1 //===-- MemorySSA.cpp - Memory SSA Builder---------------------------===// 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 the MemorySSA class. 11 // 12 //===----------------------------------------------------------------===// 13 #include "llvm/Transforms/Utils/MemorySSA.h" 14 #include "llvm/ADT/DenseMap.h" 15 #include "llvm/ADT/DenseSet.h" 16 #include "llvm/ADT/DepthFirstIterator.h" 17 #include "llvm/ADT/GraphTraits.h" 18 #include "llvm/ADT/PostOrderIterator.h" 19 #include "llvm/ADT/STLExtras.h" 20 #include "llvm/ADT/SmallPtrSet.h" 21 #include "llvm/ADT/SmallSet.h" 22 #include "llvm/ADT/Statistic.h" 23 #include "llvm/Analysis/AliasAnalysis.h" 24 #include "llvm/Analysis/CFG.h" 25 #include "llvm/Analysis/GlobalsModRef.h" 26 #include "llvm/Analysis/IteratedDominanceFrontier.h" 27 #include "llvm/Analysis/MemoryLocation.h" 28 #include "llvm/Analysis/PHITransAddr.h" 29 #include "llvm/IR/AssemblyAnnotationWriter.h" 30 #include "llvm/IR/DataLayout.h" 31 #include "llvm/IR/Dominators.h" 32 #include "llvm/IR/GlobalVariable.h" 33 #include "llvm/IR/IRBuilder.h" 34 #include "llvm/IR/IntrinsicInst.h" 35 #include "llvm/IR/LLVMContext.h" 36 #include "llvm/IR/Metadata.h" 37 #include "llvm/IR/Module.h" 38 #include "llvm/IR/PatternMatch.h" 39 #include "llvm/Support/Debug.h" 40 #include "llvm/Support/FormattedStream.h" 41 #include "llvm/Transforms/Scalar.h" 42 #include <algorithm> 43 44 #define DEBUG_TYPE "memoryssa" 45 using namespace llvm; 46 STATISTIC(NumClobberCacheLookups, "Number of Memory SSA version cache lookups"); 47 STATISTIC(NumClobberCacheHits, "Number of Memory SSA version cache hits"); 48 STATISTIC(NumClobberCacheInserts, "Number of MemorySSA version cache inserts"); 49 50 INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false, 51 true) 52 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 53 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 54 INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false, 55 true) 56 57 INITIALIZE_PASS_BEGIN(MemorySSAPrinterLegacyPass, "print-memoryssa", 58 "Memory SSA Printer", false, false) 59 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) 60 INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa", 61 "Memory SSA Printer", false, false) 62 63 static cl::opt<bool> 64 VerifyMemorySSA("verify-memoryssa", cl::init(false), cl::Hidden, 65 cl::desc("Verify MemorySSA in legacy printer pass.")); 66 67 namespace llvm { 68 /// \brief An assembly annotator class to print Memory SSA information in 69 /// comments. 70 class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter { 71 friend class MemorySSA; 72 const MemorySSA *MSSA; 73 74 public: 75 MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {} 76 77 virtual void emitBasicBlockStartAnnot(const BasicBlock *BB, 78 formatted_raw_ostream &OS) { 79 if (MemoryAccess *MA = MSSA->getMemoryAccess(BB)) 80 OS << "; " << *MA << "\n"; 81 } 82 83 virtual void emitInstructionAnnot(const Instruction *I, 84 formatted_raw_ostream &OS) { 85 if (MemoryAccess *MA = MSSA->getMemoryAccess(I)) 86 OS << "; " << *MA << "\n"; 87 } 88 }; 89 90 /// \brief A MemorySSAWalker that does AA walks and caching of lookups to 91 /// disambiguate accesses. 92 /// 93 /// FIXME: The current implementation of this can take quadratic space in rare 94 /// cases. This can be fixed, but it is something to note until it is fixed. 95 /// 96 /// In order to trigger this behavior, you need to store to N distinct locations 97 /// (that AA can prove don't alias), perform M stores to other memory 98 /// locations that AA can prove don't alias any of the initial N locations, and 99 /// then load from all of the N locations. In this case, we insert M cache 100 /// entries for each of the N loads. 101 /// 102 /// For example: 103 /// define i32 @foo() { 104 /// %a = alloca i32, align 4 105 /// %b = alloca i32, align 4 106 /// store i32 0, i32* %a, align 4 107 /// store i32 0, i32* %b, align 4 108 /// 109 /// ; Insert M stores to other memory that doesn't alias %a or %b here 110 /// 111 /// %c = load i32, i32* %a, align 4 ; Caches M entries in 112 /// ; CachedUpwardsClobberingAccess for the 113 /// ; MemoryLocation %a 114 /// %d = load i32, i32* %b, align 4 ; Caches M entries in 115 /// ; CachedUpwardsClobberingAccess for the 116 /// ; MemoryLocation %b 117 /// 118 /// ; For completeness' sake, loading %a or %b again would not cache *another* 119 /// ; M entries. 120 /// %r = add i32 %c, %d 121 /// ret i32 %r 122 /// } 123 class MemorySSA::CachingWalker final : public MemorySSAWalker { 124 public: 125 CachingWalker(MemorySSA *, AliasAnalysis *, DominatorTree *); 126 ~CachingWalker() override; 127 128 MemoryAccess *getClobberingMemoryAccess(const Instruction *) override; 129 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, 130 MemoryLocation &) override; 131 void invalidateInfo(MemoryAccess *) override; 132 133 protected: 134 struct UpwardsMemoryQuery; 135 MemoryAccess *doCacheLookup(const MemoryAccess *, const UpwardsMemoryQuery &, 136 const MemoryLocation &); 137 138 void doCacheInsert(const MemoryAccess *, MemoryAccess *, 139 const UpwardsMemoryQuery &, const MemoryLocation &); 140 141 void doCacheRemove(const MemoryAccess *, const UpwardsMemoryQuery &, 142 const MemoryLocation &); 143 144 private: 145 MemoryAccessPair UpwardsDFSWalk(MemoryAccess *, const MemoryLocation &, 146 UpwardsMemoryQuery &, bool); 147 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *, UpwardsMemoryQuery &); 148 bool instructionClobbersQuery(const MemoryDef *, UpwardsMemoryQuery &, 149 const MemoryLocation &Loc) const; 150 void verifyRemoved(MemoryAccess *); 151 SmallDenseMap<ConstMemoryAccessPair, MemoryAccess *> 152 CachedUpwardsClobberingAccess; 153 DenseMap<const MemoryAccess *, MemoryAccess *> CachedUpwardsClobberingCall; 154 AliasAnalysis *AA; 155 DominatorTree *DT; 156 }; 157 } 158 159 namespace { 160 struct RenamePassData { 161 DomTreeNode *DTN; 162 DomTreeNode::const_iterator ChildIt; 163 MemoryAccess *IncomingVal; 164 165 RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It, 166 MemoryAccess *M) 167 : DTN(D), ChildIt(It), IncomingVal(M) {} 168 void swap(RenamePassData &RHS) { 169 std::swap(DTN, RHS.DTN); 170 std::swap(ChildIt, RHS.ChildIt); 171 std::swap(IncomingVal, RHS.IncomingVal); 172 } 173 }; 174 } 175 176 namespace llvm { 177 /// \brief Rename a single basic block into MemorySSA form. 178 /// Uses the standard SSA renaming algorithm. 