1 //===- MemorySSA.cpp - Memory SSA Builder ---------------------------------===// 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 the MemorySSA class. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "llvm/Analysis/MemorySSA.h" 14 #include "llvm/ADT/DenseMap.h" 15 #include "llvm/ADT/DenseMapInfo.h" 16 #include "llvm/ADT/DenseSet.h" 17 #include "llvm/ADT/DepthFirstIterator.h" 18 #include "llvm/ADT/Hashing.h" 19 #include "llvm/ADT/STLExtras.h" 20 #include "llvm/ADT/SmallPtrSet.h" 21 #include "llvm/ADT/SmallVector.h" 22 #include "llvm/ADT/StringExtras.h" 23 #include "llvm/ADT/iterator.h" 24 #include "llvm/ADT/iterator_range.h" 25 #include "llvm/Analysis/AliasAnalysis.h" 26 #include "llvm/Analysis/CFGPrinter.h" 27 #include "llvm/Analysis/IteratedDominanceFrontier.h" 28 #include "llvm/Analysis/MemoryLocation.h" 29 #include "llvm/Config/llvm-config.h" 30 #include "llvm/IR/AssemblyAnnotationWriter.h" 31 #include "llvm/IR/BasicBlock.h" 32 #include "llvm/IR/Dominators.h" 33 #include "llvm/IR/Function.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/Operator.h" 39 #include "llvm/IR/PassManager.h" 40 #include "llvm/IR/Use.h" 41 #include "llvm/InitializePasses.h" 42 #include "llvm/Pass.h" 43 #include "llvm/Support/AtomicOrdering.h" 44 #include "llvm/Support/Casting.h" 45 #include "llvm/Support/CommandLine.h" 46 #include "llvm/Support/Compiler.h" 47 #include "llvm/Support/Debug.h" 48 #include "llvm/Support/ErrorHandling.h" 49 #include "llvm/Support/FormattedStream.h" 50 #include "llvm/Support/GraphWriter.h" 51 #include "llvm/Support/raw_ostream.h" 52 #include <algorithm> 53 #include <cassert> 54 #include <iterator> 55 #include <memory> 56 #include <utility> 57 58 using namespace llvm; 59 60 #define DEBUG_TYPE "memoryssa" 61 62 static cl::opt<std::string> 63 DotCFGMSSA("dot-cfg-mssa", 64 cl::value_desc("file name for generated dot file"), 65 cl::desc("file name for generated dot file"), cl::init("")); 66 67 INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false, 68 true) 69 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 70 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 71 INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false, 72 true) 73 74 static cl::opt<unsigned> MaxCheckLimit( 75 "memssa-check-limit", cl::Hidden, cl::init(100), 76 cl::desc("The maximum number of stores/phis MemorySSA" 77 "will consider trying to walk past (default = 100)")); 78 79 // Always verify MemorySSA if expensive checking is enabled. 80 #ifdef EXPENSIVE_CHECKS 81 bool llvm::VerifyMemorySSA = true; 82 #else 83 bool llvm::VerifyMemorySSA = false; 84 #endif 85 86 static cl::opt<bool, true> 87 VerifyMemorySSAX("verify-memoryssa", cl::location(VerifyMemorySSA), 88 cl::Hidden, cl::desc("Enable verification of MemorySSA.")); 89 90 const static char LiveOnEntryStr[] = "liveOnEntry"; 91 92 namespace { 93 94 /// An assembly annotator class to print Memory SSA information in 95 /// comments. 96 class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter { 97 const MemorySSA *MSSA; 98 99 public: 100 MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {} 101 102 void emitBasicBlockStartAnnot(const BasicBlock *BB, 103 formatted_raw_ostream &OS) override { 104 if (MemoryAccess *MA = MSSA->getMemoryAccess(BB)) 105 OS << "; " << *MA << "\n"; 106 } 107 108 void emitInstructionAnnot(const Instruction *I, 109 formatted_raw_ostream &OS) override { 110 if (MemoryAccess *MA = MSSA->getMemoryAccess(I)) 111 OS << "; " << *MA << "\n"; 112 } 113 }; 114 115 /// An assembly annotator class to print Memory SSA information in 116 /// comments. 117 class MemorySSAWalkerAnnotatedWriter : public AssemblyAnnotationWriter { 118 MemorySSA *MSSA; 119 MemorySSAWalker *Walker; 120 BatchAAResults BAA; 121 122 public: 123 MemorySSAWalkerAnnotatedWriter(MemorySSA *M) 124 : MSSA(M), Walker(M->getWalker()), BAA(M->getAA()) {} 125 126 void emitBasicBlockStartAnnot(const BasicBlock *BB, 127 formatted_raw_ostream &OS) override { 128 if (MemoryAccess *MA = MSSA->getMemoryAccess(BB)) 129 OS << "; " << *MA << "\n"; 130 } 131 132 void emitInstructionAnnot(const Instruction *I, 133 formatted_raw_ostream &OS) override { 134 if (MemoryAccess *MA = MSSA->getMemoryAccess(I)) { 135 MemoryAccess *Clobber = Walker->getClobberingMemoryAccess(MA, BAA); 136 OS << "; " << *MA; 137 if (Clobber) { 138 OS << " - clobbered by "; 139 if (MSSA->isLiveOnEntryDef(Clobber)) 140 OS << LiveOnEntryStr; 141 else 142 OS << *Clobber; 143 } 144 OS << "\n"; 145 } 146 } 147 }; 148 149 } // namespace 150 151 namespace { 152 153 /// Our current alias analysis API differentiates heavily between calls and 154 /// non-calls, and functions called on one usually assert on the other. 155 /// This class encapsulates the distinction to simplify other code that wants 156 /// "Memory affecting instructions and related data" to use as a key. 157 /// For example, this class is used as a densemap key in the use optimizer. 158 class MemoryLocOrCall { 159 public: 160 bool IsCall = false; 161 162 MemoryLocOrCall(MemoryUseOrDef *MUD) 163 : MemoryLocOrCall(MUD->getMemoryInst()) {} 164 MemoryLocOrCall(const MemoryUseOrDef *MUD) 165 : MemoryLocOrCall(MUD->getMemoryInst()) {} 166 167 MemoryLocOrCall(Instruction *Inst) { 168 if (auto *C = dyn_cast<CallBase>(Inst)) { 169 IsCall = true; 170 Call = C; 171 } else { 172 IsCall = false; 173 // There is no such thing as a memorylocation for a fence inst, and it is 174 // unique in that regard. 175 if (!isa<FenceInst>(Inst)) 176 Loc = MemoryLocation::get(Inst); 177 } 178 } 179 180 explicit MemoryLocOrCall(const MemoryLocation &Loc) : Loc(Loc) {} 181 182 const CallBase *getCall() const { 183 assert(IsCall); 184 return Call; 185 } 186 187 MemoryLocation getLoc() const { 188 assert(!IsCall); 189 return Loc; 190 } 191 192 bool operator==(const MemoryLocOrCall &Other) const { 193 if (IsCall != Other.IsCall) 194 return false; 195 196 if (!IsCall) 197 return Loc == Other.Loc; 198 199 if (Call->getCalledOperand() != Other.Call->getCalledOperand()) 200 return false; 201 202 return Call->arg_size() == Other.Call->arg_size() && 203 std::equal(Call->arg_begin(), Call->arg_end(), 204 Other.Call->arg_begin()); 205 } 206 207 private: 208 union { 209 const CallBase *Call; 210 MemoryLocation Loc; 211 }; 212 }; 213 214 } // end anonymous namespace 215 216 namespace llvm { 217 218 template <> struct DenseMapInfo<MemoryLocOrCall> { 219 static inline MemoryLocOrCall getEmptyKey() { 220 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey()); 221 } 222 223 static inline MemoryLocOrCall getTombstoneKey() { 224 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey()); 225 } 226 227 static unsigned getHashValue(const MemoryLocOrCall &MLOC) { 228 if (!MLOC.IsCall) 229 return hash_combine( 230 MLOC.IsCall, 231 DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc())); 232 233 hash_code hash = 234 hash_combine(MLOC.IsCall, DenseMapInfo<const Value *>::getHashValue( 235 MLOC.getCall()->getCalledOperand())); 236 237 for (const Value *Arg : MLOC.getCall()->args()) 238 hash = hash_combine(hash, DenseMapInfo<const Value *>::getHashValue(Arg)); 239 return hash; 240 } 241 242 static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) { 243 return LHS == RHS; 244 } 245 }; 246 247 } // end namespace llvm 248 249 /// This does one-way checks to see if Use could theoretically be hoisted above 250 /// MayClobber. This will not check the other way around. 251 /// 252 /// This assumes that, for the purposes of MemorySSA, Use comes directly after 253 /// MayClobber, with no potentially clobbering operations in between them. 254 /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.) 255 static bool areLoadsReorderable(const LoadInst *Use, 256 const LoadInst *MayClobber) { 257 bool VolatileUse = Use->isVolatile(); 258 bool VolatileClobber = MayClobber->isVolatile(); 259 // Volatile operations may never be reordered with other volatile operations. 260 if (VolatileUse && VolatileClobber) 261 return false; 262 // Otherwise, volatile doesn't matter here. From the language reference: 263 // 'optimizers may change the order of volatile operations relative to 264 // non-volatile operations.'" 265 266 // If a load is seq_cst, it cannot be moved above other loads. If its ordering 267 // is weaker, it can be moved above other loads. We just need to be sure that 268 // MayClobber isn't an acquire load, because loads can't be moved above 269 // acquire loads. 270 // 271 // Note that this explicitly *does* allow the free reordering of monotonic (or 272 // weaker) loads of the same address. 273 bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent; 274 bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(), 275 AtomicOrdering::Acquire); 276 return !(SeqCstUse || MayClobberIsAcquire); 277 } 278 279 template <typename AliasAnalysisType> 280 static bool 281 instructionClobbersQuery(const MemoryDef *MD, const MemoryLocation &UseLoc, 282 const Instruction *UseInst, AliasAnalysisType &AA) { 283 Instruction *DefInst = MD->getMemoryInst(); 284 assert(DefInst && "Defining instruction not actually an instruction"); 285 286 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) { 287 // These intrinsics will show up as affecting memory, but they are just 288 // markers, mostly. 289 // 290 // FIXME: We probably don't actually want MemorySSA to model these at all 291 // (including creating MemoryAccesses for them): we just end up inventing 292 // clobbers where they don't really exist at all. Please see D43269 for 293 // context. 294 switch (II->getIntrinsicID()) { 295 case Intrinsic::invariant_start: 296 case Intrinsic::invariant_end: 297 case Intrinsic::assume: 298 case Intrinsic::experimental_noalias_scope_decl: 299 case Intrinsic::pseudoprobe: 300 return false; 301 case Intrinsic::dbg_declare: 302 case Intrinsic::dbg_label: 303 case Intrinsic::dbg_value: 304 llvm_unreachable("debuginfo shouldn't have associated defs!"); 305 default: 306 break; 307 } 308 } 309 310 if (auto *CB = dyn_cast_or_null<CallBase>(UseInst)) { 311 ModRefInfo I = AA.getModRefInfo(DefInst, CB); 312 return isModOrRefSet(I); 313 } 314 315 if (auto *DefLoad = dyn_cast<LoadInst>(DefInst)) 316 if (auto *UseLoad = dyn_cast_or_null<LoadInst>(UseInst)) 317 return !areLoadsReorderable(UseLoad, DefLoad); 318 319 ModRefInfo I = AA.getModRefInfo(DefInst, UseLoc); 320 return isModSet(I); 321 } 322 323 template <typename AliasAnalysisType> 324 static bool instructionClobbersQuery(MemoryDef *MD, const MemoryUseOrDef *MU, 325 const MemoryLocOrCall &UseMLOC, 326 AliasAnalysisType &AA) { 327 // FIXME: This is a temporary hack to allow a single instructionClobbersQuery 328 // to exist while MemoryLocOrCall is pushed through places. 329 if (UseMLOC.IsCall) 330 return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(), 331 AA); 332 return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(), 333 AA); 334 } 335 336 // Return true when MD may alias MU, return false otherwise. 337 bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU, 338 AliasAnalysis &AA) { 339 return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA); 340 } 341 342 namespace { 343 344 struct UpwardsMemoryQuery { 345 // True if our original query started off as a call 346 bool IsCall = false; 347 // The pointer location we started the query with. This will be empty if 348 // IsCall is true. 349 MemoryLocation StartingLoc; 350 // This is the instruction we were querying about. 351 const Instruction *Inst = nullptr; 352 // The MemoryAccess we actually got called with, used to test local domination 353 const MemoryAccess *OriginalAccess = nullptr; 354 bool SkipSelfAccess = false; 355 356 UpwardsMemoryQuery() = default; 357 358 UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access) 359 : IsCall(isa<CallBase>(Inst)), Inst(Inst), OriginalAccess(Access) { 360 if (!IsCall) 361 StartingLoc = MemoryLocation::get(Inst); 362 } 363 }; 364 365 } // end anonymous namespace 366 367 template <typename AliasAnalysisType> 368 static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysisType &AA, 369 const Instruction *I) { 370 // If the memory can't be changed, then loads of the memory can't be 371 // clobbered. 