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/None.h" 20 #include "llvm/ADT/Optional.h" 21 #include "llvm/ADT/STLExtras.h" 22 #include "llvm/ADT/SmallPtrSet.h" 23 #include "llvm/ADT/SmallVector.h" 24 #include "llvm/ADT/StringExtras.h" 25 #include "llvm/ADT/iterator.h" 26 #include "llvm/ADT/iterator_range.h" 27 #include "llvm/Analysis/AliasAnalysis.h" 28 #include "llvm/Analysis/CFGPrinter.h" 29 #include "llvm/Analysis/IteratedDominanceFrontier.h" 30 #include "llvm/Analysis/MemoryLocation.h" 31 #include "llvm/Config/llvm-config.h" 32 #include "llvm/IR/AssemblyAnnotationWriter.h" 33 #include "llvm/IR/BasicBlock.h" 34 #include "llvm/IR/Dominators.h" 35 #include "llvm/IR/Function.h" 36 #include "llvm/IR/Instruction.h" 37 #include "llvm/IR/Instructions.h" 38 #include "llvm/IR/IntrinsicInst.h" 39 #include "llvm/IR/LLVMContext.h" 40 #include "llvm/IR/Operator.h" 41 #include "llvm/IR/PassManager.h" 42 #include "llvm/IR/Use.h" 43 #include "llvm/InitializePasses.h" 44 #include "llvm/Pass.h" 45 #include "llvm/Support/AtomicOrdering.h" 46 #include "llvm/Support/Casting.h" 47 #include "llvm/Support/CommandLine.h" 48 #include "llvm/Support/Compiler.h" 49 #include "llvm/Support/Debug.h" 50 #include "llvm/Support/ErrorHandling.h" 51 #include "llvm/Support/FormattedStream.h" 52 #include "llvm/Support/GraphWriter.h" 53 #include "llvm/Support/raw_ostream.h" 54 #include <algorithm> 55 #include <cassert> 56 #include <iterator> 57 #include <memory> 58 #include <utility> 59 60 using namespace llvm; 61 62 #define DEBUG_TYPE "memoryssa" 63 64 static cl::opt<std::string> 65 DotCFGMSSA("dot-cfg-mssa", 66 cl::value_desc("file name for generated dot file"), 67 cl::desc("file name for generated dot file"), cl::init("")); 68 69 INITIALIZE_PASS_BEGIN(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false, 70 true) 71 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 72 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 73 INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false, 74 true) 75 76 INITIALIZE_PASS_BEGIN(MemorySSAPrinterLegacyPass, "print-memoryssa", 77 "Memory SSA Printer", false, false) 78 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) 79 INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa", 80 "Memory SSA Printer", false, false) 81 82 static cl::opt<unsigned> MaxCheckLimit( 83 "memssa-check-limit", cl::Hidden, cl::init(100), 84 cl::desc("The maximum number of stores/phis MemorySSA" 85 "will consider trying to walk past (default = 100)")); 86 87 // Always verify MemorySSA if expensive checking is enabled. 88 #ifdef EXPENSIVE_CHECKS 89 bool llvm::VerifyMemorySSA = true; 90 #else 91 bool llvm::VerifyMemorySSA = false; 92 #endif 93 94 static cl::opt<bool, true> 95 VerifyMemorySSAX("verify-memoryssa", cl::location(VerifyMemorySSA), 96 cl::Hidden, cl::desc("Enable verification of MemorySSA.")); 97 98 const static char LiveOnEntryStr[] = "liveOnEntry"; 99 100 namespace { 101 102 /// An assembly annotator class to print Memory SSA information in 103 /// comments. 104 class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter { 105 const MemorySSA *MSSA; 106 107 public: 108 MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {} 109 110 void emitBasicBlockStartAnnot(const BasicBlock *BB, 111 formatted_raw_ostream &OS) override { 112 if (MemoryAccess *MA = MSSA->getMemoryAccess(BB)) 113 OS << "; " << *MA << "\n"; 114 } 115 116 void emitInstructionAnnot(const Instruction *I, 117 formatted_raw_ostream &OS) override { 118 if (MemoryAccess *MA = MSSA->getMemoryAccess(I)) 119 OS << "; " << *MA << "\n"; 120 } 121 }; 122 123 /// An assembly annotator class to print Memory SSA information in 124 /// comments. 125 class MemorySSAWalkerAnnotatedWriter : public AssemblyAnnotationWriter { 126 MemorySSA *MSSA; 127 MemorySSAWalker *Walker; 128 129 public: 130 MemorySSAWalkerAnnotatedWriter(MemorySSA *M) 131 : MSSA(M), Walker(M->getWalker()) {} 132 133 void emitBasicBlockStartAnnot(const BasicBlock *BB, 134 formatted_raw_ostream &OS) override { 135 if (MemoryAccess *MA = MSSA->getMemoryAccess(BB)) 136 OS << "; " << *MA << "\n"; 137 } 138 139 void emitInstructionAnnot(const Instruction *I, 140 formatted_raw_ostream &OS) override { 141 if (MemoryAccess *MA = MSSA->getMemoryAccess(I)) { 142 MemoryAccess *Clobber = Walker->getClobberingMemoryAccess(MA); 143 OS << "; " << *MA; 144 if (Clobber) { 145 OS << " - clobbered by "; 146 if (MSSA->isLiveOnEntryDef(Clobber)) 147 OS << LiveOnEntryStr; 148 else 149 OS << *Clobber; 150 } 151 OS << "\n"; 152 } 153 } 154 }; 155 156 } // namespace 157 158 namespace { 159 160 /// Our current alias analysis API differentiates heavily between calls and 161 /// non-calls, and functions called on one usually assert on the other. 162 /// This class encapsulates the distinction to simplify other code that wants 163 /// "Memory affecting instructions and related data" to use as a key. 164 /// For example, this class is used as a densemap key in the use optimizer. 165 class MemoryLocOrCall { 166 public: 167 bool IsCall = false; 168 169 MemoryLocOrCall(MemoryUseOrDef *MUD) 170 : MemoryLocOrCall(MUD->getMemoryInst()) {} 171 MemoryLocOrCall(const MemoryUseOrDef *MUD) 172 : MemoryLocOrCall(MUD->getMemoryInst()) {} 173 174 MemoryLocOrCall(Instruction *Inst) { 175 if (auto *C = dyn_cast<CallBase>(Inst)) { 176 IsCall = true; 177 Call = C; 178 } else { 179 IsCall = false; 180 // There is no such thing as a memorylocation for a fence inst, and it is 181 // unique in that regard. 182 if (!isa<FenceInst>(Inst)) 183 Loc = MemoryLocation::get(Inst); 184 } 185 } 186 187 explicit MemoryLocOrCall(const MemoryLocation &Loc) : Loc(Loc) {} 188 189 const CallBase *getCall() const { 190 assert(IsCall); 191 return Call; 192 } 193 194 MemoryLocation getLoc() const { 195 assert(!IsCall); 196 return Loc; 197 } 198 199 bool operator==(const MemoryLocOrCall &Other) const { 200 if (IsCall != Other.IsCall) 201 return false; 202 203 if (!IsCall) 204 return Loc == Other.Loc; 205 206 if (Call->getCalledOperand() != Other.Call->getCalledOperand()) 207 return false; 208 209 return Call->arg_size() == Other.Call->arg_size() && 210 std::equal(Call->arg_begin(), Call->arg_end(), 211 Other.Call->arg_begin()); 212 } 213 214 private: 215 union { 216 const CallBase *Call; 217 MemoryLocation Loc; 218 }; 219 }; 220 221 } // end anonymous namespace 222 223 namespace llvm { 224 225 template <> struct DenseMapInfo<MemoryLocOrCall> { 226 static inline MemoryLocOrCall getEmptyKey() { 227 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey()); 228 } 229 230 static inline MemoryLocOrCall getTombstoneKey() { 231 return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey()); 232 } 233 234 static unsigned getHashValue(const MemoryLocOrCall &MLOC) { 235 if (!MLOC.IsCall) 236 return hash_combine( 237 MLOC.IsCall, 238 DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc())); 239 240 hash_code hash = 241 hash_combine(MLOC.IsCall, DenseMapInfo<const Value *>::getHashValue( 242 MLOC.getCall()->getCalledOperand())); 243 244 for (const Value *Arg : MLOC.getCall()->args()) 245 hash = hash_combine(hash, DenseMapInfo<const Value *>::getHashValue(Arg)); 246 return hash; 247 } 248 249 static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) { 250 return LHS == RHS; 251 } 252 }; 253 254 } // end namespace llvm 255 256 /// This does one-way checks to see if Use could theoretically be hoisted above 257 /// MayClobber. This will not check the other way around. 258 /// 259 /// This assumes that, for the purposes of MemorySSA, Use comes directly after 260 /// MayClobber, with no potentially clobbering operations in between them. 261 /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.) 262 static bool areLoadsReorderable(const LoadInst *Use, 263 const LoadInst *MayClobber) { 264 bool VolatileUse = Use->isVolatile(); 265 bool VolatileClobber = MayClobber->isVolatile(); 266 // Volatile operations may never be reordered with other volatile operations. 267 if (VolatileUse && VolatileClobber) 268 return false; 269 // Otherwise, volatile doesn't matter here. From the language reference: 270 // 'optimizers may change the order of volatile operations relative to 271 // non-volatile operations.'" 272 273 // If a load is seq_cst, it cannot be moved above other loads. If its ordering 274 // is weaker, it can be moved above other loads. We just need to be sure that 275 // MayClobber isn't an acquire load, because loads can't be moved above 276 // acquire loads. 277 // 278 // Note that this explicitly *does* allow the free reordering of monotonic (or 279 // weaker) loads of the same address. 280 bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent; 281 bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(), 282 AtomicOrdering::Acquire); 283 return !(SeqCstUse || MayClobberIsAcquire); 284 } 285 286 namespace { 287 288 struct ClobberAlias { 289 bool IsClobber; 290 Optional<AliasResult> AR; 291 }; 292 293 } // end anonymous namespace 294 295 // Return a pair of {IsClobber (bool), AR (AliasResult)}. It relies on AR being 296 // ignored if IsClobber = false. 297 template <typename AliasAnalysisType> 298 static ClobberAlias 299 instructionClobbersQuery(const MemoryDef *MD, const MemoryLocation &UseLoc, 300 const Instruction *UseInst, AliasAnalysisType &AA) { 301 Instruction *DefInst = MD->getMemoryInst(); 302 assert(DefInst && "Defining instruction not actually an instruction"); 303 Optional<AliasResult> AR; 304 305 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) { 306 // These intrinsics will show up as affecting memory, but they are just 307 // markers, mostly. 308 // 309 // FIXME: We probably don't actually want MemorySSA to model these at all 310 // (including creating MemoryAccesses for them): we just end up inventing 311 // clobbers where they don't really exist at all. Please see D43269 for 312 // context. 