179 /// \returns The new incoming value. 180 MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, 181 MemoryAccess *IncomingVal) { 182 auto It = PerBlockAccesses.find(BB); 183 // Skip most processing if the list is empty. 184 if (It != PerBlockAccesses.end()) { 185 AccessList *Accesses = It->second.get(); 186 for (MemoryAccess &L : *Accesses) { 187 switch (L.getValueID()) { 188 case Value::MemoryUseVal: 189 cast<MemoryUse>(&L)->setDefiningAccess(IncomingVal); 190 break; 191 case Value::MemoryDefVal: 192 // We can't legally optimize defs, because we only allow single 193 // memory phis/uses on operations, and if we optimize these, we can 194 // end up with multiple reaching defs. Uses do not have this 195 // problem, since they do not produce a value 196 cast<MemoryDef>(&L)->setDefiningAccess(IncomingVal); 197 IncomingVal = &L; 198 break; 199 case Value::MemoryPhiVal: 200 IncomingVal = &L; 201 break; 202 } 203 } 204 } 205 206 // Pass through values to our successors 207 for (const BasicBlock *S : successors(BB)) { 208 auto It = PerBlockAccesses.find(S); 209 // Rename the phi nodes in our successor block 210 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front())) 211 continue; 212 AccessList *Accesses = It->second.get(); 213 auto *Phi = cast<MemoryPhi>(&Accesses->front()); 214 Phi->addIncoming(IncomingVal, BB); 215 } 216 217 return IncomingVal; 218 } 219 220 /// \brief This is the standard SSA renaming algorithm. 221 /// 222 /// We walk the dominator tree in preorder, renaming accesses, and then filling 223 /// in phi nodes in our successors. 224 void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal, 225 SmallPtrSet<BasicBlock *, 16> &Visited) { 226 SmallVector<RenamePassData, 32> WorkStack; 227 IncomingVal = renameBlock(Root->getBlock(), IncomingVal); 228 WorkStack.push_back({Root, Root->begin(), IncomingVal}); 229 Visited.insert(Root->getBlock()); 230 231 while (!WorkStack.empty()) { 232 DomTreeNode *Node = WorkStack.back().DTN; 233 DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt; 234 IncomingVal = WorkStack.back().IncomingVal; 235 236 if (ChildIt == Node->end()) { 237 WorkStack.pop_back(); 238 } else { 239 DomTreeNode *Child = *ChildIt; 240 ++WorkStack.back().ChildIt; 241 BasicBlock *BB = Child->getBlock(); 242 Visited.insert(BB); 243 IncomingVal = renameBlock(BB, IncomingVal); 244 WorkStack.push_back({Child, Child->begin(), IncomingVal}); 245 } 246 } 247 } 248 249 /// \brief Compute dominator levels, used by the phi insertion algorithm above. 250 void MemorySSA::computeDomLevels(DenseMap<DomTreeNode *, unsigned> &DomLevels) { 251 for (auto DFI = df_begin(DT->getRootNode()), DFE = df_end(DT->getRootNode()); 252 DFI != DFE; ++DFI) 253 DomLevels[*DFI] = DFI.getPathLength() - 1; 254 } 255 256 /// \brief This handles unreachable block accesses by deleting phi nodes in 257 /// unreachable blocks, and marking all other unreachable MemoryAccess's as 258 /// being uses of the live on entry definition. 259 void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) { 260 assert(!DT->isReachableFromEntry(BB) && 261 "Reachable block found while handling unreachable blocks"); 262 263 // Make sure phi nodes in our reachable successors end up with a 264 // LiveOnEntryDef for our incoming edge, even though our block is forward 265 // unreachable. We could just disconnect these blocks from the CFG fully, 266 // but we do not right now. 267 for (const BasicBlock *S : successors(BB)) { 268 if (!DT->isReachableFromEntry(S)) 269 continue; 270 auto It = PerBlockAccesses.find(S); 271 // Rename the phi nodes in our successor block 272 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front())) 273 continue; 274 AccessList *Accesses = It->second.get(); 275 auto *Phi = cast<MemoryPhi>(&Accesses->front()); 276 Phi->addIncoming(LiveOnEntryDef.get(), BB); 277 } 278 279 auto It = PerBlockAccesses.find(BB); 280 if (It == PerBlockAccesses.end()) 281 return; 282 283 auto &Accesses = It->second; 284 for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) { 285 auto Next = std::next(AI); 286 // If we have a phi, just remove it. We are going to replace all 287 // users with live on entry. 288 if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI)) 289 UseOrDef->setDefiningAccess(LiveOnEntryDef.get()); 290 else 291 Accesses->erase(AI); 292 AI = Next; 293 } 294 } 295 296 MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT) 297 : AA(AA), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr), 298 NextID(0) { 299 buildMemorySSA(); 300 } 301 302 MemorySSA::MemorySSA(MemorySSA &&MSSA) 303 : AA(MSSA.AA), DT(MSSA.DT), F(MSSA.F), 304 ValueToMemoryAccess(std::move(MSSA.ValueToMemoryAccess)), 305 PerBlockAccesses(std::move(MSSA.PerBlockAccesses)), 306 LiveOnEntryDef(std::move(MSSA.LiveOnEntryDef)), 307 Walker(std::move(MSSA.Walker)), NextID(MSSA.NextID) { 308 // Update the Walker MSSA pointer so it doesn't point to the moved-from MSSA 309 // object any more. 310 Walker->MSSA = this; 311 } 312 313 MemorySSA::~MemorySSA() { 314 // Drop all our references 315 for (const auto &Pair : PerBlockAccesses) 316 for (MemoryAccess &MA : *Pair.second) 317 MA.dropAllReferences(); 318 } 319 320 MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) { 321 auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr)); 322 323 if (Res.second) 324 Res.first->second = make_unique<AccessList>(); 325 return Res.first->second.get(); 326 } 327 328 void MemorySSA::buildMemorySSA() { 329 // We create an access to represent "live on entry", for things like 330 // arguments or users of globals, where the memory they use is defined before 331 // the beginning of the function. We do not actually insert it into the IR. 332 // We do not define a live on exit for the immediate uses, and thus our 333 // semantics do *not* imply that something with no immediate uses can simply 334 // be removed. 335 BasicBlock &StartingPoint = F.getEntryBlock(); 336 LiveOnEntryDef = make_unique<MemoryDef>(F.getContext(), nullptr, nullptr, 337 &StartingPoint, NextID++); 338 339 // We maintain lists of memory accesses per-block, trading memory for time. We 340 // could just look up the memory access for every possible instruction in the 341 // stream. 