372 if (auto *LI = dyn_cast<LoadInst>(I)) { 373 return I->hasMetadata(LLVMContext::MD_invariant_load) || 374 !isModSet(AA.getModRefInfoMask(MemoryLocation::get(LI))); 375 } 376 return false; 377 } 378 379 /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing 380 /// inbetween `Start` and `ClobberAt` can clobbers `Start`. 381 /// 382 /// This is meant to be as simple and self-contained as possible. Because it 383 /// uses no cache, etc., it can be relatively expensive. 384 /// 385 /// \param Start The MemoryAccess that we want to walk from. 386 /// \param ClobberAt A clobber for Start. 387 /// \param StartLoc The MemoryLocation for Start. 388 /// \param MSSA The MemorySSA instance that Start and ClobberAt belong to. 389 /// \param Query The UpwardsMemoryQuery we used for our search. 390 /// \param AA The AliasAnalysis we used for our search. 391 /// \param AllowImpreciseClobber Always false, unless we do relaxed verify. 392 393 LLVM_ATTRIBUTE_UNUSED static void 394 checkClobberSanity(const MemoryAccess *Start, MemoryAccess *ClobberAt, 395 const MemoryLocation &StartLoc, const MemorySSA &MSSA, 396 const UpwardsMemoryQuery &Query, BatchAAResults &AA, 397 bool AllowImpreciseClobber = false) { 398 assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?"); 399 400 if (MSSA.isLiveOnEntryDef(Start)) { 401 assert(MSSA.isLiveOnEntryDef(ClobberAt) && 402 "liveOnEntry must clobber itself"); 403 return; 404 } 405 406 bool FoundClobber = false; 407 DenseSet<ConstMemoryAccessPair> VisitedPhis; 408 SmallVector<ConstMemoryAccessPair, 8> Worklist; 409 Worklist.emplace_back(Start, StartLoc); 410 // Walk all paths from Start to ClobberAt, while looking for clobbers. If one 411 // is found, complain. 412 while (!Worklist.empty()) { 413 auto MAP = Worklist.pop_back_val(); 414 // All we care about is that nothing from Start to ClobberAt clobbers Start. 415 // We learn nothing from revisiting nodes. 416 if (!VisitedPhis.insert(MAP).second) 417 continue; 418 419 for (const auto *MA : def_chain(MAP.first)) { 420 if (MA == ClobberAt) { 421 if (const auto *MD = dyn_cast<MemoryDef>(MA)) { 422 // instructionClobbersQuery isn't essentially free, so don't use `|=`, 423 // since it won't let us short-circuit. 424 // 425 // Also, note that this can't be hoisted out of the `Worklist` loop, 426 // since MD may only act as a clobber for 1 of N MemoryLocations. 427 FoundClobber = FoundClobber || MSSA.isLiveOnEntryDef(MD); 428 if (!FoundClobber) { 429 if (instructionClobbersQuery(MD, MAP.second, Query.Inst, AA)) 430 FoundClobber = true; 431 } 432 } 433 break; 434 } 435 436 // We should never hit liveOnEntry, unless it's the clobber. 437 assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?"); 438 439 if (const auto *MD = dyn_cast<MemoryDef>(MA)) { 440 // If Start is a Def, skip self. 441 if (MD == Start) 442 continue; 443 444 assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) && 445 "Found clobber before reaching ClobberAt!"); 446 continue; 447 } 448 449 if (const auto *MU = dyn_cast<MemoryUse>(MA)) { 450 (void)MU; 451 assert (MU == Start && 452 "Can only find use in def chain if Start is a use"); 453 continue; 454 } 455 456 assert(isa<MemoryPhi>(MA)); 457 458 // Add reachable phi predecessors 459 for (auto ItB = upward_defs_begin( 460 {const_cast<MemoryAccess *>(MA), MAP.second}, 461 MSSA.getDomTree()), 462 ItE = upward_defs_end(); 463 ItB != ItE; ++ItB) 464 if (MSSA.getDomTree().isReachableFromEntry(ItB.getPhiArgBlock())) 465 Worklist.emplace_back(*ItB); 466 } 467 } 468 469 // If the verify is done following an optimization, it's possible that 470 // ClobberAt was a conservative clobbering, that we can now infer is not a 471 // true clobbering access. Don't fail the verify if that's the case. 472 // We do have accesses that claim they're optimized, but could be optimized 473 // further. Updating all these can be expensive, so allow it for now (FIXME). 474 if (AllowImpreciseClobber) 475 return; 476 477 // If ClobberAt is a MemoryPhi, we can assume something above it acted as a 478 // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point. 479 assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) && 480 "ClobberAt never acted as a clobber"); 481 } 482 483 namespace { 484 485 /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up 486 /// in one class. 487 class ClobberWalker { 488 /// Save a few bytes by using unsigned instead of size_t. 489 using ListIndex = unsigned; 490 491 /// Represents a span of contiguous MemoryDefs, potentially ending in a 492 /// MemoryPhi. 493 struct DefPath { 494 MemoryLocation Loc; 495 // Note that, because we always walk in reverse, Last will always dominate 496 // First. Also note that First and Last are inclusive. 497 MemoryAccess *First; 498 MemoryAccess *Last; 499 std::optional<ListIndex> Previous; 500 501 DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last, 502 std::optional<ListIndex> Previous) 503 : Loc(Loc), First(First), Last(Last), Previous(Previous) {} 504 505 DefPath(const MemoryLocation &Loc, MemoryAccess *Init, 506 std::optional<ListIndex> Previous) 507 : DefPath(Loc, Init, Init, Previous) {} 508 }; 509 510 const MemorySSA &MSSA; 511 DominatorTree &DT; 512 BatchAAResults *AA; 513 UpwardsMemoryQuery *Query; 514 unsigned *UpwardWalkLimit; 515 516 // Phi optimization bookkeeping: 517 // List of DefPath to process during the current phi optimization walk. 518 SmallVector<DefPath, 32> Paths; 519 // List of visited <Access, Location> pairs; we can skip paths already 520 // visited with the same memory location. 521 DenseSet<ConstMemoryAccessPair> VisitedPhis; 522 523 /// Find the nearest def or phi that `From` can legally be optimized to. 524 const MemoryAccess *getWalkTarget(const MemoryPhi *From) const { 525 assert(From->getNumOperands() && "Phi with no operands?"); 526 527 BasicBlock *BB = From->getBlock(); 528 MemoryAccess *Result = MSSA.getLiveOnEntryDef(); 529 DomTreeNode *Node = DT.getNode(BB); 530 while ((Node = Node->getIDom())) { 531 auto *Defs = MSSA.getBlockDefs(Node->getBlock()); 532 if (Defs) 533 return &*Defs->rbegin(); 534 } 535 return Result; 536 } 537 538 /// Result of calling walkToPhiOrClobber. 539 struct UpwardsWalkResult { 540 /// The "Result" of the walk. Either a clobber, the last thing we walked, or 541 /// both. Include alias info when clobber found. 542 MemoryAccess *Result; 543 bool IsKnownClobber; 544 }; 545 546 /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last. 547 /// This will update Desc.Last as it walks. It will (optionally) also stop at 548 /// StopAt. 549 /// 550 /// This does not test for whether StopAt is a clobber 551 UpwardsWalkResult 552 walkToPhiOrClobber(DefPath &Desc, const MemoryAccess *StopAt = nullptr, 553 const MemoryAccess *SkipStopAt = nullptr) const { 554 assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world"); 555 assert(UpwardWalkLimit && "Need a valid walk limit"); 556 bool LimitAlreadyReached = false; 557 // (*UpwardWalkLimit) may be 0 here, due to the loop in tryOptimizePhi. Set 558 // it to 1. This will not do any alias() calls. It either returns in the 559 // first iteration in the loop below, or is set back to 0 if all def chains 560 // are free of MemoryDefs. 561 if (!*UpwardWalkLimit) { 562 *UpwardWalkLimit = 1; 563 LimitAlreadyReached = true; 564 } 565 566 for (MemoryAccess *Current : def_chain(Desc.Last)) { 567 Desc.Last = Current; 568 if (Current == StopAt || Current == SkipStopAt) 569 return {Current, false}; 570 571 if (auto *MD = dyn_cast<MemoryDef>(Current)) { 572 if (MSSA.isLiveOnEntryDef(MD)) 573 return {MD, true}; 574 575 if (!--*UpwardWalkLimit) 576 return {Current, true}; 577 578 if (instructionClobbersQuery(MD, Desc.Loc, Query->Inst, *AA)) 579 return {MD, true}; 580 } 581 } 582 583 if (LimitAlreadyReached) 584 *UpwardWalkLimit = 0; 585 586 assert(isa<MemoryPhi>(Desc.Last) && 587 "Ended at a non-clobber that's not a phi?"); 588 return {Desc.Last, false}; 589 } 590 591 void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches, 592 ListIndex PriorNode) { 593 auto UpwardDefsBegin = upward_defs_begin({Phi, Paths[PriorNode].Loc}, DT); 594 auto UpwardDefs = make_range(UpwardDefsBegin, upward_defs_end()); 595 for (const MemoryAccessPair &P : UpwardDefs) { 596 PausedSearches.push_back(Paths.size()); 597 Paths.emplace_back(P.second, P.first, PriorNode); 598 } 599 } 600 601 /// Represents a search that terminated after finding a clobber. This clobber 602 /// may or may not be present in the path of defs from LastNode..SearchStart, 603 /// since it may have been retrieved from cache. 604 struct TerminatedPath { 605 MemoryAccess *Clobber; 606 ListIndex LastNode; 607 }; 608 609 /// Get an access that keeps us from optimizing to the given phi. 610 /// 611 /// PausedSearches is an array of indices into the Paths array. Its incoming 612 /// value is the indices of searches that stopped at the last phi optimization 613 /// target. It's left in an unspecified state. 614 /// 615 /// If this returns std::nullopt, NewPaused is a vector of searches that 616 /// terminated at StopWhere. Otherwise, NewPaused is left in an unspecified 617 /// state. 618 std::optional<TerminatedPath> 619 getBlockingAccess(const MemoryAccess *StopWhere, 620 SmallVectorImpl<ListIndex> &PausedSearches, 621 SmallVectorImpl<ListIndex> &NewPaused, 622 SmallVectorImpl<TerminatedPath> &Terminated) { 623 assert(!PausedSearches.empty() && "No searches to continue?"); 624 625 // BFS vs DFS really doesn't make a difference here, so just do a DFS with 626 // PausedSearches as our stack. 627 while (!PausedSearches.empty()) { 628 ListIndex PathIndex = PausedSearches.pop_back_val(); 629 DefPath &Node = Paths[PathIndex]; 630 631 // If we've already visited this path with this MemoryLocation, we don't 632 // need to do so again. 633 // 634 // NOTE: That we just drop these paths on the ground makes caching 635 // behavior sporadic. e.g. given a diamond: 636 // A 637 // B C 638 // D 639 // 640 // ...If we walk D, B, A, C, we'll only cache the result of phi 641 // optimization for A, B, and D; C will be skipped because it dies here. 642 // This arguably isn't the worst thing ever, since: 643 // - We generally query things in a top-down order, so if we got below D 644 // without needing cache entries for {C, MemLoc}, then chances are 645 // that those cache entries would end up ultimately unused. 646 // - We still cache things for A, so C only needs to walk up a bit. 647 // If this behavior becomes problematic, we can fix without a ton of extra 648 // work. 649 if (!VisitedPhis.insert({Node.Last, Node.Loc}).second) 650 continue; 651 652 const MemoryAccess *SkipStopWhere = nullptr; 653 if (Query->SkipSelfAccess && Node.Loc == Query->StartingLoc) { 654 assert(isa<MemoryDef>(Query->OriginalAccess)); 655 SkipStopWhere = Query->OriginalAccess; 656 } 657 658 UpwardsWalkResult Res = walkToPhiOrClobber(Node, 659 /*StopAt=*/StopWhere, 660 /*SkipStopAt=*/SkipStopWhere); 661 if (Res.IsKnownClobber) { 662 assert(Res.Result != StopWhere && Res.Result != SkipStopWhere); 663 664 // If this wasn't a cache hit, we hit a clobber when walking. That's a 665 // failure. 