313 switch (II->getIntrinsicID()) { 314 case Intrinsic::invariant_start: 315 case Intrinsic::invariant_end: 316 case Intrinsic::assume: 317 case Intrinsic::experimental_noalias_scope_decl: 318 case Intrinsic::pseudoprobe: 319 return {false, AliasResult(AliasResult::NoAlias)}; 320 case Intrinsic::dbg_addr: 321 case Intrinsic::dbg_declare: 322 case Intrinsic::dbg_label: 323 case Intrinsic::dbg_value: 324 llvm_unreachable("debuginfo shouldn't have associated defs!"); 325 default: 326 break; 327 } 328 } 329 330 if (auto *CB = dyn_cast_or_null<CallBase>(UseInst)) { 331 ModRefInfo I = AA.getModRefInfo(DefInst, CB); 332 AR = isMustSet(I) ? AliasResult::MustAlias : AliasResult::MayAlias; 333 return {isModOrRefSet(I), AR}; 334 } 335 336 if (auto *DefLoad = dyn_cast<LoadInst>(DefInst)) 337 if (auto *UseLoad = dyn_cast_or_null<LoadInst>(UseInst)) 338 return {!areLoadsReorderable(UseLoad, DefLoad), 339 AliasResult(AliasResult::MayAlias)}; 340 341 ModRefInfo I = AA.getModRefInfo(DefInst, UseLoc); 342 AR = isMustSet(I) ? AliasResult::MustAlias : AliasResult::MayAlias; 343 return {isModSet(I), AR}; 344 } 345 346 template <typename AliasAnalysisType> 347 static ClobberAlias instructionClobbersQuery(MemoryDef *MD, 348 const MemoryUseOrDef *MU, 349 const MemoryLocOrCall &UseMLOC, 350 AliasAnalysisType &AA) { 351 // FIXME: This is a temporary hack to allow a single instructionClobbersQuery 352 // to exist while MemoryLocOrCall is pushed through places. 353 if (UseMLOC.IsCall) 354 return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(), 355 AA); 356 return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(), 357 AA); 358 } 359 360 // Return true when MD may alias MU, return false otherwise. 361 bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU, 362 AliasAnalysis &AA) { 363 return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA).IsClobber; 364 } 365 366 namespace { 367 368 struct UpwardsMemoryQuery { 369 // True if our original query started off as a call 370 bool IsCall = false; 371 // The pointer location we started the query with. This will be empty if 372 // IsCall is true. 373 MemoryLocation StartingLoc; 374 // This is the instruction we were querying about. 375 const Instruction *Inst = nullptr; 376 // The MemoryAccess we actually got called with, used to test local domination 377 const MemoryAccess *OriginalAccess = nullptr; 378 Optional<AliasResult> AR = AliasResult(AliasResult::MayAlias); 379 bool SkipSelfAccess = false; 380 381 UpwardsMemoryQuery() = default; 382 383 UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access) 384 : IsCall(isa<CallBase>(Inst)), Inst(Inst), OriginalAccess(Access) { 385 if (!IsCall) 386 StartingLoc = MemoryLocation::get(Inst); 387 } 388 }; 389 390 } // end anonymous namespace 391 392 template <typename AliasAnalysisType> 393 static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysisType &AA, 394 const Instruction *I) { 395 // If the memory can't be changed, then loads of the memory can't be 396 // clobbered. 397 if (auto *LI = dyn_cast<LoadInst>(I)) 398 return I->hasMetadata(LLVMContext::MD_invariant_load) || 399 AA.pointsToConstantMemory(MemoryLocation::get(LI)); 400 return false; 401 } 402 403 /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing 404 /// inbetween `Start` and `ClobberAt` can clobbers `Start`. 405 /// 406 /// This is meant to be as simple and self-contained as possible. Because it 407 /// uses no cache, etc., it can be relatively expensive. 408 /// 409 /// \param Start The MemoryAccess that we want to walk from. 410 /// \param ClobberAt A clobber for Start. 411 /// \param StartLoc The MemoryLocation for Start. 412 /// \param MSSA The MemorySSA instance that Start and ClobberAt belong to. 413 /// \param Query The UpwardsMemoryQuery we used for our search. 414 /// \param AA The AliasAnalysis we used for our search. 415 /// \param AllowImpreciseClobber Always false, unless we do relaxed verify. 416 417 template <typename AliasAnalysisType> 418 LLVM_ATTRIBUTE_UNUSED static void 419 checkClobberSanity(const MemoryAccess *Start, MemoryAccess *ClobberAt, 420 const MemoryLocation &StartLoc, const MemorySSA &MSSA, 421 const UpwardsMemoryQuery &Query, AliasAnalysisType &AA, 422 bool AllowImpreciseClobber = false) { 423 assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?"); 424 425 if (MSSA.isLiveOnEntryDef(Start)) { 426 assert(MSSA.isLiveOnEntryDef(ClobberAt) && 427 "liveOnEntry must clobber itself"); 428 return; 429 } 430 431 bool FoundClobber = false; 432 DenseSet<ConstMemoryAccessPair> VisitedPhis; 433 SmallVector<ConstMemoryAccessPair, 8> Worklist; 434 Worklist.emplace_back(Start, StartLoc); 435 // Walk all paths from Start to ClobberAt, while looking for clobbers. If one 436 // is found, complain. 437 while (!Worklist.empty()) { 438 auto MAP = Worklist.pop_back_val(); 439 // All we care about is that nothing from Start to ClobberAt clobbers Start. 440 // We learn nothing from revisiting nodes. 441 if (!VisitedPhis.insert(MAP).second) 442 continue; 443 444 for (const auto *MA : def_chain(MAP.first)) { 445 if (MA == ClobberAt) { 446 if (const auto *MD = dyn_cast<MemoryDef>(MA)) { 447 // instructionClobbersQuery isn't essentially free, so don't use `|=`, 448 // since it won't let us short-circuit. 449 // 450 // Also, note that this can't be hoisted out of the `Worklist` loop, 451 // since MD may only act as a clobber for 1 of N MemoryLocations. 452 FoundClobber = FoundClobber || MSSA.isLiveOnEntryDef(MD); 453 if (!FoundClobber) { 454 ClobberAlias CA = 455 instructionClobbersQuery(MD, MAP.second, Query.Inst, AA); 456 if (CA.IsClobber) { 457 FoundClobber = true; 458 // Not used: CA.AR; 459 } 460 } 461 } 462 break; 463 } 464 465 // We should never hit liveOnEntry, unless it's the clobber. 466 assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?"); 467 468 if (const auto *MD = dyn_cast<MemoryDef>(MA)) { 469 // If Start is a Def, skip self. 470 if (MD == Start) 471 continue; 472 473 assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA) 474 .IsClobber && 475 "Found clobber before reaching ClobberAt!"); 476 continue; 477 } 478 479 if (const auto *MU = dyn_cast<MemoryUse>(MA)) { 480 (void)MU; 481 assert (MU == Start && 482 "Can only find use in def chain if Start is a use"); 483 continue; 484 } 485 486 assert(isa<MemoryPhi>(MA)); 487 488 // Add reachable phi predecessors 489 for (auto ItB = upward_defs_begin( 490 {const_cast<MemoryAccess *>(MA), MAP.second}, 491 MSSA.getDomTree()), 492 ItE = upward_defs_end(); 493 ItB != ItE; ++ItB) 494 if (MSSA.getDomTree().isReachableFromEntry(ItB.getPhiArgBlock())) 495 Worklist.emplace_back(*ItB); 496 } 497 } 498 499 // If the verify is done following an optimization, it's possible that 500 // ClobberAt was a conservative clobbering, that we can now infer is not a 501 // true clobbering access. Don't fail the verify if that's the case. 502 // We do have accesses that claim they're optimized, but could be optimized 503 // further. Updating all these can be expensive, so allow it for now (FIXME). 504 if (AllowImpreciseClobber) 505 return; 506 507 // If ClobberAt is a MemoryPhi, we can assume something above it acted as a 508 // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point. 509 assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) && 510 "ClobberAt never acted as a clobber"); 511 } 512 513 namespace { 514 515 /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up 516 /// in one class. 517 template <class AliasAnalysisType> class ClobberWalker { 518 /// Save a few bytes by using unsigned instead of size_t. 519 using ListIndex = unsigned; 520 521 /// Represents a span of contiguous MemoryDefs, potentially ending in a 522 /// MemoryPhi. 523 struct DefPath { 524 MemoryLocation Loc; 525 // Note that, because we always walk in reverse, Last will always dominate 526 // First. Also note that First and Last are inclusive. 527 MemoryAccess *First; 528 MemoryAccess *Last; 529 Optional<ListIndex> Previous; 530 531 DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last, 532 Optional<ListIndex> Previous) 533 : Loc(Loc), First(First), Last(Last), Previous(Previous) {} 534 535 DefPath(const MemoryLocation &Loc, MemoryAccess *Init, 536 Optional<ListIndex> Previous) 537 : DefPath(Loc, Init, Init, Previous) {} 538 }; 539 540 const MemorySSA &MSSA; 541 AliasAnalysisType &AA; 542 DominatorTree &DT; 543 UpwardsMemoryQuery *Query; 544 unsigned *UpwardWalkLimit; 545 546 // Phi optimization bookkeeping: 547 // List of DefPath to process during the current phi optimization walk. 548 SmallVector<DefPath, 32> Paths; 549 // List of visited <Access, Location> pairs; we can skip paths already 550 // visited with the same memory location. 551 DenseSet<ConstMemoryAccessPair> VisitedPhis; 552 // Record if phi translation has been performed during the current phi 553 // optimization walk, as merging alias results after phi translation can 554 // yield incorrect results. Context in PR46156. 555 bool PerformedPhiTranslation = false; 556 557 /// Find the nearest def or phi that `From` can legally be optimized to. 558 const MemoryAccess *getWalkTarget(const MemoryPhi *From) const { 559 assert(From->getNumOperands() && "Phi with no operands?"); 560 561 BasicBlock *BB = From->getBlock(); 562 MemoryAccess *Result = MSSA.getLiveOnEntryDef(); 563 DomTreeNode *Node = DT.getNode(BB); 564 while ((Node = Node->getIDom())) { 565 auto *Defs = MSSA.getBlockDefs(Node->getBlock()); 566 if (Defs) 567 return &*Defs->rbegin(); 568 } 569 return Result; 570 } 571 572 /// Result of calling walkToPhiOrClobber. 573 struct UpwardsWalkResult { 574 /// The "Result" of the walk. Either a clobber, the last thing we walked, or 575 /// both. Include alias info when clobber found. 