342 SmallPtrSet<BasicBlock *, 32> DefiningBlocks; 343 SmallPtrSet<BasicBlock *, 32> DefUseBlocks; 344 // Go through each block, figure out where defs occur, and chain together all 345 // the accesses. 346 for (BasicBlock &B : F) { 347 bool InsertIntoDef = false; 348 AccessList *Accesses = nullptr; 349 for (Instruction &I : B) { 350 MemoryUseOrDef *MUD = createNewAccess(&I); 351 if (!MUD) 352 continue; 353 InsertIntoDef |= isa<MemoryDef>(MUD); 354 355 if (!Accesses) 356 Accesses = getOrCreateAccessList(&B); 357 Accesses->push_back(MUD); 358 } 359 if (InsertIntoDef) 360 DefiningBlocks.insert(&B); 361 if (Accesses) 362 DefUseBlocks.insert(&B); 363 } 364 365 // Compute live-in. 366 // Live in is normally defined as "all the blocks on the path from each def to 367 // each of it's uses". 368 // MemoryDef's are implicit uses of previous state, so they are also uses. 369 // This means we don't really have def-only instructions. The only 370 // MemoryDef's that are not really uses are those that are of the LiveOnEntry 371 // variable (because LiveOnEntry can reach anywhere, and every def is a 372 // must-kill of LiveOnEntry). 373 // In theory, you could precisely compute live-in by using alias-analysis to 374 // disambiguate defs and uses to see which really pair up with which. 375 // In practice, this would be really expensive and difficult. So we simply 376 // assume all defs are also uses that need to be kept live. 377 // Because of this, the end result of this live-in computation will be "the 378 // entire set of basic blocks that reach any use". 379 380 SmallPtrSet<BasicBlock *, 32> LiveInBlocks; 381 SmallVector<BasicBlock *, 64> LiveInBlockWorklist(DefUseBlocks.begin(), 382 DefUseBlocks.end()); 383 // Now that we have a set of blocks where a value is live-in, recursively add 384 // predecessors until we find the full region the value is live. 385 while (!LiveInBlockWorklist.empty()) { 386 BasicBlock *BB = LiveInBlockWorklist.pop_back_val(); 387 388 // The block really is live in here, insert it into the set. If already in 389 // the set, then it has already been processed. 390 if (!LiveInBlocks.insert(BB).second) 391 continue; 392 393 // Since the value is live into BB, it is either defined in a predecessor or 394 // live into it to. 395 LiveInBlockWorklist.append(pred_begin(BB), pred_end(BB)); 396 } 397 398 // Determine where our MemoryPhi's should go 399 ForwardIDFCalculator IDFs(*DT); 400 IDFs.setDefiningBlocks(DefiningBlocks); 401 IDFs.setLiveInBlocks(LiveInBlocks); 402 SmallVector<BasicBlock *, 32> IDFBlocks; 403 IDFs.calculate(IDFBlocks); 404 405 // Now place MemoryPhi nodes. 406 for (auto &BB : IDFBlocks) { 407 // Insert phi node 408 AccessList *Accesses = getOrCreateAccessList(BB); 409 MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++); 410 ValueToMemoryAccess.insert(std::make_pair(BB, Phi)); 411 // Phi's always are placed at the front of the block. 412 Accesses->push_front(Phi); 413 } 414 415 // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get 416 // filled in with all blocks. 417 SmallPtrSet<BasicBlock *, 16> Visited; 418 renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited); 419 420 MemorySSAWalker *Walker = getWalker(); 421 422 // Now optimize the MemoryUse's defining access to point to the nearest 423 // dominating clobbering def. 424 // This ensures that MemoryUse's that are killed by the same store are 425 // immediate users of that store, one of the invariants we guarantee. 426 for (auto DomNode : depth_first(DT)) { 427 BasicBlock *BB = DomNode->getBlock(); 428 auto AI = PerBlockAccesses.find(BB); 429 if (AI == PerBlockAccesses.end()) 430 continue; 431 AccessList *Accesses = AI->second.get(); 432 for (auto &MA : *Accesses) { 433 if (auto *MU = dyn_cast<MemoryUse>(&MA)) { 434 Instruction *Inst = MU->getMemoryInst(); 435 MU->setDefiningAccess(Walker->getClobberingMemoryAccess(Inst)); 436 } 437 } 438 } 439 440 // Mark the uses in unreachable blocks as live on entry, so that they go 441 // somewhere. 442 for (auto &BB : F) 443 if (!Visited.count(&BB)) 444 markUnreachableAsLiveOnEntry(&BB); 445 } 446 447 MemorySSAWalker *MemorySSA::getWalker() { 448 if (Walker) 449 return Walker.get(); 450 451 Walker = make_unique<CachingWalker>(this, AA, DT); 452 return Walker.get(); 453 } 454 455 MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) { 456 assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB"); 457 AccessList *Accesses = getOrCreateAccessList(BB); 458 MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++); 459 ValueToMemoryAccess.insert(std::make_pair(BB, Phi)); 460 // Phi's always are placed at the front of the block. 461 Accesses->push_front(Phi); 462 return Phi; 463 } 464 465 MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I, 466 MemoryAccess *Definition) { 467 assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI"); 468 MemoryUseOrDef *NewAccess = createNewAccess(I); 469 assert( 470 NewAccess != nullptr && 471 "Tried to create a memory access for a non-memory touching instruction"); 472 NewAccess->setDefiningAccess(Definition); 473 return NewAccess; 474 } 475 476 MemoryAccess *MemorySSA::createMemoryAccessInBB(Instruction *I, 477 MemoryAccess *Definition, 478 const BasicBlock *BB, 479 InsertionPlace Point) { 480 MemoryUseOrDef *NewAccess = createDefinedAccess(I, Definition); 481 auto *Accesses = getOrCreateAccessList(BB); 482 if (Point == Beginning) { 483 // It goes after any phi nodes 484 auto AI = std::find_if( 485 Accesses->begin(), Accesses->end(), 486 [](const MemoryAccess &MA) { return !isa<MemoryPhi>(MA); }); 487 488 Accesses->insert(AI, NewAccess); 489 } else { 490 Accesses->push_back(NewAccess); 491 } 492 493 return NewAccess; 494 } 495 MemoryAccess *MemorySSA::createMemoryAccessBefore(Instruction *I, 496 MemoryAccess *Definition, 497 MemoryAccess *InsertPt) { 498 assert(I->getParent() == InsertPt->getBlock() && 499 "New and old access must be in the same block"); 500 MemoryUseOrDef *NewAccess = createDefinedAccess(I, Definition); 501 auto *Accesses = getOrCreateAccessList(InsertPt->getBlock()); 502 Accesses->insert(AccessList::iterator(InsertPt), NewAccess); 503 return NewAccess; 504 } 505 506 MemoryAccess *MemorySSA::createMemoryAccessAfter(Instruction *I, 507 MemoryAccess *Definition, 508 MemoryAccess *InsertPt) { 509 assert(I->getParent() == InsertPt->getBlock() && 510 "New and old access must be in the same block"); 511 MemoryUseOrDef *NewAccess = createDefinedAccess(I, Definition); 512 auto *Accesses = getOrCreateAccessList(InsertPt->getBlock()); 513 Accesses->insertAfter(AccessList::iterator(InsertPt), NewAccess); 514 return NewAccess; 515 } 516 517 /// \brief Helper function to create new memory accesses 518 MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I) { 519 // The assume intrinsic has a control dependency which we model by claiming 520 // that it writes arbitrarily. Ignore that fake memory dependency here. 521 // FIXME: Replace this special casing with a more accurate modelling of 522 // assume's control dependency. 