666 TerminatedPath Term{Res.Result, PathIndex}; 667 if (!MSSA.dominates(Res.Result, StopWhere)) 668 return Term; 669 670 // Otherwise, it's a valid thing to potentially optimize to. 671 Terminated.push_back(Term); 672 continue; 673 } 674 675 if (Res.Result == StopWhere || Res.Result == SkipStopWhere) { 676 // We've hit our target. Save this path off for if we want to continue 677 // walking. If we are in the mode of skipping the OriginalAccess, and 678 // we've reached back to the OriginalAccess, do not save path, we've 679 // just looped back to self. 680 if (Res.Result != SkipStopWhere) 681 NewPaused.push_back(PathIndex); 682 continue; 683 } 684 685 assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber"); 686 addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex); 687 } 688 689 return std::nullopt; 690 } 691 692 template <typename T, typename Walker> 693 struct generic_def_path_iterator 694 : public iterator_facade_base<generic_def_path_iterator<T, Walker>, 695 std::forward_iterator_tag, T *> { 696 generic_def_path_iterator() = default; 697 generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {} 698 699 T &operator*() const { return curNode(); } 700 701 generic_def_path_iterator &operator++() { 702 N = curNode().Previous; 703 return *this; 704 } 705 706 bool operator==(const generic_def_path_iterator &O) const { 707 if (N.has_value() != O.N.has_value()) 708 return false; 709 return !N || *N == *O.N; 710 } 711 712 private: 713 T &curNode() const { return W->Paths[*N]; } 714 715 Walker *W = nullptr; 716 std::optional<ListIndex> N; 717 }; 718 719 using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>; 720 using const_def_path_iterator = 721 generic_def_path_iterator<const DefPath, const ClobberWalker>; 722 723 iterator_range<def_path_iterator> def_path(ListIndex From) { 724 return make_range(def_path_iterator(this, From), def_path_iterator()); 725 } 726 727 iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const { 728 return make_range(const_def_path_iterator(this, From), 729 const_def_path_iterator()); 730 } 731 732 struct OptznResult { 733 /// The path that contains our result. 734 TerminatedPath PrimaryClobber; 735 /// The paths that we can legally cache back from, but that aren't 736 /// necessarily the result of the Phi optimization. 737 SmallVector<TerminatedPath, 4> OtherClobbers; 738 }; 739 740 ListIndex defPathIndex(const DefPath &N) const { 741 // The assert looks nicer if we don't need to do &N 742 const DefPath *NP = &N; 743 assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() && 744 "Out of bounds DefPath!"); 745 return NP - &Paths.front(); 746 } 747 748 /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths 749 /// that act as legal clobbers. Note that this won't return *all* clobbers. 750 /// 751 /// Phi optimization algorithm tl;dr: 752 /// - Find the earliest def/phi, A, we can optimize to 753 /// - Find if all paths from the starting memory access ultimately reach A 754 /// - If not, optimization isn't possible. 755 /// - Otherwise, walk from A to another clobber or phi, A'. 756 /// - If A' is a def, we're done. 757 /// - If A' is a phi, try to optimize it. 758 /// 759 /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path 760 /// terminates when a MemoryAccess that clobbers said MemoryLocation is found. 761 OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start, 762 const MemoryLocation &Loc) { 763 assert(Paths.empty() && VisitedPhis.empty() && 764 "Reset the optimization state."); 765 766 Paths.emplace_back(Loc, Start, Phi, std::nullopt); 767 // Stores how many "valid" optimization nodes we had prior to calling 768 // addSearches/getBlockingAccess. Necessary for caching if we had a blocker. 769 auto PriorPathsSize = Paths.size(); 770 771 SmallVector<ListIndex, 16> PausedSearches; 772 SmallVector<ListIndex, 8> NewPaused; 773 SmallVector<TerminatedPath, 4> TerminatedPaths; 774 775 addSearches(Phi, PausedSearches, 0); 776 777 // Moves the TerminatedPath with the "most dominated" Clobber to the end of 778 // Paths. 779 auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) { 780 assert(!Paths.empty() && "Need a path to move"); 781 auto Dom = Paths.begin(); 782 for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I) 783 if (!MSSA.dominates(I->Clobber, Dom->Clobber)) 784 Dom = I; 785 auto Last = Paths.end() - 1; 786 if (Last != Dom) 787 std::iter_swap(Last, Dom); 788 }; 789 790 MemoryPhi *Current = Phi; 791 while (true) { 792 assert(!MSSA.isLiveOnEntryDef(Current) && 793 "liveOnEntry wasn't treated as a clobber?"); 794 795 const auto *Target = getWalkTarget(Current); 796 // If a TerminatedPath doesn't dominate Target, then it wasn't a legal 797 // optimization for the prior phi. 798 assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) { 799 return MSSA.dominates(P.Clobber, Target); 800 })); 801 802 // FIXME: This is broken, because the Blocker may be reported to be 803 // liveOnEntry, and we'll happily wait for that to disappear (read: never) 804 // For the moment, this is fine, since we do nothing with blocker info. 805 if (std::optional<TerminatedPath> Blocker = getBlockingAccess( 806 Target, PausedSearches, NewPaused, TerminatedPaths)) { 807 808 // Find the node we started at. We can't search based on N->Last, since 809 // we may have gone around a loop with a different MemoryLocation. 810 auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) { 811 return defPathIndex(N) < PriorPathsSize; 812 }); 813 assert(Iter != def_path_iterator()); 814 815 DefPath &CurNode = *Iter; 816 assert(CurNode.Last == Current); 817 818 // Two things: 819 // A. We can't reliably cache all of NewPaused back. Consider a case 820 // where we have two paths in NewPaused; one of which can't optimize 821 // above this phi, whereas the other can. If we cache the second path 822 // back, we'll end up with suboptimal cache entries. We can handle 823 // cases like this a bit better when we either try to find all 824 // clobbers that block phi optimization, or when our cache starts 825 // supporting unfinished searches. 826 // B. We can't reliably cache TerminatedPaths back here without doing 827 // extra checks; consider a case like: 828 // T 829 // / \ 830 // D C 831 // \ / 832 // S 833 // Where T is our target, C is a node with a clobber on it, D is a 834 // diamond (with a clobber *only* on the left or right node, N), and 835 // S is our start. Say we walk to D, through the node opposite N 836 // (read: ignoring the clobber), and see a cache entry in the top 837 // node of D. That cache entry gets put into TerminatedPaths. We then 838 // walk up to C (N is later in our worklist), find the clobber, and 839 // quit. If we append TerminatedPaths to OtherClobbers, we'll cache 840 // the bottom part of D to the cached clobber, ignoring the clobber 841 // in N. Again, this problem goes away if we start tracking all 842 // blockers for a given phi optimization. 843 TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)}; 844 return {Result, {}}; 845 } 846 847 // If there's nothing left to search, then all paths led to valid clobbers 848 // that we got from our cache; pick the nearest to the start, and allow 849 // the rest to be cached back. 850 if (NewPaused.empty()) { 851 MoveDominatedPathToEnd(TerminatedPaths); 852 TerminatedPath Result = TerminatedPaths.pop_back_val(); 853 return {Result, std::move(TerminatedPaths)}; 854 } 855 856 MemoryAccess *DefChainEnd = nullptr; 857 SmallVector<TerminatedPath, 4> Clobbers; 858 for (ListIndex Paused : NewPaused) { 859 UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]); 860 if (WR.IsKnownClobber) 861 Clobbers.push_back({WR.Result, Paused}); 862 else 863 // Micro-opt: If we hit the end of the chain, save it. 864 DefChainEnd = WR.Result; 865 } 866 867 if (!TerminatedPaths.empty()) { 868 // If we couldn't find the dominating phi/liveOnEntry in the above loop, 869 // do it now. 870 if (!DefChainEnd) 871 for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target))) 872 DefChainEnd = MA; 873 assert(DefChainEnd && "Failed to find dominating phi/liveOnEntry"); 874 875 // If any of the terminated paths don't dominate the phi we'll try to 876 // optimize, we need to figure out what they are and quit. 877 const BasicBlock *ChainBB = DefChainEnd->getBlock(); 878 for (const TerminatedPath &TP : TerminatedPaths) { 879 // Because we know that DefChainEnd is as "high" as we can go, we 880 // don't need local dominance checks; BB dominance is sufficient. 881 if (DT.dominates(ChainBB, TP.Clobber->getBlock())) 882 Clobbers.push_back(TP); 883 } 884 } 885 886 // If we have clobbers in the def chain, find the one closest to Current 887 // and quit. 888 if (!Clobbers.empty()) { 889 MoveDominatedPathToEnd(Clobbers); 890 TerminatedPath Result = Clobbers.pop_back_val(); 891 return {Result, std::move(Clobbers)}; 892 } 893 894 assert(all_of(NewPaused, 895 [&](ListIndex I) { return Paths[I].Last == DefChainEnd; })); 896 897 // Because liveOnEntry is a clobber, this must be a phi. 898 auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd); 899 900 PriorPathsSize = Paths.size(); 901 PausedSearches.clear(); 902 for (ListIndex I : NewPaused) 903 addSearches(DefChainPhi, PausedSearches, I); 904 NewPaused.clear(); 905 906 Current = DefChainPhi; 907 } 908 } 909 910 void verifyOptResult(const OptznResult &R) const { 911 assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) { 912 return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber); 913 })); 914 } 915 916 void resetPhiOptznState() { 917 Paths.clear(); 918 VisitedPhis.clear(); 919 } 920 921 public: 922 ClobberWalker(const MemorySSA &MSSA, DominatorTree &DT) 923 : MSSA(MSSA), DT(DT) {} 924 925 /// Finds the nearest clobber for the given query, optimizing phis if 926 /// possible. 927 MemoryAccess *findClobber(BatchAAResults &BAA, MemoryAccess *Start, 928 UpwardsMemoryQuery &Q, unsigned &UpWalkLimit) { 929 AA = &BAA; 930 Query = &Q; 931 UpwardWalkLimit = &UpWalkLimit; 932 // Starting limit must be > 0. 933 if (!UpWalkLimit) 934 UpWalkLimit++; 935 936 MemoryAccess *Current = Start; 937 // This walker pretends uses don't exist. If we're handed one, silently grab 938 // its def. (This has the nice side-effect of ensuring we never cache uses) 939 if (auto *MU = dyn_cast<MemoryUse>(Start)) 940 Current = MU->getDefiningAccess(); 941 942 DefPath FirstDesc(Q.StartingLoc, Current, Current, std::nullopt); 943 // Fast path for the overly-common case (no crazy phi optimization 944 // necessary) 945 UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc); 946 MemoryAccess *Result; 947 if (WalkResult.IsKnownClobber) { 948 Result = WalkResult.Result; 949 } else { 950 OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last), 951 Current, Q.StartingLoc); 952 verifyOptResult(OptRes); 953 resetPhiOptznState(); 954 Result = OptRes.PrimaryClobber.Clobber; 955 } 956 957 #ifdef EXPENSIVE_CHECKS 958 if (!Q.SkipSelfAccess && *UpwardWalkLimit > 0) 959 checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, BAA); 960 #endif 961 return Result; 962 } 963 }; 964 965 struct RenamePassData { 966 DomTreeNode *DTN; 967 DomTreeNode::const_iterator ChildIt; 968 MemoryAccess *IncomingVal; 969 970 RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It, 971 MemoryAccess *M) 972 : DTN(D), ChildIt(It), IncomingVal(M) {} 973 974 void swap(RenamePassData &RHS) { 975 std::swap(DTN, RHS.DTN); 976 std::swap(ChildIt, RHS.ChildIt); 977 std::swap(IncomingVal, RHS.IncomingVal); 978 } 979 }; 980 981 } // end anonymous namespace 982 983 namespace llvm { 984 985 class MemorySSA::ClobberWalkerBase { 986 ClobberWalker Walker; 987 MemorySSA *MSSA; 988 989 public: 990 ClobberWalkerBase(MemorySSA *M, DominatorTree *D) : Walker(*M, *D), MSSA(M) {} 991 992 MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *, 993 const MemoryLocation &, 994 BatchAAResults &, unsigned &); 995 // Third argument (bool), defines whether the clobber search should skip the 996 // original queried access. If true, there will be a follow-up query searching 997 // for a clobber access past "self". Note that the Optimized access is not 998 // updated if a new clobber is found by this SkipSelf search. If this 999 // additional query becomes heavily used we may decide to cache the result. 1000 // Walker instantiations will decide how to set the SkipSelf bool. 1001 MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *, BatchAAResults &, 1002 unsigned &, bool, 1003 bool UseInvariantGroup = true); 1004 }; 1005 1006 /// A MemorySSAWalker that does AA walks to disambiguate accesses. It no 1007 /// longer does caching on its own, but the name has been retained for the 1008 /// moment. 