576 MemoryAccess *Result; 577 bool IsKnownClobber; 578 Optional<AliasResult> AR; 579 }; 580 581 /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last. 582 /// This will update Desc.Last as it walks. It will (optionally) also stop at 583 /// StopAt. 584 /// 585 /// This does not test for whether StopAt is a clobber 586 UpwardsWalkResult 587 walkToPhiOrClobber(DefPath &Desc, const MemoryAccess *StopAt = nullptr, 588 const MemoryAccess *SkipStopAt = nullptr) const { 589 assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world"); 590 assert(UpwardWalkLimit && "Need a valid walk limit"); 591 bool LimitAlreadyReached = false; 592 // (*UpwardWalkLimit) may be 0 here, due to the loop in tryOptimizePhi. Set 593 // it to 1. This will not do any alias() calls. It either returns in the 594 // first iteration in the loop below, or is set back to 0 if all def chains 595 // are free of MemoryDefs. 596 if (!*UpwardWalkLimit) { 597 *UpwardWalkLimit = 1; 598 LimitAlreadyReached = true; 599 } 600 601 for (MemoryAccess *Current : def_chain(Desc.Last)) { 602 Desc.Last = Current; 603 if (Current == StopAt || Current == SkipStopAt) 604 return {Current, false, AliasResult(AliasResult::MayAlias)}; 605 606 if (auto *MD = dyn_cast<MemoryDef>(Current)) { 607 if (MSSA.isLiveOnEntryDef(MD)) 608 return {MD, true, AliasResult(AliasResult::MustAlias)}; 609 610 if (!--*UpwardWalkLimit) 611 return {Current, true, AliasResult(AliasResult::MayAlias)}; 612 613 ClobberAlias CA = 614 instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA); 615 if (CA.IsClobber) 616 return {MD, true, CA.AR}; 617 } 618 } 619 620 if (LimitAlreadyReached) 621 *UpwardWalkLimit = 0; 622 623 assert(isa<MemoryPhi>(Desc.Last) && 624 "Ended at a non-clobber that's not a phi?"); 625 return {Desc.Last, false, AliasResult(AliasResult::MayAlias)}; 626 } 627 628 void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches, 629 ListIndex PriorNode) { 630 auto UpwardDefsBegin = upward_defs_begin({Phi, Paths[PriorNode].Loc}, DT, 631 &PerformedPhiTranslation); 632 auto UpwardDefs = make_range(UpwardDefsBegin, upward_defs_end()); 633 for (const MemoryAccessPair &P : UpwardDefs) { 634 PausedSearches.push_back(Paths.size()); 635 Paths.emplace_back(P.second, P.first, PriorNode); 636 } 637 } 638 639 /// Represents a search that terminated after finding a clobber. This clobber 640 /// may or may not be present in the path of defs from LastNode..SearchStart, 641 /// since it may have been retrieved from cache. 642 struct TerminatedPath { 643 MemoryAccess *Clobber; 644 ListIndex LastNode; 645 }; 646 647 /// Get an access that keeps us from optimizing to the given phi. 648 /// 649 /// PausedSearches is an array of indices into the Paths array. Its incoming 650 /// value is the indices of searches that stopped at the last phi optimization 651 /// target. It's left in an unspecified state. 652 /// 653 /// If this returns None, NewPaused is a vector of searches that terminated 654 /// at StopWhere. Otherwise, NewPaused is left in an unspecified state. 655 Optional<TerminatedPath> 656 getBlockingAccess(const MemoryAccess *StopWhere, 657 SmallVectorImpl<ListIndex> &PausedSearches, 658 SmallVectorImpl<ListIndex> &NewPaused, 659 SmallVectorImpl<TerminatedPath> &Terminated) { 660 assert(!PausedSearches.empty() && "No searches to continue?"); 661 662 // BFS vs DFS really doesn't make a difference here, so just do a DFS with 663 // PausedSearches as our stack. 664 while (!PausedSearches.empty()) { 665 ListIndex PathIndex = PausedSearches.pop_back_val(); 666 DefPath &Node = Paths[PathIndex]; 667 668 // If we've already visited this path with this MemoryLocation, we don't 669 // need to do so again. 670 // 671 // NOTE: That we just drop these paths on the ground makes caching 672 // behavior sporadic. e.g. given a diamond: 673 // A 674 // B C 675 // D 676 // 677 // ...If we walk D, B, A, C, we'll only cache the result of phi 678 // optimization for A, B, and D; C will be skipped because it dies here. 679 // This arguably isn't the worst thing ever, since: 680 // - We generally query things in a top-down order, so if we got below D 681 // without needing cache entries for {C, MemLoc}, then chances are 682 // that those cache entries would end up ultimately unused. 683 // - We still cache things for A, so C only needs to walk up a bit. 684 // If this behavior becomes problematic, we can fix without a ton of extra 685 // work. 686 if (!VisitedPhis.insert({Node.Last, Node.Loc}).second) { 687 if (PerformedPhiTranslation) { 688 // If visiting this path performed Phi translation, don't continue, 689 // since it may not be correct to merge results from two paths if one 690 // relies on the phi translation. 691 TerminatedPath Term{Node.Last, PathIndex}; 692 return Term; 693 } 694 continue; 695 } 696 697 const MemoryAccess *SkipStopWhere = nullptr; 698 if (Query->SkipSelfAccess && Node.Loc == Query->StartingLoc) { 699 assert(isa<MemoryDef>(Query->OriginalAccess)); 700 SkipStopWhere = Query->OriginalAccess; 701 } 702 703 UpwardsWalkResult Res = walkToPhiOrClobber(Node, 704 /*StopAt=*/StopWhere, 705 /*SkipStopAt=*/SkipStopWhere); 706 if (Res.IsKnownClobber) { 707 assert(Res.Result != StopWhere && Res.Result != SkipStopWhere); 708 709 // If this wasn't a cache hit, we hit a clobber when walking. That's a 710 // failure. 711 TerminatedPath Term{Res.Result, PathIndex}; 712 if (!MSSA.dominates(Res.Result, StopWhere)) 713 return Term; 714 715 // Otherwise, it's a valid thing to potentially optimize to. 716 Terminated.push_back(Term); 717 continue; 718 } 719 720 if (Res.Result == StopWhere || Res.Result == SkipStopWhere) { 721 // We've hit our target. Save this path off for if we want to continue 722 // walking. If we are in the mode of skipping the OriginalAccess, and 723 // we've reached back to the OriginalAccess, do not save path, we've 724 // just looped back to self. 725 if (Res.Result != SkipStopWhere) 726 NewPaused.push_back(PathIndex); 727 continue; 728 } 729 730 assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber"); 731 addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex); 732 } 733 734 return None; 735 } 736 737 template <typename T, typename Walker> 738 struct generic_def_path_iterator 739 : public iterator_facade_base<generic_def_path_iterator<T, Walker>, 740 std::forward_iterator_tag, T *> { 741 generic_def_path_iterator() = default; 742 generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {} 743 744 T &operator*() const { return curNode(); } 745 746 generic_def_path_iterator &operator++() { 747 N = curNode().Previous; 748 return *this; 749 } 750 751 bool operator==(const generic_def_path_iterator &O) const { 752 if (N.has_value() != O.N.has_value()) 753 return false; 754 return !N || *N == *O.N; 755 } 756 757 private: 758 T &curNode() const { return W->Paths[*N]; } 759 760 Walker *W = nullptr; 761 Optional<ListIndex> N = None; 762 }; 763 764 using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>; 765 using const_def_path_iterator = 766 generic_def_path_iterator<const DefPath, const ClobberWalker>; 767 768 iterator_range<def_path_iterator> def_path(ListIndex From) { 769 return make_range(def_path_iterator(this, From), def_path_iterator()); 770 } 771 772 iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const { 773 return make_range(const_def_path_iterator(this, From), 774 const_def_path_iterator()); 775 } 776 777 struct OptznResult { 778 /// The path that contains our result. 779 TerminatedPath PrimaryClobber; 780 /// The paths that we can legally cache back from, but that aren't 781 /// necessarily the result of the Phi optimization. 782 SmallVector<TerminatedPath, 4> OtherClobbers; 783 }; 784 785 ListIndex defPathIndex(const DefPath &N) const { 786 // The assert looks nicer if we don't need to do &N 787 const DefPath *NP = &N; 788 assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() && 789 "Out of bounds DefPath!"); 790 return NP - &Paths.front(); 791 } 792 793 /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths 794 /// that act as legal clobbers. Note that this won't return *all* clobbers. 795 /// 796 /// Phi optimization algorithm tl;dr: 797 /// - Find the earliest def/phi, A, we can optimize to 798 /// - Find if all paths from the starting memory access ultimately reach A 799 /// - If not, optimization isn't possible. 800 /// - Otherwise, walk from A to another clobber or phi, A'. 801 /// - If A' is a def, we're done. 802 /// - If A' is a phi, try to optimize it. 803 /// 804 /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path 805 /// terminates when a MemoryAccess that clobbers said MemoryLocation is found. 806 OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start, 807 const MemoryLocation &Loc) { 808 assert(Paths.empty() && VisitedPhis.empty() && !PerformedPhiTranslation && 809 "Reset the optimization state."); 810 811 Paths.emplace_back(Loc, Start, Phi, None); 812 // Stores how many "valid" optimization nodes we had prior to calling 813 // addSearches/getBlockingAccess. Necessary for caching if we had a blocker. 814 auto PriorPathsSize = Paths.size(); 815 816 SmallVector<ListIndex, 16> PausedSearches; 817 SmallVector<ListIndex, 8> NewPaused; 818 SmallVector<TerminatedPath, 4> TerminatedPaths; 819 820 addSearches(Phi, PausedSearches, 0); 821 822 // Moves the TerminatedPath with the "most dominated" Clobber to the end of 823 // Paths. 