768 // Check whether our analysis is preserved. 1769 auto PAC = PA.getChecker<MemoryDependenceAnalysis>(); 1770 if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Function>>()) 1771 // If not, give up now. 1772 return true; 1773 1774 // Check whether the analyses we depend on became invalid for any reason. 1775 if (Inv.invalidate<AAManager>(F, PA) || 1776 Inv.invalidate<AssumptionAnalysis>(F, PA) || 1777 Inv.invalidate<DominatorTreeAnalysis>(F, PA) || 1778 Inv.invalidate<PhiValuesAnalysis>(F, PA)) 1779 return true; 1780 1781 // Otherwise this analysis result remains valid. 1782 return false; 1783 } 1784 1785 unsigned MemoryDependenceResults::getDefaultBlockScanLimit() const { 1786 return DefaultBlockScanLimit; 1787 } 1788 1789 bool MemoryDependenceWrapperPass::runOnFunction(Function &F) { 1790 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); 1791 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 1792 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 1793 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1794 auto &PV = getAnalysis<PhiValuesWrapperPass>().getResult(); 1795 MemDep.emplace(AA, AC, TLI, DT, PV, BlockScanLimit); 1796 return false; 1797 } 1798