1009 class MemorySSA::CachingWalker final : public MemorySSAWalker { 1010 ClobberWalkerBase *Walker; 1011 1012 public: 1013 CachingWalker(MemorySSA *M, ClobberWalkerBase *W) 1014 : MemorySSAWalker(M), Walker(W) {} 1015 ~CachingWalker() override = default; 1016 1017 using MemorySSAWalker::getClobberingMemoryAccess; 1018 1019 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, BatchAAResults &BAA, 1020 unsigned &UWL) { 1021 return Walker->getClobberingMemoryAccessBase(MA, BAA, UWL, false); 1022 } 1023 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, 1024 const MemoryLocation &Loc, 1025 BatchAAResults &BAA, unsigned &UWL) { 1026 return Walker->getClobberingMemoryAccessBase(MA, Loc, BAA, UWL); 1027 } 1028 // This method is not accessible outside of this file. 1029 MemoryAccess *getClobberingMemoryAccessWithoutInvariantGroup( 1030 MemoryAccess *MA, BatchAAResults &BAA, unsigned &UWL) { 1031 return Walker->getClobberingMemoryAccessBase(MA, BAA, UWL, false, false); 1032 } 1033 1034 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, 1035 BatchAAResults &BAA) override { 1036 unsigned UpwardWalkLimit = MaxCheckLimit; 1037 return getClobberingMemoryAccess(MA, BAA, UpwardWalkLimit); 1038 } 1039 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, 1040 const MemoryLocation &Loc, 1041 BatchAAResults &BAA) override { 1042 unsigned UpwardWalkLimit = MaxCheckLimit; 1043 return getClobberingMemoryAccess(MA, Loc, BAA, UpwardWalkLimit); 1044 } 1045 1046 void invalidateInfo(MemoryAccess *MA) override { 1047 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) 1048 MUD->resetOptimized(); 1049 } 1050 }; 1051 1052 class MemorySSA::SkipSelfWalker final : public MemorySSAWalker { 1053 ClobberWalkerBase *Walker; 1054 1055 public: 1056 SkipSelfWalker(MemorySSA *M, ClobberWalkerBase *W) 1057 : MemorySSAWalker(M), Walker(W) {} 1058 ~SkipSelfWalker() override = default; 1059 1060 using MemorySSAWalker::getClobberingMemoryAccess; 1061 1062 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, BatchAAResults &BAA, 1063 unsigned &UWL) { 1064 return Walker->getClobberingMemoryAccessBase(MA, BAA, UWL, true); 1065 } 1066 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, 1067 const MemoryLocation &Loc, 1068 BatchAAResults &BAA, unsigned &UWL) { 1069 return Walker->getClobberingMemoryAccessBase(MA, Loc, BAA, UWL); 1070 } 1071 1072 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, 1073 BatchAAResults &BAA) override { 1074 unsigned UpwardWalkLimit = MaxCheckLimit; 1075 return getClobberingMemoryAccess(MA, BAA, UpwardWalkLimit); 1076 } 1077 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, 1078 const MemoryLocation &Loc, 1079 BatchAAResults &BAA) override { 1080 unsigned UpwardWalkLimit = MaxCheckLimit; 1081 return getClobberingMemoryAccess(MA, Loc, BAA, UpwardWalkLimit); 1082 } 1083 1084 void invalidateInfo(MemoryAccess *MA) override { 1085 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) 1086 MUD->resetOptimized(); 1087 } 1088 }; 1089 1090 } // end namespace llvm 1091 1092 void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal, 1093 bool RenameAllUses) { 1094 // Pass through values to our successors 1095 for (const BasicBlock *S : successors(BB)) { 1096 auto It = PerBlockAccesses.find(S); 1097 // Rename the phi nodes in our successor block 1098 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front())) 1099 continue; 1100 AccessList *Accesses = It->second.get(); 1101 auto *Phi = cast<MemoryPhi>(&Accesses->front()); 1102 if (RenameAllUses) { 1103 bool ReplacementDone = false; 1104 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) 1105 if (Phi->getIncomingBlock(I) == BB) { 1106 Phi->setIncomingValue(I, IncomingVal); 1107 ReplacementDone = true; 1108 } 1109 (void) ReplacementDone; 1110 assert(ReplacementDone && "Incomplete phi during partial rename"); 1111 } else 1112 Phi->addIncoming(IncomingVal, BB); 1113 } 1114 } 1115 1116 /// Rename a single basic block into MemorySSA form. 1117 /// Uses the standard SSA renaming algorithm. 1118 /// \returns The new incoming value. 1119 MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal, 1120 bool RenameAllUses) { 1121 auto It = PerBlockAccesses.find(BB); 1122 // Skip most processing if the list is empty. 1123 if (It != PerBlockAccesses.end()) { 1124 AccessList *Accesses = It->second.get(); 1125 for (MemoryAccess &L : *Accesses) { 1126 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) { 1127 if (MUD->getDefiningAccess() == nullptr || RenameAllUses) 1128 MUD->setDefiningAccess(IncomingVal); 1129 if (isa<MemoryDef>(&L)) 1130 IncomingVal = &L; 1131 } else { 1132 IncomingVal = &L; 1133 } 1134 } 1135 } 1136 return IncomingVal; 1137 } 1138 1139 /// This is the standard SSA renaming algorithm. 1140 /// 1141 /// We walk the dominator tree in preorder, renaming accesses, and then filling 1142 /// in phi nodes in our successors. 1143 void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal, 1144 SmallPtrSetImpl<BasicBlock *> &Visited, 1145 bool SkipVisited, bool RenameAllUses) { 1146 assert(Root && "Trying to rename accesses in an unreachable block"); 1147 1148 SmallVector<RenamePassData, 32> WorkStack; 1149 // Skip everything if we already renamed this block and we are skipping. 1150 // Note: You can't sink this into the if, because we need it to occur 1151 // regardless of whether we skip blocks or not. 1152 bool AlreadyVisited = !Visited.insert(Root->getBlock()).second; 1153 if (SkipVisited && AlreadyVisited) 1154 return; 1155 1156 IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses); 1157 renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses); 1158 WorkStack.push_back({Root, Root->begin(), IncomingVal}); 1159 1160 while (!WorkStack.empty()) { 1161 DomTreeNode *Node = WorkStack.back().DTN; 1162 DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt; 1163 IncomingVal = WorkStack.back().IncomingVal; 1164 1165 if (ChildIt == Node->end()) { 1166 WorkStack.pop_back(); 1167 } else { 1168 DomTreeNode *Child = *ChildIt; 1169 ++WorkStack.back().ChildIt; 1170 BasicBlock *BB = Child->getBlock(); 1171 // Note: You can't sink this into the if, because we need it to occur 1172 // regardless of whether we skip blocks or not. 1173 AlreadyVisited = !Visited.insert(BB).second; 1174 if (SkipVisited && AlreadyVisited) { 1175 // We already visited this during our renaming, which can happen when 1176 // being asked to rename multiple blocks. Figure out the incoming val, 1177 // which is the last def. 1178 // Incoming value can only change if there is a block def, and in that 1179 // case, it's the last block def in the list. 1180 if (auto *BlockDefs = getWritableBlockDefs(BB)) 1181 IncomingVal = &*BlockDefs->rbegin(); 1182 } else 1183 IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses); 1184 renameSuccessorPhis(BB, IncomingVal, RenameAllUses); 1185 WorkStack.push_back({Child, Child->begin(), IncomingVal}); 1186 } 1187 } 1188 } 1189 1190 /// This handles unreachable block accesses by deleting phi nodes in 1191 /// unreachable blocks, and marking all other unreachable MemoryAccess's as 1192 /// being uses of the live on entry definition. 1193 void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) { 1194 assert(!DT->isReachableFromEntry(BB) && 1195 "Reachable block found while handling unreachable blocks"); 1196 1197 // Make sure phi nodes in our reachable successors end up with a 1198 // LiveOnEntryDef for our incoming edge, even though our block is forward 1199 // unreachable. We could just disconnect these blocks from the CFG fully, 1200 // but we do not right now. 1201 for (const BasicBlock *S : successors(BB)) { 1202 if (!DT->isReachableFromEntry(S)) 1203 continue; 1204 auto It = PerBlockAccesses.find(S); 1205 // Rename the phi nodes in our successor block 1206 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front())) 1207 continue; 1208 AccessList *Accesses = It->second.get(); 1209 auto *Phi = cast<MemoryPhi>(&Accesses->front()); 1210 Phi->addIncoming(LiveOnEntryDef.get(), BB); 1211 } 1212 1213 auto It = PerBlockAccesses.find(BB); 1214 if (It == PerBlockAccesses.end()) 1215 return; 1216 1217 auto &Accesses = It->second; 1218 for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) { 1219 auto Next = std::next(AI); 1220 // If we have a phi, just remove it. We are going to replace all 1221 // users with live on entry. 1222 if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI)) 1223 UseOrDef->setDefiningAccess(LiveOnEntryDef.get()); 1224 else 1225 Accesses->erase(AI); 1226 AI = Next; 1227 } 1228 } 1229 1230 MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT) 1231 : DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr), 1232 SkipWalker(nullptr) { 1233 // Build MemorySSA using a batch alias analysis. This reuses the internal 1234 // state that AA collects during an alias()/getModRefInfo() call. This is 1235 // safe because there are no CFG changes while building MemorySSA and can 1236 // significantly reduce the time spent by the compiler in AA, because we will 1237 // make queries about all the instructions in the Function. 1238 assert(AA && "No alias analysis?"); 1239 BatchAAResults BatchAA(*AA); 1240 buildMemorySSA(BatchAA); 1241 // Intentionally leave AA to nullptr while building so we don't accidently 1242 // use non-batch AliasAnalysis. 1243 this->AA = AA; 1244 // Also create the walker here. 1245 getWalker(); 1246 } 1247 1248 MemorySSA::~MemorySSA() { 1249 // Drop all our references 1250 for (const auto &Pair : PerBlockAccesses) 1251 for (MemoryAccess &MA : *Pair.second) 1252 MA.dropAllReferences(); 1253 } 1254 1255 MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) { 1256 auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr)); 1257 1258 if (Res.second) 1259 Res.first->second = std::make_unique<AccessList>(); 1260 return Res.first->second.get(); 1261 } 1262 1263 MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) { 1264 auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr)); 1265 1266 if (Res.second) 1267 Res.first->second = std::make_unique<DefsList>(); 1268 return Res.first->second.get(); 1269 } 1270 1271 namespace llvm { 1272 1273 /// This class is a batch walker of all MemoryUse's in the program, and points 1274 /// their defining access at the thing that actually clobbers them. Because it 1275 /// is a batch walker that touches everything, it does not operate like the 1276 /// other walkers. This walker is basically performing a top-down SSA renaming 1277 /// pass, where the version stack is used as the cache. This enables it to be 1278 /// significantly more time and memory efficient than using the regular walker, 1279 /// which is walking bottom-up. 1280 class MemorySSA::OptimizeUses { 1281 public: 1282 OptimizeUses(MemorySSA *MSSA, CachingWalker *Walker, BatchAAResults *BAA, 1283 DominatorTree *DT) 1284 : MSSA(MSSA), Walker(Walker), AA(BAA), DT(DT) {} 1285 1286 void optimizeUses(); 1287 1288 private: 1289 /// This represents where a given memorylocation is in the stack. 1290 struct MemlocStackInfo { 1291 // This essentially is keeping track of versions of the stack. Whenever 1292 // the stack changes due to pushes or pops, these versions increase. 1293 unsigned long StackEpoch; 1294 unsigned long PopEpoch; 1295 // This is the lower bound of places on the stack to check. It is equal to 1296 // the place the last stack walk ended. 1297 // Note: Correctness depends on this being initialized to 0, which densemap 1298 // does 1299 unsigned long LowerBound; 1300 const BasicBlock *LowerBoundBlock; 1301 // This is where the last walk for this memory location ended. 