824 auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) { 825 assert(!Paths.empty() && "Need a path to move"); 826 auto Dom = Paths.begin(); 827 for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I) 828 if (!MSSA.dominates(I->Clobber, Dom->Clobber)) 829 Dom = I; 830 auto Last = Paths.end() - 1; 831 if (Last != Dom) 832 std::iter_swap(Last, Dom); 833 }; 834 835 MemoryPhi *Current = Phi; 836 while (true) { 837 assert(!MSSA.isLiveOnEntryDef(Current) && 838 "liveOnEntry wasn't treated as a clobber?"); 839 840 const auto *Target = getWalkTarget(Current); 841 // If a TerminatedPath doesn't dominate Target, then it wasn't a legal 842 // optimization for the prior phi. 843 assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) { 844 return MSSA.dominates(P.Clobber, Target); 845 })); 846 847 // FIXME: This is broken, because the Blocker may be reported to be 848 // liveOnEntry, and we'll happily wait for that to disappear (read: never) 849 // For the moment, this is fine, since we do nothing with blocker info. 850 if (Optional<TerminatedPath> Blocker = getBlockingAccess( 851 Target, PausedSearches, NewPaused, TerminatedPaths)) { 852 853 // Find the node we started at. We can't search based on N->Last, since 854 // we may have gone around a loop with a different MemoryLocation. 855 auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) { 856 return defPathIndex(N) < PriorPathsSize; 857 }); 858 assert(Iter != def_path_iterator()); 859 860 DefPath &CurNode = *Iter; 861 assert(CurNode.Last == Current); 862 863 // Two things: 864 // A. We can't reliably cache all of NewPaused back. Consider a case 865 // where we have two paths in NewPaused; one of which can't optimize 866 // above this phi, whereas the other can. If we cache the second path 867 // back, we'll end up with suboptimal cache entries. We can handle 868 // cases like this a bit better when we either try to find all 869 // clobbers that block phi optimization, or when our cache starts 870 // supporting unfinished searches. 871 // B. We can't reliably cache TerminatedPaths back here without doing 872 // extra checks; consider a case like: 873 // T 874 // / \ 875 // D C 876 // \ / 877 // S 878 // Where T is our target, C is a node with a clobber on it, D is a 879 // diamond (with a clobber *only* on the left or right node, N), and 880 // S is our start. Say we walk to D, through the node opposite N 881 // (read: ignoring the clobber), and see a cache entry in the top 882 // node of D. That cache entry gets put into TerminatedPaths. We then 883 // walk up to C (N is later in our worklist), find the clobber, and 884 // quit. If we append TerminatedPaths to OtherClobbers, we'll cache 885 // the bottom part of D to the cached clobber, ignoring the clobber 886 // in N. Again, this problem goes away if we start tracking all 887 // blockers for a given phi optimization. 888 TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)}; 889 return {Result, {}}; 890 } 891 892 // If there's nothing left to search, then all paths led to valid clobbers 893 // that we got from our cache; pick the nearest to the start, and allow 894 // the rest to be cached back. 895 if (NewPaused.empty()) { 896 MoveDominatedPathToEnd(TerminatedPaths); 897 TerminatedPath Result = TerminatedPaths.pop_back_val(); 898 return {Result, std::move(TerminatedPaths)}; 899 } 900 901 MemoryAccess *DefChainEnd = nullptr; 902 SmallVector<TerminatedPath, 4> Clobbers; 903 for (ListIndex Paused : NewPaused) { 904 UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]); 905 if (WR.IsKnownClobber) 906 Clobbers.push_back({WR.Result, Paused}); 907 else 908 // Micro-opt: If we hit the end of the chain, save it. 909 DefChainEnd = WR.Result; 910 } 911 912 if (!TerminatedPaths.empty()) { 913 // If we couldn't find the dominating phi/liveOnEntry in the above loop, 914 // do it now. 915 if (!DefChainEnd) 916 for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target))) 917 DefChainEnd = MA; 918 assert(DefChainEnd && "Failed to find dominating phi/liveOnEntry"); 919 920 // If any of the terminated paths don't dominate the phi we'll try to 921 // optimize, we need to figure out what they are and quit. 922 const BasicBlock *ChainBB = DefChainEnd->getBlock(); 923 for (const TerminatedPath &TP : TerminatedPaths) { 924 // Because we know that DefChainEnd is as "high" as we can go, we 925 // don't need local dominance checks; BB dominance is sufficient. 926 if (DT.dominates(ChainBB, TP.Clobber->getBlock())) 927 Clobbers.push_back(TP); 928 } 929 } 930 931 // If we have clobbers in the def chain, find the one closest to Current 932 // and quit. 933 if (!Clobbers.empty()) { 934 MoveDominatedPathToEnd(Clobbers); 935 TerminatedPath Result = Clobbers.pop_back_val(); 936 return {Result, std::move(Clobbers)}; 937 } 938 939 assert(all_of(NewPaused, 940 [&](ListIndex I) { return Paths[I].Last == DefChainEnd; })); 941 942 // Because liveOnEntry is a clobber, this must be a phi. 943 auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd); 944 945 PriorPathsSize = Paths.size(); 946 PausedSearches.clear(); 947 for (ListIndex I : NewPaused) 948 addSearches(DefChainPhi, PausedSearches, I); 949 NewPaused.clear(); 950 951 Current = DefChainPhi; 952 } 953 } 954 955 void verifyOptResult(const OptznResult &R) const { 956 assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) { 957 return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber); 958 })); 959 } 960 961 void resetPhiOptznState() { 962 Paths.clear(); 963 VisitedPhis.clear(); 964 PerformedPhiTranslation = false; 965 } 966 967 public: 968 ClobberWalker(const MemorySSA &MSSA, AliasAnalysisType &AA, DominatorTree &DT) 969 : MSSA(MSSA), AA(AA), DT(DT) {} 970 971 AliasAnalysisType *getAA() { return &AA; } 972 /// Finds the nearest clobber for the given query, optimizing phis if 973 /// possible. 974 MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q, 975 unsigned &UpWalkLimit) { 976 Query = &Q; 977 UpwardWalkLimit = &UpWalkLimit; 978 // Starting limit must be > 0. 979 if (!UpWalkLimit) 980 UpWalkLimit++; 981 982 MemoryAccess *Current = Start; 983 // This walker pretends uses don't exist. If we're handed one, silently grab 984 // its def. (This has the nice side-effect of ensuring we never cache uses) 985 if (auto *MU = dyn_cast<MemoryUse>(Start)) 986 Current = MU->getDefiningAccess(); 987 988 DefPath FirstDesc(Q.StartingLoc, Current, Current, None); 989 // Fast path for the overly-common case (no crazy phi optimization 990 // necessary) 991 UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc); 992 MemoryAccess *Result; 993 if (WalkResult.IsKnownClobber) { 994 Result = WalkResult.Result; 995 Q.AR = WalkResult.AR; 996 } else { 997 OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last), 998 Current, Q.StartingLoc); 999 verifyOptResult(OptRes); 1000 resetPhiOptznState(); 1001 Result = OptRes.PrimaryClobber.Clobber; 1002 } 1003 1004 #ifdef EXPENSIVE_CHECKS 1005 if (!Q.SkipSelfAccess && *UpwardWalkLimit > 0) 1006 checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA); 1007 #endif 1008 return Result; 1009 } 1010 }; 1011 1012 struct RenamePassData { 1013 DomTreeNode *DTN; 1014 DomTreeNode::const_iterator ChildIt; 1015 MemoryAccess *IncomingVal; 1016 1017 RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It, 1018 MemoryAccess *M) 1019 : DTN(D), ChildIt(It), IncomingVal(M) {} 1020 1021 void swap(RenamePassData &RHS) { 1022 std::swap(DTN, RHS.DTN); 1023 std::swap(ChildIt, RHS.ChildIt); 1024 std::swap(IncomingVal, RHS.IncomingVal); 1025 } 1026 }; 1027 1028 } // end anonymous namespace 1029 1030 namespace llvm { 1031 1032 template <class AliasAnalysisType> class MemorySSA::ClobberWalkerBase { 1033 ClobberWalker<AliasAnalysisType> Walker; 1034 MemorySSA *MSSA; 1035 1036 public: 1037 ClobberWalkerBase(MemorySSA *M, AliasAnalysisType *A, DominatorTree *D) 1038 : Walker(*M, *A, *D), MSSA(M) {} 1039 1040 MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *, 1041 const MemoryLocation &, 1042 unsigned &); 1043 // Third argument (bool), defines whether the clobber search should skip the 1044 // original queried access. If true, there will be a follow-up query searching 1045 // for a clobber access past "self". Note that the Optimized access is not 1046 // updated if a new clobber is found by this SkipSelf search. If this 1047 // additional query becomes heavily used we may decide to cache the result. 1048 // Walker instantiations will decide how to set the SkipSelf bool. 1049 MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *, unsigned &, bool, 1050 bool UseInvariantGroup = true); 1051 }; 1052 1053 /// A MemorySSAWalker that does AA walks to disambiguate accesses. It no 1054 /// longer does caching on its own, but the name has been retained for the 1055 /// moment. 