523 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 524 if (II->getIntrinsicID() == Intrinsic::assume) 525 return nullptr; 526 527 // Find out what affect this instruction has on memory. 528 ModRefInfo ModRef = AA->getModRefInfo(I); 529 bool Def = bool(ModRef & MRI_Mod); 530 bool Use = bool(ModRef & MRI_Ref); 531 532 // It's possible for an instruction to not modify memory at all. During 533 // construction, we ignore them. 534 if (!Def && !Use) 535 return nullptr; 536 537 assert((Def || Use) && 538 "Trying to create a memory access with a non-memory instruction"); 539 540 MemoryUseOrDef *MUD; 541 if (Def) 542 MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++); 543 else 544 MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent()); 545 ValueToMemoryAccess.insert(std::make_pair(I, MUD)); 546 return MUD; 547 } 548 549 MemoryAccess *MemorySSA::findDominatingDef(BasicBlock *UseBlock, 550 enum InsertionPlace Where) { 551 // Handle the initial case 552 if (Where == Beginning) 553 // The only thing that could define us at the beginning is a phi node 554 if (MemoryPhi *Phi = getMemoryAccess(UseBlock)) 555 return Phi; 556 557 DomTreeNode *CurrNode = DT->getNode(UseBlock); 558 // Need to be defined by our dominator 559 if (Where == Beginning) 560 CurrNode = CurrNode->getIDom(); 561 Where = End; 562 while (CurrNode) { 563 auto It = PerBlockAccesses.find(CurrNode->getBlock()); 564 if (It != PerBlockAccesses.end()) { 565 auto &Accesses = It->second; 566 for (MemoryAccess &RA : reverse(*Accesses)) { 567 if (isa<MemoryDef>(RA) || isa<MemoryPhi>(RA)) 568 return &RA; 569 } 570 } 571 CurrNode = CurrNode->getIDom(); 572 } 573 return LiveOnEntryDef.get(); 574 } 575 576 /// \brief Returns true if \p Replacer dominates \p Replacee . 577 bool MemorySSA::dominatesUse(const MemoryAccess *Replacer, 578 const MemoryAccess *Replacee) const { 579 if (isa<MemoryUseOrDef>(Replacee)) 580 return DT->dominates(Replacer->getBlock(), Replacee->getBlock()); 581 const auto *MP = cast<MemoryPhi>(Replacee); 582 // For a phi node, the use occurs in the predecessor block of the phi node. 583 // Since we may occur multiple times in the phi node, we have to check each 584 // operand to ensure Replacer dominates each operand where Replacee occurs. 585 for (const Use &Arg : MP->operands()) { 586 if (Arg.get() != Replacee && 587 !DT->dominates(Replacer->getBlock(), MP->getIncomingBlock(Arg))) 588 return false; 589 } 590 return true; 591 } 592 593 /// \brief If all arguments of a MemoryPHI are defined by the same incoming 594 /// argument, return that argument. 595 static MemoryAccess *onlySingleValue(MemoryPhi *MP) { 596 MemoryAccess *MA = nullptr; 597 598 for (auto &Arg : MP->operands()) { 599 if (!MA) 600 MA = cast<MemoryAccess>(Arg); 601 else if (MA != Arg) 602 return nullptr; 603 } 604 return MA; 605 } 606 607 /// \brief Properly remove \p MA from all of MemorySSA's lookup tables. 608 /// 609 /// Because of the way the intrusive list and use lists work, it is important to 610 /// do removal in the right order. 611 void MemorySSA::removeFromLookups(MemoryAccess *MA) { 612 assert(MA->use_empty() && 613 "Trying to remove memory access that still has uses"); 614 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA)) 615 MUD->setDefiningAccess(nullptr); 616 // Invalidate our walker's cache if necessary 617 if (!isa<MemoryUse>(MA)) 618 Walker->invalidateInfo(MA); 619 // The call below to erase will destroy MA, so we can't change the order we 620 // are doing things here 621 Value *MemoryInst; 622 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(MA)) { 623 MemoryInst = MUD->getMemoryInst(); 624 } else { 625 MemoryInst = MA->getBlock(); 626 } 627 ValueToMemoryAccess.erase(MemoryInst); 628 629 auto AccessIt = PerBlockAccesses.find(MA->getBlock()); 630 std::unique_ptr<AccessList> &Accesses = AccessIt->second; 631 Accesses->erase(MA); 632 if (Accesses->empty()) 633 PerBlockAccesses.erase(AccessIt); 634 } 635 636 void MemorySSA::removeMemoryAccess(MemoryAccess *MA) { 637 assert(!isLiveOnEntryDef(MA) && "Trying to remove the live on entry def"); 638 // We can only delete phi nodes if they have no uses, or we can replace all 639 // uses with a single definition. 640 MemoryAccess *NewDefTarget = nullptr; 641 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(MA)) { 642 // Note that it is sufficient to know that all edges of the phi node have 643 // the same argument. If they do, by the definition of dominance frontiers 644 // (which we used to place this phi), that argument must dominate this phi, 645 // and thus, must dominate the phi's uses, and so we will not hit the assert 646 // below. 647 NewDefTarget = onlySingleValue(MP); 648 assert((NewDefTarget || MP->use_empty()) && 649 "We can't delete this memory phi"); 650 } else { 651 NewDefTarget = cast<MemoryUseOrDef>(MA)->getDefiningAccess(); 652 } 653 654 // Re-point the uses at our defining access 655 if (!MA->use_empty()) 656 MA->replaceAllUsesWith(NewDefTarget); 657 658 // The call below to erase will destroy MA, so we can't change the order we 659 // are doing things here 660 removeFromLookups(MA); 661 } 662 663 void MemorySSA::print(raw_ostream &OS) const { 664 MemorySSAAnnotatedWriter Writer(this); 665 F.print(OS, &Writer); 666 } 667 668 void MemorySSA::dump() const { 669 MemorySSAAnnotatedWriter Writer(this); 670 F.print(dbgs(), &Writer); 671 } 672 673 void MemorySSA::verifyMemorySSA() const { 674 verifyDefUses(F); 675 verifyDomination(F); 676 verifyOrdering(F); 677 } 678 679 /// \brief Verify that the order and existence of MemoryAccesses matches the 680 /// order and existence of memory affecting instructions. 