1302 unsigned long LastKill; 1303 bool LastKillValid; 1304 }; 1305 1306 void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &, 1307 SmallVectorImpl<MemoryAccess *> &, 1308 DenseMap<MemoryLocOrCall, MemlocStackInfo> &); 1309 1310 MemorySSA *MSSA; 1311 CachingWalker *Walker; 1312 BatchAAResults *AA; 1313 DominatorTree *DT; 1314 }; 1315 1316 } // end namespace llvm 1317 1318 /// Optimize the uses in a given block This is basically the SSA renaming 1319 /// algorithm, with one caveat: We are able to use a single stack for all 1320 /// MemoryUses. This is because the set of *possible* reaching MemoryDefs is 1321 /// the same for every MemoryUse. The *actual* clobbering MemoryDef is just 1322 /// going to be some position in that stack of possible ones. 1323 /// 1324 /// We track the stack positions that each MemoryLocation needs 1325 /// to check, and last ended at. This is because we only want to check the 1326 /// things that changed since last time. The same MemoryLocation should 1327 /// get clobbered by the same store (getModRefInfo does not use invariantness or 1328 /// things like this, and if they start, we can modify MemoryLocOrCall to 1329 /// include relevant data) 1330 void MemorySSA::OptimizeUses::optimizeUsesInBlock( 1331 const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch, 1332 SmallVectorImpl<MemoryAccess *> &VersionStack, 1333 DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) { 1334 1335 /// If no accesses, nothing to do. 1336 MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB); 1337 if (Accesses == nullptr) 1338 return; 1339 1340 // Pop everything that doesn't dominate the current block off the stack, 1341 // increment the PopEpoch to account for this. 1342 while (true) { 1343 assert( 1344 !VersionStack.empty() && 1345 "Version stack should have liveOnEntry sentinel dominating everything"); 1346 BasicBlock *BackBlock = VersionStack.back()->getBlock(); 1347 if (DT->dominates(BackBlock, BB)) 1348 break; 1349 while (VersionStack.back()->getBlock() == BackBlock) 1350 VersionStack.pop_back(); 1351 ++PopEpoch; 1352 } 1353 1354 for (MemoryAccess &MA : *Accesses) { 1355 auto *MU = dyn_cast<MemoryUse>(&MA); 1356 if (!MU) { 1357 VersionStack.push_back(&MA); 1358 ++StackEpoch; 1359 continue; 1360 } 1361 1362 if (MU->isOptimized()) 1363 continue; 1364 1365 MemoryLocOrCall UseMLOC(MU); 1366 auto &LocInfo = LocStackInfo[UseMLOC]; 1367 // If the pop epoch changed, it means we've removed stuff from top of 1368 // stack due to changing blocks. We may have to reset the lower bound or 1369 // last kill info. 1370 if (LocInfo.PopEpoch != PopEpoch) { 1371 LocInfo.PopEpoch = PopEpoch; 1372 LocInfo.StackEpoch = StackEpoch; 1373 // If the lower bound was in something that no longer dominates us, we 1374 // have to reset it. 1375 // We can't simply track stack size, because the stack may have had 1376 // pushes/pops in the meantime. 1377 // XXX: This is non-optimal, but only is slower cases with heavily 1378 // branching dominator trees. To get the optimal number of queries would 1379 // be to make lowerbound and lastkill a per-loc stack, and pop it until 1380 // the top of that stack dominates us. This does not seem worth it ATM. 1381 // A much cheaper optimization would be to always explore the deepest 1382 // branch of the dominator tree first. This will guarantee this resets on 1383 // the smallest set of blocks. 1384 if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB && 1385 !DT->dominates(LocInfo.LowerBoundBlock, BB)) { 1386 // Reset the lower bound of things to check. 1387 // TODO: Some day we should be able to reset to last kill, rather than 1388 // 0. 1389 LocInfo.LowerBound = 0; 1390 LocInfo.LowerBoundBlock = VersionStack[0]->getBlock(); 1391 LocInfo.LastKillValid = false; 1392 } 1393 } else if (LocInfo.StackEpoch != StackEpoch) { 1394 // If all that has changed is the StackEpoch, we only have to check the 1395 // new things on the stack, because we've checked everything before. In 1396 // this case, the lower bound of things to check remains the same. 1397 LocInfo.PopEpoch = PopEpoch; 1398 LocInfo.StackEpoch = StackEpoch; 1399 } 1400 if (!LocInfo.LastKillValid) { 1401 LocInfo.LastKill = VersionStack.size() - 1; 1402 LocInfo.LastKillValid = true; 1403 } 1404 1405 // At this point, we should have corrected last kill and LowerBound to be 1406 // in bounds. 1407 assert(LocInfo.LowerBound < VersionStack.size() && 1408 "Lower bound out of range"); 1409 assert(LocInfo.LastKill < VersionStack.size() && 1410 "Last kill info out of range"); 1411 // In any case, the new upper bound is the top of the stack. 1412 unsigned long UpperBound = VersionStack.size() - 1; 1413 1414 if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) { 1415 LLVM_DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " (" 1416 << *(MU->getMemoryInst()) << ")" 1417 << " because there are " 1418 << UpperBound - LocInfo.LowerBound 1419 << " stores to disambiguate\n"); 1420 // Because we did not walk, LastKill is no longer valid, as this may 1421 // have been a kill. 1422 LocInfo.LastKillValid = false; 1423 continue; 1424 } 1425 bool FoundClobberResult = false; 1426 unsigned UpwardWalkLimit = MaxCheckLimit; 1427 while (UpperBound > LocInfo.LowerBound) { 1428 if (isa<MemoryPhi>(VersionStack[UpperBound])) { 1429 // For phis, use the walker, see where we ended up, go there. 1430 // The invariant.group handling in MemorySSA is ad-hoc and doesn't 1431 // support updates, so don't use it to optimize uses. 1432 MemoryAccess *Result = 1433 Walker->getClobberingMemoryAccessWithoutInvariantGroup( 1434 MU, *AA, UpwardWalkLimit); 1435 // We are guaranteed to find it or something is wrong. 1436 while (VersionStack[UpperBound] != Result) { 1437 assert(UpperBound != 0); 1438 --UpperBound; 1439 } 1440 FoundClobberResult = true; 1441 break; 1442 } 1443 1444 MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]); 1445 if (instructionClobbersQuery(MD, MU, UseMLOC, *AA)) { 1446 FoundClobberResult = true; 1447 break; 1448 } 1449 --UpperBound; 1450 } 1451 1452 // At the end of this loop, UpperBound is either a clobber, or lower bound 1453 // PHI walking may cause it to be < LowerBound, and in fact, < LastKill. 1454 if (FoundClobberResult || UpperBound < LocInfo.LastKill) { 1455 MU->setDefiningAccess(VersionStack[UpperBound], true); 1456 LocInfo.LastKill = UpperBound; 1457 } else { 1458 // Otherwise, we checked all the new ones, and now we know we can get to 1459 // LastKill. 1460 MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true); 1461 } 1462 LocInfo.LowerBound = VersionStack.size() - 1; 1463 LocInfo.LowerBoundBlock = BB; 1464 } 1465 } 1466 1467 /// Optimize uses to point to their actual clobbering definitions. 1468 void MemorySSA::OptimizeUses::optimizeUses() { 1469 SmallVector<MemoryAccess *, 16> VersionStack; 1470 DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo; 1471 VersionStack.push_back(MSSA->getLiveOnEntryDef()); 1472 1473 unsigned long StackEpoch = 1; 1474 unsigned long PopEpoch = 1; 1475 // We perform a non-recursive top-down dominator tree walk. 1476 for (const auto *DomNode : depth_first(DT->getRootNode())) 1477 optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack, 1478 LocStackInfo); 1479 } 1480 1481 void MemorySSA::placePHINodes( 1482 const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks) { 1483 // Determine where our MemoryPhi's should go 1484 ForwardIDFCalculator IDFs(*DT); 1485 IDFs.setDefiningBlocks(DefiningBlocks); 1486 SmallVector<BasicBlock *, 32> IDFBlocks; 1487 IDFs.calculate(IDFBlocks); 1488 1489 // Now place MemoryPhi nodes. 1490 for (auto &BB : IDFBlocks) 1491 createMemoryPhi(BB); 1492 } 1493 1494 void MemorySSA::buildMemorySSA(BatchAAResults &BAA) { 1495 // We create an access to represent "live on entry", for things like 1496 // arguments or users of globals, where the memory they use is defined before 1497 // the beginning of the function. We do not actually insert it into the IR. 1498 // We do not define a live on exit for the immediate uses, and thus our 1499 // semantics do *not* imply that something with no immediate uses can simply 1500 // be removed. 1501 BasicBlock &StartingPoint = F.getEntryBlock(); 1502 LiveOnEntryDef.reset(new MemoryDef(F.getContext(), nullptr, nullptr, 1503 &StartingPoint, NextID++)); 1504 1505 // We maintain lists of memory accesses per-block, trading memory for time. We 1506 // could just look up the memory access for every possible instruction in the 1507 // stream. 1508 SmallPtrSet<BasicBlock *, 32> DefiningBlocks; 1509 // Go through each block, figure out where defs occur, and chain together all 1510 // the accesses. 1511 for (BasicBlock &B : F) { 1512 bool InsertIntoDef = false; 1513 AccessList *Accesses = nullptr; 1514 DefsList *Defs = nullptr; 1515 for (Instruction &I : B) { 1516 MemoryUseOrDef *MUD = createNewAccess(&I, &BAA); 1517 if (!MUD) 1518 continue; 1519 1520 if (!Accesses) 1521 Accesses = getOrCreateAccessList(&B); 1522 Accesses->push_back(MUD); 1523 if (isa<MemoryDef>(MUD)) { 1524 InsertIntoDef = true; 1525 if (!Defs) 1526 Defs = getOrCreateDefsList(&B); 1527 Defs->push_back(*MUD); 1528 } 1529 } 1530 if (InsertIntoDef) 1531 DefiningBlocks.insert(&B); 1532 } 1533 placePHINodes(DefiningBlocks); 1534 1535 // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get 1536 // filled in with all blocks. 1537 SmallPtrSet<BasicBlock *, 16> Visited; 1538 renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited); 1539 1540 // Mark the uses in unreachable blocks as live on entry, so that they go 1541 // somewhere. 1542 for (auto &BB : F) 1543 if (!Visited.count(&BB)) 1544 markUnreachableAsLiveOnEntry(&BB); 1545 } 1546 1547 MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); } 1548 1549 MemorySSA::CachingWalker *MemorySSA::getWalkerImpl() { 1550 if (Walker) 1551 return Walker.get(); 1552 1553 if (!WalkerBase) 1554 WalkerBase = std::make_unique<ClobberWalkerBase>(this, DT); 1555 1556 Walker = std::make_unique<CachingWalker>(this, WalkerBase.get()); 1557 return Walker.get(); 1558 } 1559 1560 MemorySSAWalker *MemorySSA::getSkipSelfWalker() { 1561 if (SkipWalker) 1562 return SkipWalker.get(); 1563 1564 if (!WalkerBase) 1565 WalkerBase = std::make_unique<ClobberWalkerBase>(this, DT); 1566 1567 SkipWalker = std::make_unique<SkipSelfWalker>(this, WalkerBase.get()); 1568 return SkipWalker.get(); 1569 } 1570 1571 1572 // This is a helper function used by the creation routines. It places NewAccess 1573 // into the access and defs lists for a given basic block, at the given 1574 // insertion point. 1575 void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess, 1576 const BasicBlock *BB, 1577 InsertionPlace Point) { 1578 auto *Accesses = getOrCreateAccessList(BB); 1579 if (Point == Beginning) { 1580 // If it's a phi node, it goes first, otherwise, it goes after any phi 1581 // nodes. 1582 if (isa<MemoryPhi>(NewAccess)) { 1583 Accesses->push_front(NewAccess); 1584 auto *Defs = getOrCreateDefsList(BB); 1585 Defs->push_front(*NewAccess); 1586 } else { 1587 auto AI = find_if_not( 1588 *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); }); 1589 Accesses->insert(AI, NewAccess); 1590 if (!isa<MemoryUse>(NewAccess)) { 1591 auto *Defs = getOrCreateDefsList(BB); 1592 auto DI = find_if_not( 1593 *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); }); 1594 Defs->insert(DI, *NewAccess); 1595 } 1596 } 1597 } else { 1598 Accesses->push_back(NewAccess); 1599 if (!isa<MemoryUse>(NewAccess)) { 1600 auto *Defs = getOrCreateDefsList(BB); 1601 Defs->push_back(*NewAccess); 1602 } 1603 } 1604 BlockNumberingValid.erase(BB); 1605 } 1606 1607 void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB, 1608 AccessList::iterator InsertPt) { 1609 auto *Accesses = getWritableBlockAccesses(BB); 1610 bool WasEnd = InsertPt == Accesses->end(); 1611 Accesses->insert(AccessList::iterator(InsertPt), What); 1612 if (!isa<MemoryUse>(What)) { 1613 auto *Defs = getOrCreateDefsList(BB); 1614 // If we got asked to insert at the end, we have an easy job, just shove it 1615 // at the end. If we got asked to insert before an existing def, we also get 1616 // an iterator. If we got asked to insert before a use, we have to hunt for 1617 // the next def. 1618 if (WasEnd) { 1619 Defs->push_back(*What); 1620 } else if (isa<MemoryDef>(InsertPt)) { 1621 Defs->insert(InsertPt->getDefsIterator(), *What); 1622 } else { 1623 while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt)) 1624 ++InsertPt; 1625 // Either we found a def, or we are inserting at the end 1626 if (InsertPt == Accesses->end()) 1627 Defs->push_back(*What); 1628 else 1629 Defs->insert(InsertPt->getDefsIterator(), *What); 1630 } 1631 } 1632 BlockNumberingValid.erase(BB); 1633 } 1634 1635 void MemorySSA::prepareForMoveTo(MemoryAccess *What, BasicBlock *BB) { 1636 // Keep it in the lookup tables, remove from the lists 1637 removeFromLists(What, false); 1638 1639 // Note that moving should implicitly invalidate the optimized state of a 1640 // MemoryUse (and Phis can't be optimized). However, it doesn't do so for a 1641 // MemoryDef. 