1056 template <class AliasAnalysisType> 1057 class MemorySSA::CachingWalker final : public MemorySSAWalker { 1058 ClobberWalkerBase<AliasAnalysisType> *Walker; 1059 1060 public: 1061 CachingWalker(MemorySSA *M, ClobberWalkerBase<AliasAnalysisType> *W) 1062 : MemorySSAWalker(M), Walker(W) {} 1063 ~CachingWalker() override = default; 1064 1065 using MemorySSAWalker::getClobberingMemoryAccess; 1066 1067 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, unsigned &UWL) { 1068 return Walker->getClobberingMemoryAccessBase(MA, UWL, false); 1069 } 1070 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, 1071 const MemoryLocation &Loc, 1072 unsigned &UWL) { 1073 return Walker->getClobberingMemoryAccessBase(MA, Loc, UWL); 1074 } 1075 // This method is not accessible outside of this file. 1076 MemoryAccess *getClobberingMemoryAccessWithoutInvariantGroup(MemoryAccess *MA, 1077 unsigned &UWL) { 1078 return Walker->getClobberingMemoryAccessBase(MA, UWL, false, false); 1079 } 1080 1081 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA) override { 1082 unsigned UpwardWalkLimit = MaxCheckLimit; 1083 return getClobberingMemoryAccess(MA, UpwardWalkLimit); 1084 } 1085 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, 1086 const MemoryLocation &Loc) override { 1087 unsigned UpwardWalkLimit = MaxCheckLimit; 1088 return getClobberingMemoryAccess(MA, Loc, UpwardWalkLimit); 1089 } 1090 1091 void invalidateInfo(MemoryAccess *MA) override { 1092 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) 1093 MUD->resetOptimized(); 1094 } 1095 }; 1096 1097 template <class AliasAnalysisType> 1098 class MemorySSA::SkipSelfWalker final : public MemorySSAWalker { 1099 ClobberWalkerBase<AliasAnalysisType> *Walker; 1100 1101 public: 1102 SkipSelfWalker(MemorySSA *M, ClobberWalkerBase<AliasAnalysisType> *W) 1103 : MemorySSAWalker(M), Walker(W) {} 1104 ~SkipSelfWalker() override = default; 1105 1106 using MemorySSAWalker::getClobberingMemoryAccess; 1107 1108 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, unsigned &UWL) { 1109 return Walker->getClobberingMemoryAccessBase(MA, UWL, true); 1110 } 1111 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, 1112 const MemoryLocation &Loc, 1113 unsigned &UWL) { 1114 return Walker->getClobberingMemoryAccessBase(MA, Loc, UWL); 1115 } 1116 1117 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA) override { 1118 unsigned UpwardWalkLimit = MaxCheckLimit; 1119 return getClobberingMemoryAccess(MA, UpwardWalkLimit); 1120 } 1121 MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, 1122 const MemoryLocation &Loc) override { 1123 unsigned UpwardWalkLimit = MaxCheckLimit; 1124 return getClobberingMemoryAccess(MA, Loc, UpwardWalkLimit); 1125 } 1126 1127 void invalidateInfo(MemoryAccess *MA) override { 1128 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) 1129 MUD->resetOptimized(); 1130 } 1131 }; 1132 1133 } // end namespace llvm 1134 1135 void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal, 1136 bool RenameAllUses) { 1137 // Pass through values to our successors 1138 for (const BasicBlock *S : successors(BB)) { 1139 auto It = PerBlockAccesses.find(S); 1140 // Rename the phi nodes in our successor block 1141 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front())) 1142 continue; 1143 AccessList *Accesses = It->second.get(); 1144 auto *Phi = cast<MemoryPhi>(&Accesses->front()); 1145 if (RenameAllUses) { 1146 bool ReplacementDone = false; 1147 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) 1148 if (Phi->getIncomingBlock(I) == BB) { 1149 Phi->setIncomingValue(I, IncomingVal); 1150 ReplacementDone = true; 1151 } 1152 (void) ReplacementDone; 1153 assert(ReplacementDone && "Incomplete phi during partial rename"); 1154 } else 1155 Phi->addIncoming(IncomingVal, BB); 1156 } 1157 } 1158 1159 /// Rename a single basic block into MemorySSA form. 1160 /// Uses the standard SSA renaming algorithm. 1161 /// \returns The new incoming value. 1162 MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal, 1163 bool RenameAllUses) { 1164 auto It = PerBlockAccesses.find(BB); 1165 // Skip most processing if the list is empty. 1166 if (It != PerBlockAccesses.end()) { 1167 AccessList *Accesses = It->second.get(); 1168 for (MemoryAccess &L : *Accesses) { 1169 if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) { 1170 if (MUD->getDefiningAccess() == nullptr || RenameAllUses) 1171 MUD->setDefiningAccess(IncomingVal); 1172 if (isa<MemoryDef>(&L)) 1173 IncomingVal = &L; 1174 } else { 1175 IncomingVal = &L; 1176 } 1177 } 1178 } 1179 return IncomingVal; 1180 } 1181 1182 /// This is the standard SSA renaming algorithm. 1183 /// 1184 /// We walk the dominator tree in preorder, renaming accesses, and then filling 1185 /// in phi nodes in our successors. 1186 void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal, 1187 SmallPtrSetImpl<BasicBlock *> &Visited, 1188 bool SkipVisited, bool RenameAllUses) { 1189 assert(Root && "Trying to rename accesses in an unreachable block"); 1190 1191 SmallVector<RenamePassData, 32> WorkStack; 1192 // Skip everything if we already renamed this block and we are skipping. 1193 // Note: You can't sink this into the if, because we need it to occur 1194 // regardless of whether we skip blocks or not. 1195 bool AlreadyVisited = !Visited.insert(Root->getBlock()).second; 1196 if (SkipVisited && AlreadyVisited) 1197 return; 1198 1199 IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses); 1200 renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses); 1201 WorkStack.push_back({Root, Root->begin(), IncomingVal}); 1202 1203 while (!WorkStack.empty()) { 1204 DomTreeNode *Node = WorkStack.back().DTN; 1205 DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt; 1206 IncomingVal = WorkStack.back().IncomingVal; 1207 1208 if (ChildIt == Node->end()) { 1209 WorkStack.pop_back(); 1210 } else { 1211 DomTreeNode *Child = *ChildIt; 1212 ++WorkStack.back().ChildIt; 1213 BasicBlock *BB = Child->getBlock(); 1214 // Note: You can't sink this into the if, because we need it to occur 1215 // regardless of whether we skip blocks or not. 1216 AlreadyVisited = !Visited.insert(BB).second; 1217 if (SkipVisited && AlreadyVisited) { 1218 // We already visited this during our renaming, which can happen when 1219 // being asked to rename multiple blocks. Figure out the incoming val, 1220 // which is the last def. 1221 // Incoming value can only change if there is a block def, and in that 1222 // case, it's the last block def in the list. 1223 if (auto *BlockDefs = getWritableBlockDefs(BB)) 1224 IncomingVal = &*BlockDefs->rbegin(); 1225 } else 1226 IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses); 1227 renameSuccessorPhis(BB, IncomingVal, RenameAllUses); 1228 WorkStack.push_back({Child, Child->begin(), IncomingVal}); 1229 } 1230 } 1231 } 1232 1233 /// This handles unreachable block accesses by deleting phi nodes in 1234 /// unreachable blocks, and marking all other unreachable MemoryAccess's as 1235 /// being uses of the live on entry definition. 1236 void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) { 1237 assert(!DT->isReachableFromEntry(BB) && 1238 "Reachable block found while handling unreachable blocks"); 1239 1240 // Make sure phi nodes in our reachable successors end up with a 1241 // LiveOnEntryDef for our incoming edge, even though our block is forward 1242 // unreachable. We could just disconnect these blocks from the CFG fully, 1243 // but we do not right now. 1244 for (const BasicBlock *S : successors(BB)) { 1245 if (!DT->isReachableFromEntry(S)) 1246 continue; 1247 auto It = PerBlockAccesses.find(S); 1248 // Rename the phi nodes in our successor block 1249 if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front())) 1250 continue; 1251 AccessList *Accesses = It->second.get(); 1252 auto *Phi = cast<MemoryPhi>(&Accesses->front()); 1253 Phi->addIncoming(LiveOnEntryDef.get(), BB); 1254 } 1255 1256 auto It = PerBlockAccesses.find(BB); 1257 if (It == PerBlockAccesses.end()) 1258 return; 1259 1260 auto &Accesses = It->second; 1261 for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) { 1262 auto Next = std::next(AI); 1263 // If we have a phi, just remove it. We are going to replace all 1264 // users with live on entry. 1265 if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI)) 1266 UseOrDef->setDefiningAccess(LiveOnEntryDef.get()); 1267 else 1268 Accesses->erase(AI); 1269 AI = Next; 1270 } 1271 } 1272 1273 MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT) 1274 : DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr), 1275 SkipWalker(nullptr) { 1276 // Build MemorySSA using a batch alias analysis. This reuses the internal 1277 // state that AA collects during an alias()/getModRefInfo() call. This is 1278 // safe because there are no CFG changes while building MemorySSA and can 1279 // significantly reduce the time spent by the compiler in AA, because we will 1280 // make queries about all the instructions in the Function. 1281 assert(AA && "No alias analysis?"); 1282 BatchAAResults BatchAA(*AA); 1283 buildMemorySSA(BatchAA); 1284 // Intentionally leave AA to nullptr while building so we don't accidently 1285 // use non-batch AliasAnalysis. 1286 this->AA = AA; 1287 // Also create the walker here. 1288 getWalker(); 1289 } 1290 1291 MemorySSA::~MemorySSA() { 1292 // Drop all our references 1293 for (const auto &Pair : PerBlockAccesses) 1294 for (MemoryAccess &MA : *Pair.second) 1295 MA.dropAllReferences(); 1296 } 1297 1298 MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) { 1299 auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr)); 1300 1301 if (Res.second) 1302 Res.first->second = std::make_unique<AccessList>(); 1303 return Res.first->second.get(); 1304 } 1305 1306 MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) { 1307 auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr)); 1308 1309 if (Res.second) 1310 Res.first->second = std::make_unique<DefsList>(); 1311 return Res.first->second.get(); 1312 } 1313 1314 namespace llvm { 1315 1316 /// This class is a batch walker of all MemoryUse's in the program, and points 1317 /// their defining access at the thing that actually clobbers them. Because it 1318 /// is a batch walker that touches everything, it does not operate like the 1319 /// other walkers. This walker is basically performing a top-down SSA renaming 1320 /// pass, where the version stack is used as the cache. This enables it to be 1321 /// significantly more time and memory efficient than using the regular walker, 1322 /// which is walking bottom-up. 1323 class MemorySSA::OptimizeUses { 1324 public: 1325 OptimizeUses(MemorySSA *MSSA, CachingWalker<BatchAAResults> *Walker, 1326 BatchAAResults *BAA, DominatorTree *DT) 1327 : MSSA(MSSA), Walker(Walker), AA(BAA), DT(DT) {} 1328 1329 void optimizeUses(); 1330 1331 private: 1332 /// This represents where a given memorylocation is in the stack. 1333 struct MemlocStackInfo { 1334 // This essentially is keeping track of versions of the stack. Whenever 1335 // the stack changes due to pushes or pops, these versions increase. 