681 void MemorySSA::verifyOrdering(Function &F) const { 682 // Walk all the blocks, comparing what the lookups think and what the access 683 // lists think, as well as the order in the blocks vs the order in the access 684 // lists. 685 SmallVector<MemoryAccess *, 32> ActualAccesses; 686 for (BasicBlock &B : F) { 687 const AccessList *AL = getBlockAccesses(&B); 688 MemoryAccess *Phi = getMemoryAccess(&B); 689 if (Phi) 690 ActualAccesses.push_back(Phi); 691 for (Instruction &I : B) { 692 MemoryAccess *MA = getMemoryAccess(&I); 693 assert((!MA || AL) && "We have memory affecting instructions " 694 "in this block but they are not in the " 695 "access list"); 696 if (MA) 697 ActualAccesses.push_back(MA); 698 } 699 // Either we hit the assert, really have no accesses, or we have both 700 // accesses and an access list 701 if (!AL) 702 continue; 703 assert(AL->size() == ActualAccesses.size() && 704 "We don't have the same number of accesses in the block as on the " 705 "access list"); 706 auto ALI = AL->begin(); 707 auto AAI = ActualAccesses.begin(); 708 while (ALI != AL->end() && AAI != ActualAccesses.end()) { 709 assert(&*ALI == *AAI && "Not the same accesses in the same order"); 710 ++ALI; 711 ++AAI; 712 } 713 ActualAccesses.clear(); 714 } 715 } 716 717 /// \brief Verify the domination properties of MemorySSA by checking that each 718 /// definition dominates all of its uses. 719 void MemorySSA::verifyDomination(Function &F) const { 720 for (BasicBlock &B : F) { 721 // Phi nodes are attached to basic blocks 722 if (MemoryPhi *MP = getMemoryAccess(&B)) { 723 for (User *U : MP->users()) { 724 BasicBlock *UseBlock; 725 // Phi operands are used on edges, we simulate the right domination by 726 // acting as if the use occurred at the end of the predecessor block. 727 if (MemoryPhi *P = dyn_cast<MemoryPhi>(U)) { 728 for (const auto &Arg : P->operands()) { 729 if (Arg == MP) { 730 UseBlock = P->getIncomingBlock(Arg); 731 break; 732 } 733 } 734 } else { 735 UseBlock = cast<MemoryAccess>(U)->getBlock(); 736 } 737 (void)UseBlock; 738 assert(DT->dominates(MP->getBlock(), UseBlock) && 739 "Memory PHI does not dominate it's uses"); 740 } 741 } 742 743 for (Instruction &I : B) { 744 MemoryAccess *MD = dyn_cast_or_null<MemoryDef>(getMemoryAccess(&I)); 745 if (!MD) 746 continue; 747 748 for (User *U : MD->users()) { 749 BasicBlock *UseBlock; 750 (void)UseBlock; 751 // Things are allowed to flow to phi nodes over their predecessor edge. 752 if (auto *P = dyn_cast<MemoryPhi>(U)) { 753 for (const auto &Arg : P->operands()) { 754 if (Arg == MD) { 755 UseBlock = P->getIncomingBlock(Arg); 756 break; 757 } 758 } 759 } else { 760 UseBlock = cast<MemoryAccess>(U)->getBlock(); 761 } 762 assert(DT->dominates(MD->getBlock(), UseBlock) && 763 "Memory Def does not dominate it's uses"); 764 } 765 } 766 } 767 } 768 769 /// \brief Verify the def-use lists in MemorySSA, by verifying that \p Use 770 /// appears in the use list of \p Def. 771 /// 772 /// llvm_unreachable is used instead of asserts because this may be called in 773 /// a build without asserts. In that case, we don't want this to turn into a 774 /// nop. 775 void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const { 776 // The live on entry use may cause us to get a NULL def here 777 if (!Def) { 778 if (!isLiveOnEntryDef(Use)) 779 llvm_unreachable("Null def but use not point to live on entry def"); 780 } else if (std::find(Def->user_begin(), Def->user_end(), Use) == 781 Def->user_end()) { 782 llvm_unreachable("Did not find use in def's use list"); 783 } 784 } 785 786 /// \brief Verify the immediate use information, by walking all the memory 787 /// accesses and verifying that, for each use, it appears in the 788 /// appropriate def's use list 789 void MemorySSA::verifyDefUses(Function &F) const { 790 for (BasicBlock &B : F) { 791 // Phi nodes are attached to basic blocks 792 if (MemoryPhi *Phi = getMemoryAccess(&B)) { 793 assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance( 794 pred_begin(&B), pred_end(&B))) && 795 "Incomplete MemoryPhi Node"); 796 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) 797 verifyUseInDefs(Phi->getIncomingValue(I), Phi); 798 } 799 800 for (Instruction &I : B) { 801 if (MemoryAccess *MA = getMemoryAccess(&I)) { 802 assert(isa<MemoryUseOrDef>(MA) && 803 "Found a phi node not attached to a bb"); 804 verifyUseInDefs(cast<MemoryUseOrDef>(MA)->getDefiningAccess(), MA); 805 } 806 } 807 } 808 } 809 810 MemoryAccess *MemorySSA::getMemoryAccess(const Value *I) const { 811 return ValueToMemoryAccess.lookup(I); 812 } 813 814 MemoryPhi *MemorySSA::getMemoryAccess(const BasicBlock *BB) const { 815 return cast_or_null<MemoryPhi>(getMemoryAccess((const Value *)BB)); 816 } 817 818 /// \brief Determine, for two memory accesses in the same block, 819 /// whether \p Dominator dominates \p Dominatee. 820 /// \returns True if \p Dominator dominates \p Dominatee. 821 bool MemorySSA::locallyDominates(const MemoryAccess *Dominator, 822 const MemoryAccess *Dominatee) const { 823 824 assert((Dominator->getBlock() == Dominatee->getBlock()) && 825 "Asking for local domination when accesses are in different blocks!"); 826 827 // A node dominates itself. 828 if (Dominatee == Dominator) 829 return true; 830 831 // When Dominatee is defined on function entry, it is not dominated by another 832 // memory access. 833 if (isLiveOnEntryDef(Dominatee)) 834 return false; 835 836 // When Dominator is defined on function entry, it dominates the other memory 837 // access. 