1642 if (auto *MD = dyn_cast<MemoryDef>(What)) 1643 MD->resetOptimized(); 1644 What->setBlock(BB); 1645 } 1646 1647 // Move What before Where in the IR. The end result is that What will belong to 1648 // the right lists and have the right Block set, but will not otherwise be 1649 // correct. It will not have the right defining access, and if it is a def, 1650 // things below it will not properly be updated. 1651 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB, 1652 AccessList::iterator Where) { 1653 prepareForMoveTo(What, BB); 1654 insertIntoListsBefore(What, BB, Where); 1655 } 1656 1657 void MemorySSA::moveTo(MemoryAccess *What, BasicBlock *BB, 1658 InsertionPlace Point) { 1659 if (isa<MemoryPhi>(What)) { 1660 assert(Point == Beginning && 1661 "Can only move a Phi at the beginning of the block"); 1662 // Update lookup table entry 1663 ValueToMemoryAccess.erase(What->getBlock()); 1664 bool Inserted = ValueToMemoryAccess.insert({BB, What}).second; 1665 (void)Inserted; 1666 assert(Inserted && "Cannot move a Phi to a block that already has one"); 1667 } 1668 1669 prepareForMoveTo(What, BB); 1670 insertIntoListsForBlock(What, BB, Point); 1671 } 1672 1673 MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) { 1674 assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB"); 1675 MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++); 1676 // Phi's always are placed at the front of the block. 1677 insertIntoListsForBlock(Phi, BB, Beginning); 1678 ValueToMemoryAccess[BB] = Phi; 1679 return Phi; 1680 } 1681 1682 MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I, 1683 MemoryAccess *Definition, 1684 const MemoryUseOrDef *Template, 1685 bool CreationMustSucceed) { 1686 assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI"); 1687 MemoryUseOrDef *NewAccess = createNewAccess(I, AA, Template); 1688 if (CreationMustSucceed) 1689 assert(NewAccess != nullptr && "Tried to create a memory access for a " 1690 "non-memory touching instruction"); 1691 if (NewAccess) { 1692 assert((!Definition || !isa<MemoryUse>(Definition)) && 1693 "A use cannot be a defining access"); 1694 NewAccess->setDefiningAccess(Definition); 1695 } 1696 return NewAccess; 1697 } 1698 1699 // Return true if the instruction has ordering constraints. 1700 // Note specifically that this only considers stores and loads 1701 // because others are still considered ModRef by getModRefInfo. 1702 static inline bool isOrdered(const Instruction *I) { 1703 if (auto *SI = dyn_cast<StoreInst>(I)) { 1704 if (!SI->isUnordered()) 1705 return true; 1706 } else if (auto *LI = dyn_cast<LoadInst>(I)) { 1707 if (!LI->isUnordered()) 1708 return true; 1709 } 1710 return false; 1711 } 1712 1713 /// Helper function to create new memory accesses 1714 template <typename AliasAnalysisType> 1715 MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I, 1716 AliasAnalysisType *AAP, 1717 const MemoryUseOrDef *Template) { 1718 // The assume intrinsic has a control dependency which we model by claiming 1719 // that it writes arbitrarily. Debuginfo intrinsics may be considered 1720 // clobbers when we have a nonstandard AA pipeline. Ignore these fake memory 1721 // dependencies here. 1722 // FIXME: Replace this special casing with a more accurate modelling of 1723 // assume's control dependency. 1724 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1725 switch (II->getIntrinsicID()) { 1726 default: 1727 break; 1728 case Intrinsic::assume: 1729 case Intrinsic::experimental_noalias_scope_decl: 1730 case Intrinsic::pseudoprobe: 1731 return nullptr; 1732 } 1733 } 1734 1735 // Using a nonstandard AA pipelines might leave us with unexpected modref 1736 // results for I, so add a check to not model instructions that may not read 1737 // from or write to memory. This is necessary for correctness. 1738 if (!I->mayReadFromMemory() && !I->mayWriteToMemory()) 1739 return nullptr; 1740 1741 bool Def, Use; 1742 if (Template) { 1743 Def = isa<MemoryDef>(Template); 1744 Use = isa<MemoryUse>(Template); 1745 #if !defined(NDEBUG) 1746 ModRefInfo ModRef = AAP->getModRefInfo(I, std::nullopt); 1747 bool DefCheck, UseCheck; 1748 DefCheck = isModSet(ModRef) || isOrdered(I); 1749 UseCheck = isRefSet(ModRef); 1750 // Memory accesses should only be reduced and can not be increased since AA 1751 // just might return better results as a result of some transformations. 1752 assert((Def == DefCheck || !DefCheck) && 1753 "Memory accesses should only be reduced"); 1754 if (!Def && Use != UseCheck) { 1755 // New Access should not have more power than template access 1756 assert(!UseCheck && "Invalid template"); 1757 } 1758 #endif 1759 } else { 1760 // Find out what affect this instruction has on memory. 1761 ModRefInfo ModRef = AAP->getModRefInfo(I, std::nullopt); 1762 // The isOrdered check is used to ensure that volatiles end up as defs 1763 // (atomics end up as ModRef right now anyway). Until we separate the 1764 // ordering chain from the memory chain, this enables people to see at least 1765 // some relative ordering to volatiles. Note that getClobberingMemoryAccess 1766 // will still give an answer that bypasses other volatile loads. TODO: 1767 // Separate memory aliasing and ordering into two different chains so that 1768 // we can precisely represent both "what memory will this read/write/is 1769 // clobbered by" and "what instructions can I move this past". 1770 Def = isModSet(ModRef) || isOrdered(I); 1771 Use = isRefSet(ModRef); 1772 } 1773 1774 // It's possible for an instruction to not modify memory at all. During 1775 // construction, we ignore them. 1776 if (!Def && !Use) 1777 return nullptr; 1778 1779 MemoryUseOrDef *MUD; 1780 if (Def) { 1781 MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++); 1782 } else { 1783 MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent()); 1784 if (isUseTriviallyOptimizableToLiveOnEntry(*AAP, I)) { 1785 MemoryAccess *LiveOnEntry = getLiveOnEntryDef(); 1786 MUD->setOptimized(LiveOnEntry); 1787 } 1788 } 1789 ValueToMemoryAccess[I] = MUD; 1790 return MUD; 1791 } 1792 1793 /// Properly remove \p MA from all of MemorySSA's lookup tables. 1794 void MemorySSA::removeFromLookups(MemoryAccess *MA) { 1795 assert(MA->use_empty() && 1796 "Trying to remove memory access that still has uses"); 1797 BlockNumbering.erase(MA); 1798 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) 1799 MUD->setDefiningAccess(nullptr); 1800 // Invalidate our walker's cache if necessary 1801 if (!isa<MemoryUse>(MA)) 1802 getWalker()->invalidateInfo(MA); 1803 1804 Value *MemoryInst; 1805 if (const auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) 1806 MemoryInst = MUD->getMemoryInst(); 1807 else 1808 MemoryInst = MA->getBlock(); 1809 1810 auto VMA = ValueToMemoryAccess.find(MemoryInst); 1811 if (VMA->second == MA) 1812 ValueToMemoryAccess.erase(VMA); 1813 } 1814 1815 /// Properly remove \p MA from all of MemorySSA's lists. 1816 /// 1817 /// Because of the way the intrusive list and use lists work, it is important to 1818 /// do removal in the right order. 1819 /// ShouldDelete defaults to true, and will cause the memory access to also be 1820 /// deleted, not just removed. 1821 void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) { 1822 BasicBlock *BB = MA->getBlock(); 1823 // The access list owns the reference, so we erase it from the non-owning list 1824 // first. 1825 if (!isa<MemoryUse>(MA)) { 1826 auto DefsIt = PerBlockDefs.find(BB); 1827 std::unique_ptr<DefsList> &Defs = DefsIt->second; 1828 Defs->remove(*MA); 1829 if (Defs->empty()) 1830 PerBlockDefs.erase(DefsIt); 1831 } 1832 1833 // The erase call here will delete it. If we don't want it deleted, we call 1834 // remove instead. 1835 auto AccessIt = PerBlockAccesses.find(BB); 1836 std::unique_ptr<AccessList> &Accesses = AccessIt->second; 1837 if (ShouldDelete) 1838 Accesses->erase(MA); 1839 else 1840 Accesses->remove(MA); 1841 1842 if (Accesses->empty()) { 1843 PerBlockAccesses.erase(AccessIt); 1844 BlockNumberingValid.erase(BB); 1845 } 1846 } 1847 1848 void MemorySSA::print(raw_ostream &OS) const { 1849 MemorySSAAnnotatedWriter Writer(this); 1850 F.print(OS, &Writer); 1851 } 1852 1853 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1854 LLVM_DUMP_METHOD void MemorySSA::dump() const { print(dbgs()); } 1855 #endif 1856 1857 void MemorySSA::verifyMemorySSA(VerificationLevel VL) const { 1858 #if !defined(NDEBUG) && defined(EXPENSIVE_CHECKS) 1859 VL = VerificationLevel::Full; 1860 #endif 1861 1862 #ifndef NDEBUG 1863 verifyOrderingDominationAndDefUses(F, VL); 1864 verifyDominationNumbers(F); 1865 if (VL == VerificationLevel::Full) 1866 verifyPrevDefInPhis(F); 1867 #endif 1868 // Previously, the verification used to also verify that the clobberingAccess 1869 // cached by MemorySSA is the same as the clobberingAccess found at a later 1870 // query to AA. This does not hold true in general due to the current fragility 1871 // of BasicAA which has arbitrary caps on the things it analyzes before giving 1872 // up. As a result, transformations that are correct, will lead to BasicAA 1873 // returning different Alias answers before and after that transformation. 1874 // Invalidating MemorySSA is not an option, as the results in BasicAA can be so 1875 // random, in the worst case we'd need to rebuild MemorySSA from scratch after 1876 // every transformation, which defeats the purpose of using it. For such an 1877 // example, see test4 added in D51960. 1878 } 1879 1880 void MemorySSA::verifyPrevDefInPhis(Function &F) const { 1881 for (const BasicBlock &BB : F) { 1882 if (MemoryPhi *Phi = getMemoryAccess(&BB)) { 1883 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) { 1884 auto *Pred = Phi->getIncomingBlock(I); 1885 auto *IncAcc = Phi->getIncomingValue(I); 1886 // If Pred has no unreachable predecessors, get last def looking at 1887 // IDoms. If, while walkings IDoms, any of these has an unreachable 1888 // predecessor, then the incoming def can be any access. 1889 if (auto *DTNode = DT->getNode(Pred)) { 1890 while (DTNode) { 1891 if (auto *DefList = getBlockDefs(DTNode->getBlock())) { 1892 auto *LastAcc = &*(--DefList->end()); 1893 assert(LastAcc == IncAcc && 1894 "Incorrect incoming access into phi."); 1895 (void)IncAcc; 1896 (void)LastAcc; 1897 break; 1898 } 1899 DTNode = DTNode->getIDom(); 1900 } 1901 } else { 1902 // If Pred has unreachable predecessors, but has at least a Def, the 1903 // incoming access can be the last Def in Pred, or it could have been 1904 // optimized to LoE. After an update, though, the LoE may have been 1905 // replaced by another access, so IncAcc may be any access. 1906 // If Pred has unreachable predecessors and no Defs, incoming access 1907 // should be LoE; However, after an update, it may be any access. 1908 } 1909 } 1910 } 1911 } 1912 } 1913 1914 /// Verify that all of the blocks we believe to have valid domination numbers 1915 /// actually have valid domination numbers. 1916 void MemorySSA::verifyDominationNumbers(const Function &F) const { 1917 if (BlockNumberingValid.empty()) 1918 return; 1919 1920 SmallPtrSet<const BasicBlock *, 16> ValidBlocks = BlockNumberingValid; 1921 for (const BasicBlock &BB : F) { 1922 if (!ValidBlocks.count(&BB)) 1923 continue; 1924 1925 ValidBlocks.erase(&BB); 1926 1927 const AccessList *Accesses = getBlockAccesses(&BB); 1928 // It's correct to say an empty block has valid numbering. 1929 if (!Accesses) 1930 continue; 1931 1932 // Block numbering starts at 1. 