1336 unsigned long StackEpoch; 1337 unsigned long PopEpoch; 1338 // This is the lower bound of places on the stack to check. It is equal to 1339 // the place the last stack walk ended. 1340 // Note: Correctness depends on this being initialized to 0, which densemap 1341 // does 1342 unsigned long LowerBound; 1343 const BasicBlock *LowerBoundBlock; 1344 // This is where the last walk for this memory location ended. 1345 unsigned long LastKill; 1346 bool LastKillValid; 1347 Optional<AliasResult> AR; 1348 }; 1349 1350 void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &, 1351 SmallVectorImpl<MemoryAccess *> &, 1352 DenseMap<MemoryLocOrCall, MemlocStackInfo> &); 1353 1354 MemorySSA *MSSA; 1355 CachingWalker<BatchAAResults> *Walker; 1356 BatchAAResults *AA; 1357 DominatorTree *DT; 1358 }; 1359 1360 } // end namespace llvm 1361 1362 /// Optimize the uses in a given block This is basically the SSA renaming 1363 /// algorithm, with one caveat: We are able to use a single stack for all 1364 /// MemoryUses. This is because the set of *possible* reaching MemoryDefs is 1365 /// the same for every MemoryUse. The *actual* clobbering MemoryDef is just 1366 /// going to be some position in that stack of possible ones. 1367 /// 1368 /// We track the stack positions that each MemoryLocation needs 1369 /// to check, and last ended at. This is because we only want to check the 1370 /// things that changed since last time. The same MemoryLocation should 1371 /// get clobbered by the same store (getModRefInfo does not use invariantness or 1372 /// things like this, and if they start, we can modify MemoryLocOrCall to 1373 /// include relevant data) 1374 void MemorySSA::OptimizeUses::optimizeUsesInBlock( 1375 const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch, 1376 SmallVectorImpl<MemoryAccess *> &VersionStack, 1377 DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) { 1378 1379 /// If no accesses, nothing to do. 1380 MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB); 1381 if (Accesses == nullptr) 1382 return; 1383 1384 // Pop everything that doesn't dominate the current block off the stack, 1385 // increment the PopEpoch to account for this. 1386 while (true) { 1387 assert( 1388 !VersionStack.empty() && 1389 "Version stack should have liveOnEntry sentinel dominating everything"); 1390 BasicBlock *BackBlock = VersionStack.back()->getBlock(); 1391 if (DT->dominates(BackBlock, BB)) 1392 break; 1393 while (VersionStack.back()->getBlock() == BackBlock) 1394 VersionStack.pop_back(); 1395 ++PopEpoch; 1396 } 1397 1398 for (MemoryAccess &MA : *Accesses) { 1399 auto *MU = dyn_cast<MemoryUse>(&MA); 1400 if (!MU) { 1401 VersionStack.push_back(&MA); 1402 ++StackEpoch; 1403 continue; 1404 } 1405 1406 if (MU->isOptimized()) 1407 continue; 1408 1409 if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) { 1410 MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true, None); 1411 continue; 1412 } 1413 1414 MemoryLocOrCall UseMLOC(MU); 1415 auto &LocInfo = LocStackInfo[UseMLOC]; 1416 // If the pop epoch changed, it means we've removed stuff from top of 1417 // stack due to changing blocks. We may have to reset the lower bound or 1418 // last kill info. 1419 if (LocInfo.PopEpoch != PopEpoch) { 1420 LocInfo.PopEpoch = PopEpoch; 1421 LocInfo.StackEpoch = StackEpoch; 1422 // If the lower bound was in something that no longer dominates us, we 1423 // have to reset it. 1424 // We can't simply track stack size, because the stack may have had 1425 // pushes/pops in the meantime. 1426 // XXX: This is non-optimal, but only is slower cases with heavily 1427 // branching dominator trees. To get the optimal number of queries would 1428 // be to make lowerbound and lastkill a per-loc stack, and pop it until 1429 // the top of that stack dominates us. This does not seem worth it ATM. 1430 // A much cheaper optimization would be to always explore the deepest 1431 // branch of the dominator tree first. This will guarantee this resets on 1432 // the smallest set of blocks. 1433 if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB && 1434 !DT->dominates(LocInfo.LowerBoundBlock, BB)) { 1435 // Reset the lower bound of things to check. 1436 // TODO: Some day we should be able to reset to last kill, rather than 1437 // 0. 1438 LocInfo.LowerBound = 0; 1439 LocInfo.LowerBoundBlock = VersionStack[0]->getBlock(); 1440 LocInfo.LastKillValid = false; 1441 } 1442 } else if (LocInfo.StackEpoch != StackEpoch) { 1443 // If all that has changed is the StackEpoch, we only have to check the 1444 // new things on the stack, because we've checked everything before. In 1445 // this case, the lower bound of things to check remains the same. 1446 LocInfo.PopEpoch = PopEpoch; 1447 LocInfo.StackEpoch = StackEpoch; 1448 } 1449 if (!LocInfo.LastKillValid) { 1450 LocInfo.LastKill = VersionStack.size() - 1; 1451 LocInfo.LastKillValid = true; 1452 LocInfo.AR = AliasResult::MayAlias; 1453 } 1454 1455 // At this point, we should have corrected last kill and LowerBound to be 1456 // in bounds. 1457 assert(LocInfo.LowerBound < VersionStack.size() && 1458 "Lower bound out of range"); 1459 assert(LocInfo.LastKill < VersionStack.size() && 1460 "Last kill info out of range"); 1461 // In any case, the new upper bound is the top of the stack. 1462 unsigned long UpperBound = VersionStack.size() - 1; 1463 1464 if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) { 1465 LLVM_DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " (" 1466 << *(MU->getMemoryInst()) << ")" 1467 << " because there are " 1468 << UpperBound - LocInfo.LowerBound 1469 << " stores to disambiguate\n"); 1470 // Because we did not walk, LastKill is no longer valid, as this may 1471 // have been a kill. 1472 LocInfo.LastKillValid = false; 1473 continue; 1474 } 1475 bool FoundClobberResult = false; 1476 unsigned UpwardWalkLimit = MaxCheckLimit; 1477 while (UpperBound > LocInfo.LowerBound) { 1478 if (isa<MemoryPhi>(VersionStack[UpperBound])) { 1479 // For phis, use the walker, see where we ended up, go there. 1480 // The invariant.group handling in MemorySSA is ad-hoc and doesn't 1481 // support updates, so don't use it to optimize uses. 1482 MemoryAccess *Result = 1483 Walker->getClobberingMemoryAccessWithoutInvariantGroup( 1484 MU, UpwardWalkLimit); 1485 // We are guaranteed to find it or something is wrong. 1486 while (VersionStack[UpperBound] != Result) { 1487 assert(UpperBound != 0); 1488 --UpperBound; 1489 } 1490 FoundClobberResult = true; 1491 break; 1492 } 1493 1494 MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]); 1495 ClobberAlias CA = instructionClobbersQuery(MD, MU, UseMLOC, *AA); 1496 if (CA.IsClobber) { 1497 FoundClobberResult = true; 1498 LocInfo.AR = CA.AR; 1499 break; 1500 } 1501 --UpperBound; 1502 } 1503 1504 // Note: Phis always have AliasResult AR set to MayAlias ATM. 1505 1506 // At the end of this loop, UpperBound is either a clobber, or lower bound 1507 // PHI walking may cause it to be < LowerBound, and in fact, < LastKill. 1508 if (FoundClobberResult || UpperBound < LocInfo.LastKill) { 1509 // We were last killed now by where we got to 1510 if (MSSA->isLiveOnEntryDef(VersionStack[UpperBound])) 1511 LocInfo.AR = None; 1512 MU->setDefiningAccess(VersionStack[UpperBound], true, LocInfo.AR); 1513 LocInfo.LastKill = UpperBound; 1514 } else { 1515 // Otherwise, we checked all the new ones, and now we know we can get to 1516 // LastKill. 1517 MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true, LocInfo.AR); 1518 } 1519 LocInfo.LowerBound = VersionStack.size() - 1; 1520 LocInfo.LowerBoundBlock = BB; 1521 } 1522 } 1523 1524 /// Optimize uses to point to their actual clobbering definitions. 1525 void MemorySSA::OptimizeUses::optimizeUses() { 1526 SmallVector<MemoryAccess *, 16> VersionStack; 1527 DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo; 1528 VersionStack.push_back(MSSA->getLiveOnEntryDef()); 1529 1530 unsigned long StackEpoch = 1; 1531 unsigned long PopEpoch = 1; 1532 // We perform a non-recursive top-down dominator tree walk. 1533 for (const auto *DomNode : depth_first(DT->getRootNode())) 1534 optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack, 1535 LocStackInfo); 1536 } 1537 1538 void MemorySSA::placePHINodes( 1539 const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks) { 1540 // Determine where our MemoryPhi's should go 1541 ForwardIDFCalculator IDFs(*DT); 1542 IDFs.setDefiningBlocks(DefiningBlocks); 1543 SmallVector<BasicBlock *, 32> IDFBlocks; 1544 IDFs.calculate(IDFBlocks); 1545 1546 // Now place MemoryPhi nodes. 1547 for (auto &BB : IDFBlocks) 1548 createMemoryPhi(BB); 1549 } 1550 1551 void MemorySSA::buildMemorySSA(BatchAAResults &BAA) { 1552 // We create an access to represent "live on entry", for things like 1553 // arguments or users of globals, where the memory they use is defined before 1554 // the beginning of the function. We do not actually insert it into the IR. 1555 // We do not define a live on exit for the immediate uses, and thus our 1556 // semantics do *not* imply that something with no immediate uses can simply 1557 // be removed. 1558 BasicBlock &StartingPoint = F.getEntryBlock(); 1559 LiveOnEntryDef.reset(new MemoryDef(F.getContext(), nullptr, nullptr, 1560 &StartingPoint, NextID++)); 1561 1562 // We maintain lists of memory accesses per-block, trading memory for time. We 1563 // could just look up the memory access for every possible instruction in the 1564 // stream. 1565 SmallPtrSet<BasicBlock *, 32> DefiningBlocks; 1566 // Go through each block, figure out where defs occur, and chain together all 1567 // the accesses. 