838 if (isLiveOnEntryDef(Dominator)) 839 return true; 840 841 // Get the access list for the block 842 const AccessList *AccessList = getBlockAccesses(Dominator->getBlock()); 843 AccessList::const_reverse_iterator It(Dominator->getIterator()); 844 845 // If we hit the beginning of the access list before we hit dominatee, we must 846 // dominate it 847 return std::none_of(It, AccessList->rend(), 848 [&](const MemoryAccess &MA) { return &MA == Dominatee; }); 849 } 850 851 const static char LiveOnEntryStr[] = "liveOnEntry"; 852 853 void MemoryDef::print(raw_ostream &OS) const { 854 MemoryAccess *UO = getDefiningAccess(); 855 856 OS << getID() << " = MemoryDef("; 857 if (UO && UO->getID()) 858 OS << UO->getID(); 859 else 860 OS << LiveOnEntryStr; 861 OS << ')'; 862 } 863 864 void MemoryPhi::print(raw_ostream &OS) const { 865 bool First = true; 866 OS << getID() << " = MemoryPhi("; 867 for (const auto &Op : operands()) { 868 BasicBlock *BB = getIncomingBlock(Op); 869 MemoryAccess *MA = cast<MemoryAccess>(Op); 870 if (!First) 871 OS << ','; 872 else 873 First = false; 874 875 OS << '{'; 876 if (BB->hasName()) 877 OS << BB->getName(); 878 else 879 BB->printAsOperand(OS, false); 880 OS << ','; 881 if (unsigned ID = MA->getID()) 882 OS << ID; 883 else 884 OS << LiveOnEntryStr; 885 OS << '}'; 886 } 887 OS << ')'; 888 } 889 890 MemoryAccess::~MemoryAccess() {} 891 892 void MemoryUse::print(raw_ostream &OS) const { 893 MemoryAccess *UO = getDefiningAccess(); 894 OS << "MemoryUse("; 895 if (UO && UO->getID()) 896 OS << UO->getID(); 897 else 898 OS << LiveOnEntryStr; 899 OS << ')'; 900 } 901 902 void MemoryAccess::dump() const { 903 print(dbgs()); 904 dbgs() << "\n"; 905 } 906 907 char MemorySSAPrinterLegacyPass::ID = 0; 908 909 MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) { 910 initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry()); 911 } 912 913 void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const { 914 AU.setPreservesAll(); 915 AU.addRequired<MemorySSAWrapperPass>(); 916 AU.addPreserved<MemorySSAWrapperPass>(); 917 } 918 919 bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) { 920 auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA(); 921 MSSA.print(dbgs()); 922 if (VerifyMemorySSA) 923 MSSA.verifyMemorySSA(); 924 return false; 925 } 926 927 char MemorySSAAnalysis::PassID; 928 929 MemorySSA MemorySSAAnalysis::run(Function &F, AnalysisManager<Function> &AM) { 930 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 931 auto &AA = AM.getResult<AAManager>(F); 932 return MemorySSA(F, &AA, &DT); 933 } 934 935 PreservedAnalyses MemorySSAPrinterPass::run(Function &F, 936 FunctionAnalysisManager &AM) { 937 OS << "MemorySSA for function: " << F.getName() << "\n"; 938 AM.getResult<MemorySSAAnalysis>(F).print(OS); 939 940 return PreservedAnalyses::all(); 941 } 942 943 PreservedAnalyses MemorySSAVerifierPass::run(Function &F, 944 FunctionAnalysisManager &AM) { 945 AM.getResult<MemorySSAAnalysis>(F).verifyMemorySSA(); 946 947 return PreservedAnalyses::all(); 948 } 949 950 char MemorySSAWrapperPass::ID = 0; 951 952 MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) { 953 initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry()); 954 } 955 956 void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); } 957 958 void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 959 AU.setPreservesAll(); 960 AU.addRequiredTransitive<DominatorTreeWrapperPass>(); 961 AU.addRequiredTransitive<AAResultsWrapperPass>(); 962 } 963 964 bool MemorySSAWrapperPass::runOnFunction(Function &F) { 965 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 966 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); 967 MSSA.reset(new MemorySSA(F, &AA, &DT)); 968 return false; 969 } 970 971 void MemorySSAWrapperPass::verifyAnalysis() const { MSSA->verifyMemorySSA(); } 972 973 void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const { 974 MSSA->print(OS); 975 } 976 977 MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {} 978 979 MemorySSA::CachingWalker::CachingWalker(MemorySSA *M, AliasAnalysis *A, 980 DominatorTree *D) 981 : MemorySSAWalker(M), AA(A), DT(D) {} 982 983 MemorySSA::CachingWalker::~CachingWalker() {} 984 985 struct MemorySSA::CachingWalker::UpwardsMemoryQuery { 986 // True if we saw a phi whose predecessor was a backedge 987 bool SawBackedgePhi; 988 // True if our original query started off as a call 989 bool IsCall; 990 // The pointer location we started the query with. This will be empty if 991 // IsCall is true. 992 MemoryLocation StartingLoc; 993 // This is the instruction we were querying about. 994 const Instruction *Inst; 995 // Set of visited Instructions for this query. 996 DenseSet<MemoryAccessPair> Visited; 997 // Vector of visited call accesses for this query. This is separated out 998 // because you can always cache and lookup the result of call queries (IE when 999 // IsCall == true) for every call in the chain. The calls have no AA location 1000 // associated with them with them, and thus, no context dependence. 1001 SmallVector<const MemoryAccess *, 32> VisitedCalls; 1002 // The MemoryAccess we actually got called with, used to test local domination 1003 const MemoryAccess *OriginalAccess; 1004 1005 UpwardsMemoryQuery() 1006 : SawBackedgePhi(false), IsCall(false), Inst(nullptr), 1007 OriginalAccess(nullptr) {} 1008 1009 UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access) 1010 : SawBackedgePhi(false), IsCall(ImmutableCallSite(Inst)), Inst(Inst), 1011 OriginalAccess(Access) {} 1012 }; 1013 1014 void MemorySSA::CachingWalker::invalidateInfo(MemoryAccess *MA) { 1015 1016 // TODO: We can do much better cache invalidation with differently stored 1017 // caches. For now, for MemoryUses, we simply remove them 1018 // from the cache, and kill the entire call/non-call cache for everything 1019 // else. The problem is for phis or defs, currently we'd need to follow use 1020 // chains down and invalidate anything below us in the chain that currently 1021 // terminates at this access. 1022 1023 // See if this is a MemoryUse, if so, just remove the cached info. MemoryUse 1024 // is by definition never a barrier, so nothing in the cache could point to 1025 // this use. In that case, we only need invalidate the info for the use 1026 // itself. 1027 1028 if (MemoryUse *MU = dyn_cast<MemoryUse>(MA)) { 1029 UpwardsMemoryQuery Q; 1030 Instruction *I = MU->getMemoryInst(); 1031 Q.IsCall = bool(ImmutableCallSite(I)); 1032 Q.Inst = I; 1033 if (!Q.IsCall) 1034 Q.StartingLoc = MemoryLocation::get(I); 1035 doCacheRemove(MA, Q, Q.StartingLoc); 1036 } else { 1037 // If it is not a use, the best we can do right now is destroy the cache. 