1933 unsigned long LastNumber = 0; 1934 for (const MemoryAccess &MA : *Accesses) { 1935 auto ThisNumberIter = BlockNumbering.find(&MA); 1936 assert(ThisNumberIter != BlockNumbering.end() && 1937 "MemoryAccess has no domination number in a valid block!"); 1938 1939 unsigned long ThisNumber = ThisNumberIter->second; 1940 assert(ThisNumber > LastNumber && 1941 "Domination numbers should be strictly increasing!"); 1942 (void)LastNumber; 1943 LastNumber = ThisNumber; 1944 } 1945 } 1946 1947 assert(ValidBlocks.empty() && 1948 "All valid BasicBlocks should exist in F -- dangling pointers?"); 1949 } 1950 1951 /// Verify ordering: the order and existence of MemoryAccesses matches the 1952 /// order and existence of memory affecting instructions. 1953 /// Verify domination: each definition dominates all of its uses. 1954 /// Verify def-uses: the immediate use information - walk all the memory 1955 /// accesses and verifying that, for each use, it appears in the appropriate 1956 /// def's use list 1957 void MemorySSA::verifyOrderingDominationAndDefUses(Function &F, 1958 VerificationLevel VL) const { 1959 // Walk all the blocks, comparing what the lookups think and what the access 1960 // lists think, as well as the order in the blocks vs the order in the access 1961 // lists. 1962 SmallVector<MemoryAccess *, 32> ActualAccesses; 1963 SmallVector<MemoryAccess *, 32> ActualDefs; 1964 for (BasicBlock &B : F) { 1965 const AccessList *AL = getBlockAccesses(&B); 1966 const auto *DL = getBlockDefs(&B); 1967 MemoryPhi *Phi = getMemoryAccess(&B); 1968 if (Phi) { 1969 // Verify ordering. 1970 ActualAccesses.push_back(Phi); 1971 ActualDefs.push_back(Phi); 1972 // Verify domination 1973 for (const Use &U : Phi->uses()) { 1974 assert(dominates(Phi, U) && "Memory PHI does not dominate it's uses"); 1975 (void)U; 1976 } 1977 // Verify def-uses for full verify. 1978 if (VL == VerificationLevel::Full) { 1979 assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance( 1980 pred_begin(&B), pred_end(&B))) && 1981 "Incomplete MemoryPhi Node"); 1982 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) { 1983 verifyUseInDefs(Phi->getIncomingValue(I), Phi); 1984 assert(is_contained(predecessors(&B), Phi->getIncomingBlock(I)) && 1985 "Incoming phi block not a block predecessor"); 1986 } 1987 } 1988 } 1989 1990 for (Instruction &I : B) { 1991 MemoryUseOrDef *MA = getMemoryAccess(&I); 1992 assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) && 1993 "We have memory affecting instructions " 1994 "in this block but they are not in the " 1995 "access list or defs list"); 1996 if (MA) { 1997 // Verify ordering. 1998 ActualAccesses.push_back(MA); 1999 if (MemoryAccess *MD = dyn_cast<MemoryDef>(MA)) { 2000 // Verify ordering. 2001 ActualDefs.push_back(MA); 2002 // Verify domination. 2003 for (const Use &U : MD->uses()) { 2004 assert(dominates(MD, U) && 2005 "Memory Def does not dominate it's uses"); 2006 (void)U; 2007 } 2008 } 2009 // Verify def-uses for full verify. 2010 if (VL == VerificationLevel::Full) 2011 verifyUseInDefs(MA->getDefiningAccess(), MA); 2012 } 2013 } 2014 // Either we hit the assert, really have no accesses, or we have both 2015 // accesses and an access list. Same with defs. 2016 if (!AL && !DL) 2017 continue; 2018 // Verify ordering. 2019 assert(AL->size() == ActualAccesses.size() && 2020 "We don't have the same number of accesses in the block as on the " 2021 "access list"); 2022 assert((DL || ActualDefs.size() == 0) && 2023 "Either we should have a defs list, or we should have no defs"); 2024 assert((!DL || DL->size() == ActualDefs.size()) && 2025 "We don't have the same number of defs in the block as on the " 2026 "def list"); 2027 auto ALI = AL->begin(); 2028 auto AAI = ActualAccesses.begin(); 2029 while (ALI != AL->end() && AAI != ActualAccesses.end()) { 2030 assert(&*ALI == *AAI && "Not the same accesses in the same order"); 2031 ++ALI; 2032 ++AAI; 2033 } 2034 ActualAccesses.clear(); 2035 if (DL) { 2036 auto DLI = DL->begin(); 2037 auto ADI = ActualDefs.begin(); 2038 while (DLI != DL->end() && ADI != ActualDefs.end()) { 2039 assert(&*DLI == *ADI && "Not the same defs in the same order"); 2040 ++DLI; 2041 ++ADI; 2042 } 2043 } 2044 ActualDefs.clear(); 2045 } 2046 } 2047 2048 /// Verify the def-use lists in MemorySSA, by verifying that \p Use 2049 /// appears in the use list of \p Def. 2050 void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const { 2051 // The live on entry use may cause us to get a NULL def here 2052 if (!Def) 2053 assert(isLiveOnEntryDef(Use) && 2054 "Null def but use not point to live on entry def"); 2055 else 2056 assert(is_contained(Def->users(), Use) && 2057 "Did not find use in def's use list"); 2058 } 2059 2060 /// Perform a local numbering on blocks so that instruction ordering can be 2061 /// determined in constant time. 2062 /// TODO: We currently just number in order. If we numbered by N, we could 2063 /// allow at least N-1 sequences of insertBefore or insertAfter (and at least 2064 /// log2(N) sequences of mixed before and after) without needing to invalidate 2065 /// the numbering. 2066 void MemorySSA::renumberBlock(const BasicBlock *B) const { 2067 // The pre-increment ensures the numbers really start at 1. 2068 unsigned long CurrentNumber = 0; 2069 const AccessList *AL = getBlockAccesses(B); 2070 assert(AL != nullptr && "Asking to renumber an empty block"); 2071 for (const auto &I : *AL) 2072 BlockNumbering[&I] = ++CurrentNumber; 2073 BlockNumberingValid.insert(B); 2074 } 2075 2076 /// Determine, for two memory accesses in the same block, 2077 /// whether \p Dominator dominates \p Dominatee. 2078 /// \returns True if \p Dominator dominates \p Dominatee. 2079 bool MemorySSA::locallyDominates(const MemoryAccess *Dominator, 2080 const MemoryAccess *Dominatee) const { 2081 const BasicBlock *DominatorBlock = Dominator->getBlock(); 2082 2083 assert((DominatorBlock == Dominatee->getBlock()) && 2084 "Asking for local domination when accesses are in different blocks!"); 2085 // A node dominates itself. 2086 if (Dominatee == Dominator) 2087 return true; 2088 2089 // When Dominatee is defined on function entry, it is not dominated by another 2090 // memory access. 2091 if (isLiveOnEntryDef(Dominatee)) 2092 return false; 2093 2094 // When Dominator is defined on function entry, it dominates the other memory 2095 // access. 2096 if (isLiveOnEntryDef(Dominator)) 2097 return true; 2098 2099 if (!BlockNumberingValid.count(DominatorBlock)) 2100 renumberBlock(DominatorBlock); 2101 2102 unsigned long DominatorNum = BlockNumbering.lookup(Dominator); 2103 // All numbers start with 1 2104 assert(DominatorNum != 0 && "Block was not numbered properly"); 2105 unsigned long DominateeNum = BlockNumbering.lookup(Dominatee); 2106 assert(DominateeNum != 0 && "Block was not numbered properly"); 2107 return DominatorNum < DominateeNum; 2108 } 2109 2110 bool MemorySSA::dominates(const MemoryAccess *Dominator, 2111 const MemoryAccess *Dominatee) const { 2112 if (Dominator == Dominatee) 2113 return true; 2114 2115 if (isLiveOnEntryDef(Dominatee)) 2116 return false; 2117 2118 if (Dominator->getBlock() != Dominatee->getBlock()) 2119 return DT->dominates(Dominator->getBlock(), Dominatee->getBlock()); 2120 return locallyDominates(Dominator, Dominatee); 2121 } 2122 2123 bool MemorySSA::dominates(const MemoryAccess *Dominator, 2124 const Use &Dominatee) const { 2125 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) { 2126 BasicBlock *UseBB = MP->getIncomingBlock(Dominatee); 2127 // The def must dominate the incoming block of the phi. 2128 if (UseBB != Dominator->getBlock()) 2129 return DT->dominates(Dominator->getBlock(), UseBB); 2130 // If the UseBB and the DefBB are the same, compare locally. 2131 return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee)); 2132 } 2133 // If it's not a PHI node use, the normal dominates can already handle it. 2134 return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser())); 2135 } 2136 2137 void MemorySSA::ensureOptimizedUses() { 2138 if (IsOptimized) 2139 return; 2140 2141 BatchAAResults BatchAA(*AA); 2142 ClobberWalkerBase WalkerBase(this, DT); 2143 CachingWalker WalkerLocal(this, &WalkerBase); 2144 OptimizeUses(this, &WalkerLocal, &BatchAA, DT).optimizeUses(); 2145 IsOptimized = true; 2146 } 2147 2148 void MemoryAccess::print(raw_ostream &OS) const { 2149 switch (getValueID()) { 2150 case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS); 2151 case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS); 2152 case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS); 2153 } 2154 llvm_unreachable("invalid value id"); 2155 } 2156 2157 void MemoryDef::print(raw_ostream &OS) const { 2158 MemoryAccess *UO = getDefiningAccess(); 2159 2160 auto printID = [&OS](MemoryAccess *A) { 2161 if (A && A->getID()) 2162 OS << A->getID(); 2163 else 2164 OS << LiveOnEntryStr; 2165 }; 2166 2167 OS << getID() << " = MemoryDef("; 2168 printID(UO); 2169 OS << ")"; 2170 2171 if (isOptimized()) { 2172 OS << "->"; 2173 printID(getOptimized()); 2174 } 2175 } 2176 2177 void MemoryPhi::print(raw_ostream &OS) const { 2178 ListSeparator LS(","); 2179 OS << getID() << " = MemoryPhi("; 2180 for (const auto &Op : operands()) { 2181 BasicBlock *BB = getIncomingBlock(Op); 2182 MemoryAccess *MA = cast<MemoryAccess>(Op); 2183 2184 OS << LS << '{'; 2185 if (BB->hasName()) 2186 OS << BB->getName(); 2187 else 2188 BB->printAsOperand(OS, false); 2189 OS << ','; 2190 if (unsigned ID = MA->getID()) 2191 OS << ID; 2192 else 2193 OS << LiveOnEntryStr; 2194 OS << '}'; 2195 } 2196 OS << ')'; 2197 } 2198 2199 void MemoryUse::print(raw_ostream &OS) const { 2200 MemoryAccess *UO = getDefiningAccess(); 2201 OS << "MemoryUse("; 2202 if (UO && UO->getID()) 2203 OS << UO->getID(); 2204 else 2205 OS << LiveOnEntryStr; 2206 OS << ')'; 2207 } 2208 2209 void MemoryAccess::dump() const { 2210 // Cannot completely remove virtual function even in release mode. 2211 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 2212 print(dbgs()); 2213 dbgs() << "\n"; 2214 #endif 2215 } 2216 2217 class DOTFuncMSSAInfo { 2218 private: 2219 const Function &F; 2220 MemorySSAAnnotatedWriter MSSAWriter; 2221 2222 public: 2223 DOTFuncMSSAInfo(const Function &F, MemorySSA &MSSA) 2224 : F(F), MSSAWriter(&MSSA) {} 2225 2226 const Function *getFunction() { return &F; } 2227 MemorySSAAnnotatedWriter &getWriter() { return MSSAWriter; } 2228 }; 2229 2230 namespace llvm { 2231 2232 template <> 2233 struct GraphTraits<DOTFuncMSSAInfo *> : public GraphTraits<const BasicBlock *> { 2234 static NodeRef getEntryNode(DOTFuncMSSAInfo *CFGInfo) { 2235 return &(CFGInfo->getFunction()->getEntryBlock()); 2236 } 2237 2238 // nodes_iterator/begin/end - Allow iteration over all nodes in the graph 2239 using nodes_iterator = pointer_iterator<Function::const_iterator>; 2240 2241 static nodes_iterator nodes_begin(DOTFuncMSSAInfo *CFGInfo) { 2242 return nodes_iterator(CFGInfo->getFunction()->begin()); 2243 } 2244 2245 static nodes_iterator nodes_end(DOTFuncMSSAInfo *CFGInfo) { 2246 return nodes_iterator(CFGInfo->getFunction()->end()); 2247 } 2248 2249 static size_t size(DOTFuncMSSAInfo *CFGInfo) { 2250 return CFGInfo->getFunction()->size(); 2251 } 2252 }; 2253 2254 template <> 2255 struct DOTGraphTraits<DOTFuncMSSAInfo *> : public DefaultDOTGraphTraits { 2256 2257 DOTGraphTraits(bool IsSimple = false) : DefaultDOTGraphTraits(IsSimple) {} 2258 2259 static std::string getGraphName(DOTFuncMSSAInfo *CFGInfo) { 2260 return "MSSA CFG for '" + CFGInfo->getFunction()->getName().str() + 2261 "' function"; 2262 } 2263 2264 std::string getNodeLabel(const BasicBlock *Node, DOTFuncMSSAInfo *CFGInfo) { 2265 return DOTGraphTraits<DOTFuncInfo *>::getCompleteNodeLabel( 2266 Node, nullptr, 2267 [CFGInfo](raw_string_ostream &OS, const BasicBlock &BB) -> void { 2268 BB.print(OS, &CFGInfo->getWriter(), true, true); 2269 }, 2270 [](std::string &S, unsigned &I, unsigned Idx) -> void { 2271 std::string Str = S.substr(I, Idx - I); 2272 StringRef SR = Str; 2273 if (SR.count(" = MemoryDef(") || SR.count(" = MemoryPhi(") || 2274 SR.count("MemoryUse(")) 2275 return; 2276 DOTGraphTraits<DOTFuncInfo *>::eraseComment(S, I, Idx); 2277 }); 2278 } 2279 2280 static std::string getEdgeSourceLabel(const BasicBlock *Node, 2281 const_succ_iterator I) { 2282 return DOTGraphTraits<DOTFuncInfo *>::getEdgeSourceLabel(Node, I); 2283 } 2284 2285 /// Display