1568 for (BasicBlock &B : F) { 1569 bool InsertIntoDef = false; 1570 AccessList *Accesses = nullptr; 1571 DefsList *Defs = nullptr; 1572 for (Instruction &I : B) { 1573 MemoryUseOrDef *MUD = createNewAccess(&I, &BAA); 1574 if (!MUD) 1575 continue; 1576 1577 if (!Accesses) 1578 Accesses = getOrCreateAccessList(&B); 1579 Accesses->push_back(MUD); 1580 if (isa<MemoryDef>(MUD)) { 1581 InsertIntoDef = true; 1582 if (!Defs) 1583 Defs = getOrCreateDefsList(&B); 1584 Defs->push_back(*MUD); 1585 } 1586 } 1587 if (InsertIntoDef) 1588 DefiningBlocks.insert(&B); 1589 } 1590 placePHINodes(DefiningBlocks); 1591 1592 // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get 1593 // filled in with all blocks. 1594 SmallPtrSet<BasicBlock *, 16> Visited; 1595 renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited); 1596 1597 // Mark the uses in unreachable blocks as live on entry, so that they go 1598 // somewhere. 1599 for (auto &BB : F) 1600 if (!Visited.count(&BB)) 1601 markUnreachableAsLiveOnEntry(&BB); 1602 } 1603 1604 MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); } 1605 1606 MemorySSA::CachingWalker<AliasAnalysis> *MemorySSA::getWalkerImpl() { 1607 if (Walker) 1608 return Walker.get(); 1609 1610 if (!WalkerBase) 1611 WalkerBase = 1612 std::make_unique<ClobberWalkerBase<AliasAnalysis>>(this, AA, DT); 1613 1614 Walker = 1615 std::make_unique<CachingWalker<AliasAnalysis>>(this, WalkerBase.get()); 1616 return Walker.get(); 1617 } 1618 1619 MemorySSAWalker *MemorySSA::getSkipSelfWalker() { 1620 if (SkipWalker) 1621 return SkipWalker.get(); 1622 1623 if (!WalkerBase) 1624 WalkerBase = 1625 std::make_unique<ClobberWalkerBase<AliasAnalysis>>(this, AA, DT); 1626 1627 SkipWalker = 1628 std::make_unique<SkipSelfWalker<AliasAnalysis>>(this, WalkerBase.get()); 1629 return SkipWalker.get(); 1630 } 1631 1632 1633 // This is a helper function used by the creation routines. It places NewAccess 1634 // into the access and defs lists for a given basic block, at the given 1635 // insertion point. 1636 void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess, 1637 const BasicBlock *BB, 1638 InsertionPlace Point) { 1639 auto *Accesses = getOrCreateAccessList(BB); 1640 if (Point == Beginning) { 1641 // If it's a phi node, it goes first, otherwise, it goes after any phi 1642 // nodes. 1643 if (isa<MemoryPhi>(NewAccess)) { 1644 Accesses->push_front(NewAccess); 1645 auto *Defs = getOrCreateDefsList(BB); 1646 Defs->push_front(*NewAccess); 1647 } else { 1648 auto AI = find_if_not( 1649 *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); }); 1650 Accesses->insert(AI, NewAccess); 1651 if (!isa<MemoryUse>(NewAccess)) { 1652 auto *Defs = getOrCreateDefsList(BB); 1653 auto DI = find_if_not( 1654 *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); }); 1655 Defs->insert(DI, *NewAccess); 1656 } 1657 } 1658 } else { 1659 Accesses->push_back(NewAccess); 1660 if (!isa<MemoryUse>(NewAccess)) { 1661 auto *Defs = getOrCreateDefsList(BB); 1662 Defs->push_back(*NewAccess); 1663 } 1664 } 1665 BlockNumberingValid.erase(BB); 1666 } 1667 1668 void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB, 1669 AccessList::iterator InsertPt) { 1670 auto *Accesses = getWritableBlockAccesses(BB); 1671 bool WasEnd = InsertPt == Accesses->end(); 1672 Accesses->insert(AccessList::iterator(InsertPt), What); 1673 if (!isa<MemoryUse>(What)) { 1674 auto *Defs = getOrCreateDefsList(BB); 1675 // If we got asked to insert at the end, we have an easy job, just shove it 1676 // at the end. If we got asked to insert before an existing def, we also get 1677 // an iterator. If we got asked to insert before a use, we have to hunt for 1678 // the next def. 1679 if (WasEnd) { 1680 Defs->push_back(*What); 1681 } else if (isa<MemoryDef>(InsertPt)) { 1682 Defs->insert(InsertPt->getDefsIterator(), *What); 1683 } else { 1684 while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt)) 1685 ++InsertPt; 1686 // Either we found a def, or we are inserting at the end 1687 if (InsertPt == Accesses->end()) 1688 Defs->push_back(*What); 1689 else 1690 Defs->insert(InsertPt->getDefsIterator(), *What); 1691 } 1692 } 1693 BlockNumberingValid.erase(BB); 1694 } 1695 1696 void MemorySSA::prepareForMoveTo(MemoryAccess *What, BasicBlock *BB) { 1697 // Keep it in the lookup tables, remove from the lists 1698 removeFromLists(What, false); 1699 1700 // Note that moving should implicitly invalidate the optimized state of a 1701 // MemoryUse (and Phis can't be optimized). However, it doesn't do so for a 1702 // MemoryDef. 1703 if (auto *MD = dyn_cast<MemoryDef>(What)) 1704 MD->resetOptimized(); 1705 What->setBlock(BB); 1706 } 1707 1708 // Move What before Where in the IR. The end result is that What will belong to 1709 // the right lists and have the right Block set, but will not otherwise be 1710 // correct. It will not have the right defining access, and if it is a def, 1711 // things below it will not properly be updated. 1712 void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB, 1713 AccessList::iterator Where) { 1714 prepareForMoveTo(What, BB); 1715 insertIntoListsBefore(What, BB, Where); 1716 } 1717 1718 void MemorySSA::moveTo(MemoryAccess *What, BasicBlock *BB, 1719 InsertionPlace Point) { 1720 if (isa<MemoryPhi>(What)) { 1721 assert(Point == Beginning && 1722 "Can only move a Phi at the beginning of the block"); 1723 // Update lookup table entry 1724 ValueToMemoryAccess.erase(What->getBlock()); 1725 bool Inserted = ValueToMemoryAccess.insert({BB, What}).second; 1726 (void)Inserted; 1727 assert(Inserted && "Cannot move a Phi to a block that already has one"); 1728 } 1729 1730 prepareForMoveTo(What, BB); 1731 insertIntoListsForBlock(What, BB, Point); 1732 } 1733 1734 MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) { 1735 assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB"); 1736 MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++); 1737 // Phi's always are placed at the front of the block. 1738 insertIntoListsForBlock(Phi, BB, Beginning); 1739 ValueToMemoryAccess[BB] = Phi; 1740 return Phi; 1741 } 1742 1743 MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I, 1744 MemoryAccess *Definition, 1745 const MemoryUseOrDef *Template, 1746 bool CreationMustSucceed) { 1747 assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI"); 1748 MemoryUseOrDef *NewAccess = createNewAccess(I, AA, Template); 1749 if (CreationMustSucceed) 1750 assert(NewAccess != nullptr && "Tried to create a memory access for a " 1751 "non-memory touching instruction"); 1752 if (NewAccess) { 1753 assert((!Definition || !isa<MemoryUse>(Definition)) && 1754 "A use cannot be a defining access"); 1755 NewAccess->setDefiningAccess(Definition); 1756 } 1757 return NewAccess; 1758 } 1759 1760 // Return true if the instruction has ordering constraints. 1761 // Note specifically that this only considers stores and loads 1762 // because others are still considered ModRef by getModRefInfo. 1763 static inline bool isOrdered(const Instruction *I) { 1764 if (auto *SI = dyn_cast<StoreInst>(I)) { 1765 if (!SI->isUnordered()) 1766 return true; 1767 } else if (auto *LI = dyn_cast<LoadInst>(I)) { 1768 if (!LI->isUnordered()) 1769 return true; 1770 } 1771 return false; 1772 } 1773 1774 /// Helper function to create new memory accesses 1775 template <typename AliasAnalysisType> 1776 MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I, 1777 AliasAnalysisType *AAP, 1778 const MemoryUseOrDef *Template) { 1779 // The assume intrinsic has a control dependency which we model by claiming 1780 // that it writes arbitrarily. Debuginfo intrinsics may be considered 1781 // clobbers when we have a nonstandard AA pipeline. Ignore these fake memory 1782 // dependencies here. 1783 // FIXME: Replace this special casing with a more accurate modelling of 1784 // assume's control dependency. 1785 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { 1786 switch (II->getIntrinsicID()) { 1787 default: 1788 break; 1789 case Intrinsic::assume: 1790 case Intrinsic::experimental_noalias_scope_decl: 1791 case Intrinsic::pseudoprobe: 1792 return nullptr; 1793 } 1794 } 1795 1796 // Using a nonstandard AA pipelines might leave us with unexpected modref 1797 // results for I, so add a check to not model instructions that may not read 1798 // from or write to memory. This is necessary for correctness. 1799 if (!I->mayReadFromMemory() && !I->mayWriteToMemory()) 1800 return nullptr; 1801 1802 bool Def, Use; 1803 if (Template) { 1804 Def = isa<MemoryDef>(Template); 1805 Use = isa<MemoryUse>(Template); 1806 #if !defined(NDEBUG) 1807 ModRefInfo ModRef = AAP->getModRefInfo(I, None); 1808 bool DefCheck, UseCheck; 1809 DefCheck = isModSet(ModRef) || isOrdered(I); 1810 UseCheck = isRefSet(ModRef); 1811 assert(Def == DefCheck && (Def || Use == UseCheck) && "Invalid template"); 1812 #endif 1813 } else { 1814 // Find out what affect this instruction has on memory. 1815 ModRefInfo ModRef = AAP->getModRefInfo(I, None); 1816 // The isOrdered check is used to ensure that volatiles end up as defs 1817 // (atomics end up as ModRef right now anyway). Until we separate the 1818 // ordering chain from the memory chain, this enables people to see at least 1819 // some relative ordering to volatiles. Note that getClobberingMemoryAccess 1820 // will still give an answer that bypasses other volatile loads. TODO: 1821 // Separate memory aliasing and ordering into two different chains so that 1822 // we can precisely represent both "what memory will this read/write/is 1823 // clobbered by" and "what instructions can I move this past". 1824 Def = isModSet(ModRef) || isOrdered(I); 1825 Use = isRefSet(ModRef); 1826 } 1827 1828 // It's possible for an instruction to not modify memory at all. During 1829 // construction, we ignore them. 1830 if (!Def && !Use) 1831 return nullptr; 1832 1833 MemoryUseOrDef *MUD; 1834 if (Def) 1835 MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++); 1836 else 1837 MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent()); 1838 ValueToMemoryAccess[I] = MUD; 1839 return MUD; 1840 } 1841 1842 /// Properly remove \p MA from all of MemorySSA's lookup tables. 