1038 CachedUpwardsClobberingCall.clear(); 1039 CachedUpwardsClobberingAccess.clear(); 1040 } 1041 1042 #ifdef EXPENSIVE_CHECKS 1043 // Run this only when expensive checks are enabled. 1044 verifyRemoved(MA); 1045 #endif 1046 } 1047 1048 void MemorySSA::CachingWalker::doCacheRemove(const MemoryAccess *M, 1049 const UpwardsMemoryQuery &Q, 1050 const MemoryLocation &Loc) { 1051 if (Q.IsCall) 1052 CachedUpwardsClobberingCall.erase(M); 1053 else 1054 CachedUpwardsClobberingAccess.erase({M, Loc}); 1055 } 1056 1057 void MemorySSA::CachingWalker::doCacheInsert(const MemoryAccess *M, 1058 MemoryAccess *Result, 1059 const UpwardsMemoryQuery &Q, 1060 const MemoryLocation &Loc) { 1061 // This is fine for Phis, since there are times where we can't optimize them. 1062 // Making a def its own clobber is never correct, though. 1063 assert((Result != M || isa<MemoryPhi>(M)) && 1064 "Something can't clobber itself!"); 1065 ++NumClobberCacheInserts; 1066 if (Q.IsCall) 1067 CachedUpwardsClobberingCall[M] = Result; 1068 else 1069 CachedUpwardsClobberingAccess[{M, Loc}] = Result; 1070 } 1071 1072 MemoryAccess * 1073 MemorySSA::CachingWalker::doCacheLookup(const MemoryAccess *M, 1074 const UpwardsMemoryQuery &Q, 1075 const MemoryLocation &Loc) { 1076 ++NumClobberCacheLookups; 1077 MemoryAccess *Result; 1078 1079 if (Q.IsCall) 1080 Result = CachedUpwardsClobberingCall.lookup(M); 1081 else 1082 Result = CachedUpwardsClobberingAccess.lookup({M, Loc}); 1083 1084 if (Result) 1085 ++NumClobberCacheHits; 1086 return Result; 1087 } 1088 1089 bool MemorySSA::CachingWalker::instructionClobbersQuery( 1090 const MemoryDef *MD, UpwardsMemoryQuery &Q, 1091 const MemoryLocation &Loc) const { 1092 Instruction *DefMemoryInst = MD->getMemoryInst(); 1093 assert(DefMemoryInst && "Defining instruction not actually an instruction"); 1094 1095 if (!Q.IsCall) 1096 return AA->getModRefInfo(DefMemoryInst, Loc) & MRI_Mod; 1097 1098 // If this is a call, mark it for caching 1099 if (ImmutableCallSite(DefMemoryInst)) 1100 Q.VisitedCalls.push_back(MD); 1101 ModRefInfo I = AA->getModRefInfo(DefMemoryInst, ImmutableCallSite(Q.Inst)); 1102 return I != MRI_NoModRef; 1103 } 1104 1105 MemoryAccessPair MemorySSA::CachingWalker::UpwardsDFSWalk( 1106 MemoryAccess *StartingAccess, const MemoryLocation &Loc, 1107 UpwardsMemoryQuery &Q, bool FollowingBackedge) { 1108 MemoryAccess *ModifyingAccess = nullptr; 1109 1110 auto DFI = df_begin(StartingAccess); 1111 for (auto DFE = df_end(StartingAccess); DFI != DFE;) { 1112 MemoryAccess *CurrAccess = *DFI; 1113 if (MSSA->isLiveOnEntryDef(CurrAccess)) 1114 return {CurrAccess, Loc}; 1115 // If this is a MemoryDef, check whether it clobbers our current query. This 1116 // needs to be done before consulting the cache, because the cache reports 1117 // the clobber for CurrAccess. If CurrAccess is a clobber for this query, 1118 // and we ask the cache for information first, then we might skip this 1119 // clobber, which is bad. 1120 if (auto *MD = dyn_cast<MemoryDef>(CurrAccess)) { 1121 // If we hit the top, stop following this path. 1122 // While we can do lookups, we can't sanely do inserts here unless we were 1123 // to track everything we saw along the way, since we don't know where we 1124 // will stop. 1125 if (instructionClobbersQuery(MD, Q, Loc)) { 1126 ModifyingAccess = CurrAccess; 1127 break; 1128 } 1129 } 1130 if (auto CacheResult = doCacheLookup(CurrAccess, Q, Loc)) 1131 return {CacheResult, Loc}; 1132 1133 // We need to know whether it is a phi so we can track backedges. 1134 // Otherwise, walk all upward defs. 1135 if (!isa<MemoryPhi>(CurrAccess)) { 1136 ++DFI; 1137 continue; 1138 } 1139 1140 #ifndef NDEBUG 1141 // The loop below visits the phi's children for us. Because phis are the 1142 // only things with multiple edges, skipping the children should always lead 1143 // us to the end of the loop. 1144 // 1145 // Use a copy of DFI because skipChildren would kill our search stack, which 1146 // would make caching anything on the way back impossible. 1147 auto DFICopy = DFI; 1148 assert(DFICopy.skipChildren() == DFE && 1149 "Skipping phi's children doesn't end the DFS?"); 1150 #endif 1151 1152 const MemoryAccessPair PHIPair(CurrAccess, Loc); 1153 1154 // Don't try to optimize this phi again if we've already tried to do so. 1155 if (!Q.Visited.insert(PHIPair).second) { 1156 ModifyingAccess = CurrAccess; 1157 break; 1158 } 1159 1160 std::size_t InitialVisitedCallSize = Q.VisitedCalls.size(); 1161 1162 // Recurse on PHI nodes, since we need to change locations. 1163 // TODO: Allow graphtraits on pairs, which would turn this whole function 1164 // into a normal single depth first walk. 1165 MemoryAccess *FirstDef = nullptr; 1166 for (auto MPI = upward_defs_begin(PHIPair), MPE = upward_defs_end(); 1167 MPI != MPE; ++MPI) { 1168 bool Backedge = 1169 !FollowingBackedge && 1170 DT->dominates(CurrAccess->getBlock(), MPI.getPhiArgBlock()); 1171 1172 MemoryAccessPair CurrentPair = 1173 UpwardsDFSWalk(MPI->first, MPI->second, Q, Backedge); 1174 // All the phi arguments should reach the same point if we can bypass 1175 // this phi. The alternative is that they hit this phi node, which 1176 // means we can skip this argument. 1177 if (FirstDef && CurrentPair.first != PHIPair.first && 1178 CurrentPair.first != FirstDef) { 1179 ModifyingAccess = CurrAccess; 1180 break; 1181 } 1182 1183 if (!FirstDef) 1184 FirstDef = CurrentPair.first; 1185 } 1186 1187 // If we exited the loop early, go with the result it gave us. 1188 if (!ModifyingAccess) { 1189 assert(FirstDef && "Found a Phi with no upward defs?"); 1190 ModifyingAccess = FirstDef; 1191 } else { 1192 // If we can't optimize this Phi, then we can't safely cache any of the 1193 // calls we visited when trying to optimize it. Wipe them out now. 1194 Q.VisitedCalls.resize(InitialVisitedCallSize); 1195 } 1196 break; 1197 } 1198 1199 if (!ModifyingAccess) 1200 return {MSSA->getLiveOnEntryDef(), Q.StartingLoc}; 1201 1202 const BasicBlock *OriginalBlock = StartingAccess->getBlock(); 1203 assert(DFI.getPathLength() > 0 && "We dropped our path?"); 1204 unsigned N = DFI.getPathLength(); 1205 // If we found a clobbering def, the last element in the path will be our 1206 // clobber, so we don't want to cache that to itself. OTOH, if we optimized a 1207 // phi, we can add the last thing in the path to the cache, since that won't 1208 // be the result. 