the raw branch weights from PGO. 2286 std::string getEdgeAttributes(const BasicBlock *Node, const_succ_iterator I, 2287 DOTFuncMSSAInfo *CFGInfo) { 2288 return ""; 2289 } 2290 2291 std::string getNodeAttributes(const BasicBlock *Node, 2292 DOTFuncMSSAInfo *CFGInfo) { 2293 return getNodeLabel(Node, CFGInfo).find(';') != std::string::npos 2294 ? "style=filled, fillcolor=lightpink" 2295 : ""; 2296 } 2297 }; 2298 2299 } // namespace llvm 2300 2301 AnalysisKey MemorySSAAnalysis::Key; 2302 2303 MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F, 2304 FunctionAnalysisManager &AM) { 2305 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 2306 auto &AA = AM.getResult<AAManager>(F); 2307 return MemorySSAAnalysis::Result(std::make_unique<MemorySSA>(F, &AA, &DT)); 2308 } 2309 2310 bool MemorySSAAnalysis::Result::invalidate( 2311 Function &F, const PreservedAnalyses &PA, 2312 FunctionAnalysisManager::Invalidator &Inv) { 2313 auto PAC = PA.getChecker<MemorySSAAnalysis>(); 2314 return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>()) || 2315 Inv.invalidate<AAManager>(F, PA) || 2316 Inv.invalidate<DominatorTreeAnalysis>(F, PA); 2317 } 2318 2319 PreservedAnalyses MemorySSAPrinterPass::run(Function &F, 2320 FunctionAnalysisManager &AM) { 2321 auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA(); 2322 if (EnsureOptimizedUses) 2323 MSSA.ensureOptimizedUses(); 2324 if (DotCFGMSSA != "") { 2325 DOTFuncMSSAInfo CFGInfo(F, MSSA); 2326 WriteGraph(&CFGInfo, "", false, "MSSA", DotCFGMSSA); 2327 } else { 2328 OS << "MemorySSA for function: " << F.getName() << "\n"; 2329 MSSA.print(OS); 2330 } 2331 2332 return PreservedAnalyses::all(); 2333 } 2334 2335 PreservedAnalyses MemorySSAWalkerPrinterPass::run(Function &F, 2336 FunctionAnalysisManager &AM) { 2337 auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA(); 2338 OS << "MemorySSA (walker) for function: " << F.getName() << "\n"; 2339 MemorySSAWalkerAnnotatedWriter Writer(&MSSA); 2340 F.print(OS, &Writer); 2341 2342 return PreservedAnalyses::all(); 2343 } 2344 2345 PreservedAnalyses MemorySSAVerifierPass::run(Function &F, 2346 FunctionAnalysisManager &AM) { 2347 AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA(); 2348 2349 return PreservedAnalyses::all(); 2350 } 2351 2352 char MemorySSAWrapperPass::ID = 0; 2353 2354 MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) { 2355 initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry()); 2356 } 2357 2358 void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); } 2359 2360 void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 2361 AU.setPreservesAll(); 2362 AU.addRequiredTransitive<DominatorTreeWrapperPass>(); 2363 AU.addRequiredTransitive<AAResultsWrapperPass>(); 2364 } 2365 2366 bool MemorySSAWrapperPass::runOnFunction(Function &F) { 2367 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2368 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); 2369 MSSA.reset(new MemorySSA(F, &AA, &DT)); 2370 return false; 2371 } 2372 2373 void MemorySSAWrapperPass::verifyAnalysis() const { 2374 if (VerifyMemorySSA) 2375 MSSA->verifyMemorySSA(); 2376 } 2377 2378 void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const { 2379 MSSA->print(OS); 2380 } 2381 2382 MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {} 2383 2384 /// Walk the use-def chains starting at \p StartingAccess and find 2385 /// the MemoryAccess that actually clobbers Loc. 2386 /// 2387 /// \returns our clobbering memory access 2388 MemoryAccess *MemorySSA::ClobberWalkerBase::getClobberingMemoryAccessBase( 2389 MemoryAccess *StartingAccess, const MemoryLocation &Loc, 2390 BatchAAResults &BAA, unsigned &UpwardWalkLimit) { 2391 assert(!isa<MemoryUse>(StartingAccess) && "Use cannot be defining access"); 2392 2393 Instruction *I = nullptr; 2394 if (auto *StartingUseOrDef = dyn_cast<MemoryUseOrDef>(StartingAccess)) { 2395 if (MSSA->isLiveOnEntryDef(StartingUseOrDef)) 2396 return StartingUseOrDef; 2397 2398 I = StartingUseOrDef->getMemoryInst(); 2399 2400 // Conservatively, fences are always clobbers, so don't perform the walk if 2401 // we hit a fence. 2402 if (!isa<CallBase>(I) && I->isFenceLike()) 2403 return StartingUseOrDef; 2404 } 2405 2406 UpwardsMemoryQuery Q; 2407 Q.OriginalAccess = StartingAccess; 2408 Q.StartingLoc = Loc; 2409 Q.Inst = nullptr; 2410 Q.IsCall = false; 2411 2412 // Unlike the other function, do not walk to the def of a def, because we are 2413 // handed something we already believe is the clobbering access. 2414 // We never set SkipSelf to true in Q in this method. 2415 MemoryAccess *Clobber = 2416 Walker.findClobber(BAA, StartingAccess, Q, UpwardWalkLimit); 2417 LLVM_DEBUG({ 2418 dbgs() << "Clobber starting at access " << *StartingAccess << "\n"; 2419 if (I) 2420 dbgs() << " for instruction " << *I << "\n"; 2421 dbgs() << " is " << *Clobber << "\n"; 2422 }); 2423 return Clobber; 2424 } 2425 2426 static const Instruction * 2427 getInvariantGroupClobberingInstruction(Instruction &I, DominatorTree &DT) { 2428 if (!I.hasMetadata(LLVMContext::MD_invariant_group) || I.isVolatile()) 2429 return nullptr; 2430 2431 // We consider bitcasts and zero GEPs to be the same pointer value. Start by 2432 // stripping bitcasts and zero GEPs, then we will recursively look at loads 2433 // and stores through bitcasts and zero GEPs. 2434 Value *PointerOperand = getLoadStorePointerOperand(&I)->stripPointerCasts(); 2435 2436 // It's not safe to walk the use list of a global value because function 2437 // passes aren't allowed to look outside their functions. 2438 // FIXME: this could be fixed by filtering instructions from outside of 2439 // current function. 2440 if (isa<Constant>(PointerOperand)) 2441 return nullptr; 2442 2443 // Queue to process all pointers that are equivalent to load operand. 2444 SmallVector<const Value *, 8> PointerUsesQueue; 2445 PointerUsesQueue.push_back(PointerOperand); 2446 2447 const Instruction *MostDominatingInstruction = &I; 2448 2449 // FIXME: This loop is O(n^2) because dominates can be O(n) and in worst case 2450 // we will see all the instructions. It may not matter in practice. If it 2451 // does, we will have to support MemorySSA construction and updates. 2452 while (!PointerUsesQueue.empty()) { 2453 const Value *Ptr = PointerUsesQueue.pop_back_val(); 2454 assert(Ptr && !isa<GlobalValue>(Ptr) && 2455 "Null or GlobalValue should not be inserted"); 2456 2457 for (const User *Us : Ptr->users()) { 2458 auto *U = dyn_cast<Instruction>(Us); 2459 if (!U || U == &I || !DT.dominates(U, MostDominatingInstruction)) 2460 continue; 2461 2462 // Add bitcasts and zero GEPs to queue. 2463 if (isa<BitCastInst>(U)) { 2464 PointerUsesQueue.push_back(U); 2465 continue; 2466 } 2467 if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) { 2468 if (GEP->hasAllZeroIndices()) 2469 PointerUsesQueue.push_back(U); 2470 continue; 2471 } 2472 2473 // If we hit a load/store with an invariant.group metadata and the same 2474 // pointer operand, we can assume that value pointed to by the pointer 2475 // operand didn't change. 2476 if (U->hasMetadata(LLVMContext::MD_invariant_group) && 2477 getLoadStorePointerOperand(U) == Ptr && !U->isVolatile()) { 2478 MostDominatingInstruction = U; 2479 } 2480 } 2481 } 2482 return MostDominatingInstruction == &I ? nullptr : MostDominatingInstruction; 2483 } 2484 2485 MemoryAccess *MemorySSA::ClobberWalkerBase::getClobberingMemoryAccessBase( 2486 MemoryAccess *MA, BatchAAResults &BAA, unsigned &UpwardWalkLimit, 2487 bool SkipSelf, bool UseInvariantGroup) { 2488 auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA); 2489 // If this is a MemoryPhi, we can't do anything. 2490 if (!StartingAccess) 2491 return MA; 2492 2493 if (UseInvariantGroup) { 2494 if (auto *I = getInvariantGroupClobberingInstruction( 2495 *StartingAccess->getMemoryInst(), MSSA->getDomTree())) { 2496 assert(isa<LoadInst>(I) || isa<StoreInst>(I)); 2497 2498 auto *ClobberMA = MSSA->getMemoryAccess(I); 2499 assert(ClobberMA); 2500 if (isa<MemoryUse>(ClobberMA)) 2501 return ClobberMA->getDefiningAccess(); 2502 return ClobberMA; 2503 } 2504 } 2505 2506 bool IsOptimized = false; 2507 2508 // If this is an already optimized use or def, return the optimized result. 2509 // Note: Currently, we store the optimized def result in a separate field, 2510 // since we can't use the defining access. 2511 if (StartingAccess->isOptimized()) { 2512 if (!SkipSelf || !isa<MemoryDef>(StartingAccess)) 2513 return StartingAccess->getOptimized(); 2514 IsOptimized = true; 2515 } 2516 2517 const Instruction *I = StartingAccess->getMemoryInst(); 2518 // We can't sanely do anything with a fence, since they conservatively clobber 2519 // all memory, and have no locations to get pointers from to try to 2520 // disambiguate. 2521 if (!isa<CallBase>(I) && I->isFenceLike()) 2522 return StartingAccess; 2523 2524 UpwardsMemoryQuery Q(I, StartingAccess); 2525 2526 if (isUseTriviallyOptimizableToLiveOnEntry(BAA, I)) { 2527 MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef(); 2528 StartingAccess->setOptimized(LiveOnEntry); 2529 return LiveOnEntry; 2530 } 2531 2532 MemoryAccess *OptimizedAccess; 2533 if (!IsOptimized) { 2534 // Start with the thing we already think clobbers this location 2535 MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess(); 2536 2537 // At this point, DefiningAccess may be the live on entry def. 2538 // If it is, we will not get a better result. 2539 if (MSSA->isLiveOnEntryDef(DefiningAccess)) { 2540 StartingAccess->setOptimized(DefiningAccess); 2541 return DefiningAccess; 2542 } 2543 2544 OptimizedAccess = 2545 Walker.findClobber(BAA, DefiningAccess, Q, UpwardWalkLimit); 2546 StartingAccess->setOptimized(OptimizedAccess); 2547 } else 2548 OptimizedAccess = StartingAccess->getOptimized(); 2549 2550 LLVM_DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is "); 2551 LLVM_DEBUG(dbgs() << *StartingAccess << "\n"); 2552 LLVM_DEBUG(dbgs() << "Optimized Memory SSA clobber for " << *I << " is "); 2553 LLVM_DEBUG(dbgs() << *OptimizedAccess << "\n"); 2554 2555 MemoryAccess *Result; 2556 if (SkipSelf && isa<MemoryPhi>(OptimizedAccess) && 2557 isa<MemoryDef>(StartingAccess) && UpwardWalkLimit) { 2558 assert(isa<MemoryDef>(Q.OriginalAccess)); 2559 Q.SkipSelfAccess = true; 2560 Result = Walker.findClobber(BAA, OptimizedAccess, Q, UpwardWalkLimit); 2561 } else 2562 Result = OptimizedAccess; 2563 2564 LLVM_DEBUG(dbgs() << "Result Memory SSA clobber [SkipSelf = " << SkipSelf); 2565 LLVM_DEBUG(dbgs() << "] for " << *I << " is " << *Result << "\n"); 2566 2567 return Result; 2568 } 2569 2570 MemoryAccess * 2571 DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA, 2572 BatchAAResults &) { 2573 if (auto *Use = dyn_cast<MemoryUseOrDef>(MA)) 2574 return Use->getDefiningAccess(); 2575 return MA; 2576 } 2577 2578 MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess( 2579 MemoryAccess *StartingAccess, const MemoryLocation &, BatchAAResults &) { 2580 if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess)) 2581 return Use->getDefiningAccess(); 2582 return StartingAccess; 2583 } 2584 2585 void MemoryPhi::deleteMe(DerivedUser *Self) { 2586 delete static_cast<MemoryPhi *>(Self); 2587 } 2588 2589 void MemoryDef::deleteMe(DerivedUser *Self) { 2590 delete static_cast<MemoryDef *>(Self); 2591 } 2592 2593 void MemoryUse::deleteMe(DerivedUser *Self) { 2594 delete static_cast<MemoryUse *>(Self); 2595 } 2596 2597 bool upward_defs_iterator::IsGuaranteedLoopInvariant(const Value *Ptr) const { 2598 auto IsGuaranteedLoopInvariantBase = [](const Value *Ptr) { 2599 Ptr = Ptr->stripPointerCasts(); 2600 if (!isa<Instruction>(Ptr)) 2601 return true; 2602 return isa<AllocaInst>(Ptr); 2603 }; 2604 2605 Ptr = Ptr->stripPointerCasts(); 2606 if (auto *I = dyn_cast<Instruction>(Ptr)) { 2607 if (I->getParent()->isEntryBlock()) 2608 return true; 2609 } 2610 if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) { 2611 return IsGuaranteedLoopInvariantBase(GEP->getPointerOperand()) && 2612 GEP->hasAllConstantIndices(); 2613 } 2614 return IsGuaranteedLoopInvariantBase(Ptr); 2615 } 2616