1209 if (DFI.getPath(N - 1) == ModifyingAccess) 1210 --N; 1211 for (; N > 1; --N) { 1212 MemoryAccess *CacheAccess = DFI.getPath(N - 1); 1213 BasicBlock *CurrBlock = CacheAccess->getBlock(); 1214 if (!FollowingBackedge) 1215 doCacheInsert(CacheAccess, ModifyingAccess, Q, Loc); 1216 if (DT->dominates(CurrBlock, OriginalBlock) && 1217 (CurrBlock != OriginalBlock || !FollowingBackedge || 1218 MSSA->locallyDominates(CacheAccess, StartingAccess))) 1219 break; 1220 } 1221 1222 // Cache everything else on the way back. The caller should cache 1223 // StartingAccess for us. 1224 for (; N > 1; --N) { 1225 MemoryAccess *CacheAccess = DFI.getPath(N - 1); 1226 doCacheInsert(CacheAccess, ModifyingAccess, Q, Loc); 1227 } 1228 assert(Q.Visited.size() < 1000 && "Visited too much"); 1229 1230 return {ModifyingAccess, Loc}; 1231 } 1232 1233 /// \brief Walk the use-def chains starting at \p MA and find 1234 /// the MemoryAccess that actually clobbers Loc. 1235 /// 1236 /// \returns our clobbering memory access 1237 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess( 1238 MemoryAccess *StartingAccess, UpwardsMemoryQuery &Q) { 1239 return UpwardsDFSWalk(StartingAccess, Q.StartingLoc, Q, false).first; 1240 } 1241 1242 MemoryAccess *MemorySSA::CachingWalker::getClobberingMemoryAccess( 1243 MemoryAccess *StartingAccess, MemoryLocation &Loc) { 1244 if (isa<MemoryPhi>(StartingAccess)) 1245 return StartingAccess; 1246 1247 auto *StartingUseOrDef = cast<MemoryUseOrDef>(StartingAccess); 1248 if (MSSA->isLiveOnEntryDef(StartingUseOrDef)) 1249 return StartingUseOrDef; 1250 1251 Instruction *I = StartingUseOrDef->getMemoryInst(); 1252 1253 // Conservatively, fences are always clobbers, so don't perform the walk if we 1254 // hit a fence. 1255 if (isa<FenceInst>(I)) 1256 return StartingUseOrDef; 1257 1258 UpwardsMemoryQuery Q; 1259 Q.OriginalAccess = StartingUseOrDef; 1260 Q.StartingLoc = Loc; 1261 Q.Inst = StartingUseOrDef->getMemoryInst(); 1262 Q.IsCall = false; 1263 1264 if (auto CacheResult = doCacheLookup(StartingUseOrDef, Q, Q.StartingLoc)) 1265 return CacheResult; 1266 1267 // Unlike the other function, do not walk to the def of a def, because we are 1268 // handed something we already believe is the clobbering access. 1269 MemoryAccess *DefiningAccess = isa<MemoryUse>(StartingUseOrDef) 1270 ? StartingUseOrDef->getDefiningAccess() 1271 : StartingUseOrDef; 1272 1273 MemoryAccess *Clobber = getClobberingMemoryAccess(DefiningAccess, Q); 1274 // Only cache this if it wouldn't make Clobber point to itself. 1275 if (Clobber != StartingAccess) 1276 doCacheInsert(Q.OriginalAccess, Clobber, Q, Q.StartingLoc); 1277 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is "); 1278 DEBUG(dbgs() << *StartingUseOrDef << "\n"); 1279 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is "); 1280 DEBUG(dbgs() << *Clobber << "\n"); 1281 return Clobber; 1282 } 1283 1284 MemoryAccess * 1285 MemorySSA::CachingWalker::getClobberingMemoryAccess(const Instruction *I) { 1286 // There should be no way to lookup an instruction and get a phi as the 1287 // access, since we only map BB's to PHI's. So, this must be a use or def. 1288 auto *StartingAccess = cast<MemoryUseOrDef>(MSSA->getMemoryAccess(I)); 1289 1290 // We can't sanely do anything with a FenceInst, they conservatively 1291 // clobber all memory, and have no locations to get pointers from to 1292 // try to disambiguate 1293 if (isa<FenceInst>(I)) 1294 return StartingAccess; 1295 1296 UpwardsMemoryQuery Q; 1297 Q.OriginalAccess = StartingAccess; 1298 Q.IsCall = bool(ImmutableCallSite(I)); 1299 if (!Q.IsCall) 1300 Q.StartingLoc = MemoryLocation::get(I); 1301 Q.Inst = I; 1302 if (auto CacheResult = doCacheLookup(StartingAccess, Q, Q.StartingLoc)) 1303 return CacheResult; 1304 1305 // Start with the thing we already think clobbers this location 1306 MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess(); 1307 1308 // At this point, DefiningAccess may be the live on entry def. 1309 // If it is, we will not get a better result. 1310 if (MSSA->isLiveOnEntryDef(DefiningAccess)) 1311 return DefiningAccess; 1312 1313 MemoryAccess *Result = getClobberingMemoryAccess(DefiningAccess, Q); 1314 // DFS won't cache a result for DefiningAccess. So, if DefiningAccess isn't 1315 // our clobber, be sure that it gets a cache entry, too. 1316 if (Result != DefiningAccess) 1317 doCacheInsert(DefiningAccess, Result, Q, Q.StartingLoc); 1318 doCacheInsert(Q.OriginalAccess, Result, Q, Q.StartingLoc); 1319 // TODO: When this implementation is more mature, we may want to figure out 1320 // what this additional caching buys us. It's most likely A Good Thing. 1321 if (Q.IsCall) 1322 for (const MemoryAccess *MA : Q.VisitedCalls) 1323 if (MA != Result) 1324 doCacheInsert(MA, Result, Q, Q.StartingLoc); 1325 1326 DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is "); 1327 DEBUG(dbgs() << *DefiningAccess << "\n"); 1328 DEBUG(dbgs() << "Final Memory SSA clobber for " << *I << " is "); 1329 DEBUG(dbgs() << *Result << "\n"); 1330 1331 return Result; 1332 } 1333 1334 // Verify that MA doesn't exist in any of the caches. 1335 void MemorySSA::CachingWalker::verifyRemoved(MemoryAccess *MA) { 1336 #ifndef NDEBUG 1337 for (auto &P : CachedUpwardsClobberingAccess) 1338 assert(P.first.first != MA && P.second != MA && 1339 "Found removed MemoryAccess in cache."); 1340 for (auto &P : CachedUpwardsClobberingCall) 1341 assert(P.first != MA && P.second != MA && 1342 "Found removed MemoryAccess in cache."); 1343 #endif // !NDEBUG 1344 } 1345 1346 MemoryAccess * 1347 DoNothingMemorySSAWalker::getClobberingMemoryAccess(const Instruction *I) { 1348 MemoryAccess *MA = MSSA->getMemoryAccess(I); 1349 if (auto *Use = dyn_cast<MemoryUseOrDef>(MA)) 1350 return Use->getDefiningAccess(); 1351 return MA; 1352 } 1353 1354 MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess( 1355 MemoryAccess *StartingAccess, MemoryLocation &) { 1356 if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess)) 1357 return Use->getDefiningAccess(); 1358 return StartingAccess; 1359 } 1360 } 1361