1843 void MemorySSA::removeFromLookups(MemoryAccess *MA) { 1844 assert(MA->use_empty() && 1845 "Trying to remove memory access that still has uses"); 1846 BlockNumbering.erase(MA); 1847 if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) 1848 MUD->setDefiningAccess(nullptr); 1849 // Invalidate our walker's cache if necessary 1850 if (!isa<MemoryUse>(MA)) 1851 getWalker()->invalidateInfo(MA); 1852 1853 Value *MemoryInst; 1854 if (const auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) 1855 MemoryInst = MUD->getMemoryInst(); 1856 else 1857 MemoryInst = MA->getBlock(); 1858 1859 auto VMA = ValueToMemoryAccess.find(MemoryInst); 1860 if (VMA->second == MA) 1861 ValueToMemoryAccess.erase(VMA); 1862 } 1863 1864 /// Properly remove \p MA from all of MemorySSA's lists. 1865 /// 1866 /// Because of the way the intrusive list and use lists work, it is important to 1867 /// do removal in the right order. 1868 /// ShouldDelete defaults to true, and will cause the memory access to also be 1869 /// deleted, not just removed. 1870 void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) { 1871 BasicBlock *BB = MA->getBlock(); 1872 // The access list owns the reference, so we erase it from the non-owning list 1873 // first. 1874 if (!isa<MemoryUse>(MA)) { 1875 auto DefsIt = PerBlockDefs.find(BB); 1876 std::unique_ptr<DefsList> &Defs = DefsIt->second; 1877 Defs->remove(*MA); 1878 if (Defs->empty()) 1879 PerBlockDefs.erase(DefsIt); 1880 } 1881 1882 // The erase call here will delete it. If we don't want it deleted, we call 1883 // remove instead. 1884 auto AccessIt = PerBlockAccesses.find(BB); 1885 std::unique_ptr<AccessList> &Accesses = AccessIt->second; 1886 if (ShouldDelete) 1887 Accesses->erase(MA); 1888 else 1889 Accesses->remove(MA); 1890 1891 if (Accesses->empty()) { 1892 PerBlockAccesses.erase(AccessIt); 1893 BlockNumberingValid.erase(BB); 1894 } 1895 } 1896 1897 void MemorySSA::print(raw_ostream &OS) const { 1898 MemorySSAAnnotatedWriter Writer(this); 1899 F.print(OS, &Writer); 1900 } 1901 1902 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1903 LLVM_DUMP_METHOD void MemorySSA::dump() const { print(dbgs()); } 1904 #endif 1905 1906 void MemorySSA::verifyMemorySSA(VerificationLevel VL) const { 1907 #if !defined(NDEBUG) && defined(EXPENSIVE_CHECKS) 1908 VL = VerificationLevel::Full; 1909 #endif 1910 1911 #ifndef NDEBUG 1912 verifyOrderingDominationAndDefUses(F, VL); 1913 verifyDominationNumbers(F); 1914 if (VL == VerificationLevel::Full) 1915 verifyPrevDefInPhis(F); 1916 #endif 1917 // Previously, the verification used to also verify that the clobberingAccess 1918 // cached by MemorySSA is the same as the clobberingAccess found at a later 1919 // query to AA. This does not hold true in general due to the current fragility 1920 // of BasicAA which has arbitrary caps on the things it analyzes before giving 1921 // up. As a result, transformations that are correct, will lead to BasicAA 1922 // returning different Alias answers before and after that transformation. 1923 // Invalidating MemorySSA is not an option, as the results in BasicAA can be so 1924 // random, in the worst case we'd need to rebuild MemorySSA from scratch after 1925 // every transformation, which defeats the purpose of using it. For such an 1926 // example, see test4 added in D51960. 1927 } 1928 1929 void MemorySSA::verifyPrevDefInPhis(Function &F) const { 1930 for (const BasicBlock &BB : F) { 1931 if (MemoryPhi *Phi = getMemoryAccess(&BB)) { 1932 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) { 1933 auto *Pred = Phi->getIncomingBlock(I); 1934 auto *IncAcc = Phi->getIncomingValue(I); 1935 // If Pred has no unreachable predecessors, get last def looking at 1936 // IDoms. If, while walkings IDoms, any of these has an unreachable 1937 // predecessor, then the incoming def can be any access. 1938 if (auto *DTNode = DT->getNode(Pred)) { 1939 while (DTNode) { 1940 if (auto *DefList = getBlockDefs(DTNode->getBlock())) { 1941 auto *LastAcc = &*(--DefList->end()); 1942 assert(LastAcc == IncAcc && 1943 "Incorrect incoming access into phi."); 1944 (void)IncAcc; 1945 (void)LastAcc; 1946 break; 1947 } 1948 DTNode = DTNode->getIDom(); 1949 } 1950 } else { 1951 // If Pred has unreachable predecessors, but has at least a Def, the 1952 // incoming access can be the last Def in Pred, or it could have been 1953 // optimized to LoE. After an update, though, the LoE may have been 1954 // replaced by another access, so IncAcc may be any access. 1955 // If Pred has unreachable predecessors and no Defs, incoming access 1956 // should be LoE; However, after an update, it may be any access. 1957 } 1958 } 1959 } 1960 } 1961 } 1962 1963 /// Verify that all of the blocks we believe to have valid domination numbers 1964 /// actually have valid domination numbers. 1965 void MemorySSA::verifyDominationNumbers(const Function &F) const { 1966 if (BlockNumberingValid.empty()) 1967 return; 1968 1969 SmallPtrSet<const BasicBlock *, 16> ValidBlocks = BlockNumberingValid; 1970 for (const BasicBlock &BB : F) { 1971 if (!ValidBlocks.count(&BB)) 1972 continue; 1973 1974 ValidBlocks.erase(&BB); 1975 1976 const AccessList *Accesses = getBlockAccesses(&BB); 1977 // It's correct to say an empty block has valid numbering. 1978 if (!Accesses) 1979 continue; 1980 1981 // Block numbering starts at 1. 1982 unsigned long LastNumber = 0; 1983 for (const MemoryAccess &MA : *Accesses) { 1984 auto ThisNumberIter = BlockNumbering.find(&MA); 1985 assert(ThisNumberIter != BlockNumbering.end() && 1986 "MemoryAccess has no domination number in a valid block!"); 1987 1988 unsigned long ThisNumber = ThisNumberIter->second; 1989 assert(ThisNumber > LastNumber && 1990 "Domination numbers should be strictly increasing!"); 1991 (void)LastNumber; 1992 LastNumber = ThisNumber; 1993 } 1994 } 1995 1996 assert(ValidBlocks.empty() && 1997 "All valid BasicBlocks should exist in F -- dangling pointers?"); 1998 } 1999 2000 /// Verify ordering: the order and existence of MemoryAccesses matches the 2001 /// order and existence of memory affecting instructions. 2002 /// Verify domination: each definition dominates all of its uses. 2003 /// Verify def-uses: the immediate use information - walk all the memory 2004 /// accesses and verifying that, for each use, it appears in the appropriate 2005 /// def's use list 2006 void MemorySSA::verifyOrderingDominationAndDefUses(Function &F, 2007 VerificationLevel VL) const { 2008 // Walk all the blocks, comparing what the lookups think and what the access 2009 // lists think, as well as the order in the blocks vs the order in the access 2010 // lists. 2011 SmallVector<MemoryAccess *, 32> ActualAccesses; 2012 SmallVector<MemoryAccess *, 32> ActualDefs; 2013 for (BasicBlock &B : F) { 2014 const AccessList *AL = getBlockAccesses(&B); 2015 const auto *DL = getBlockDefs(&B); 2016 MemoryPhi *Phi = getMemoryAccess(&B); 2017 if (Phi) { 2018 // Verify ordering. 2019 ActualAccesses.push_back(Phi); 2020 ActualDefs.push_back(Phi); 2021 // Verify domination 2022 for (const Use &U : Phi->uses()) { 2023 assert(dominates(Phi, U) && "Memory PHI does not dominate it's uses"); 2024 (void)U; 2025 } 2026 // Verify def-uses for full verify. 2027 if (VL == VerificationLevel::Full) { 2028 assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance( 2029 pred_begin(&B), pred_end(&B))) && 2030 "Incomplete MemoryPhi Node"); 2031 for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) { 2032 verifyUseInDefs(Phi->getIncomingValue(I), Phi); 2033 assert(is_contained(predecessors(&B), Phi->getIncomingBlock(I)) && 2034 "Incoming phi block not a block predecessor"); 2035 } 2036 } 2037 } 2038 2039 for (Instruction &I : B) { 2040 MemoryUseOrDef *MA = getMemoryAccess(&I); 2041 assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) && 2042 "We have memory affecting instructions " 2043 "in this block but they are not in the " 2044 "access list or defs list"); 2045 if (MA) { 2046 // Verify ordering. 2047 ActualAccesses.push_back(MA); 2048 if (MemoryAccess *MD = dyn_cast<MemoryDef>(MA)) { 2049 // Verify ordering. 2050 ActualDefs.push_back(MA); 2051 // Verify domination. 2052 for (const Use &U : MD->uses()) { 2053 assert(dominates(MD, U) && 2054 "Memory Def does not dominate it's uses"); 2055 (void)U; 2056 } 2057 } 2058 // Verify def-uses for full verify. 2059 if (VL == VerificationLevel::Full) 2060 verifyUseInDefs(MA->getDefiningAccess(), MA); 2061 } 2062 } 2063 // Either we hit the assert, really have no accesses, or we have both 2064 // accesses and an access list. Same with defs. 2065 if (!AL && !DL) 2066 continue; 2067 // Verify ordering. 2068 assert(AL->size() == ActualAccesses.size() && 2069 "We don't have the same number of accesses in the block as on the " 2070 "access list"); 2071 assert((DL || ActualDefs.size() == 0) && 2072 "Either we should have a defs list, or we should have no defs"); 2073 assert((!DL || DL->size() == ActualDefs.size()) && 2074 "We don't have the same number of defs in the block as on the " 2075 "def list"); 2076 auto ALI = AL->begin(); 2077 auto AAI = ActualAccesses.begin(); 2078 while (ALI != AL->end() && AAI != ActualAccesses.end()) { 2079 assert(&*ALI == *AAI && "Not the same accesses in the same order"); 2080 ++ALI; 2081 ++AAI; 2082 } 2083 ActualAccesses.clear(); 2084 if (DL) { 2085 auto DLI = DL->begin(); 2086 auto ADI = ActualDefs.begin(); 2087 while (DLI != DL->end() && ADI != ActualDefs.end()) { 2088 assert(&*DLI == *ADI && "Not the same defs in the same order"); 2089 ++DLI; 2090 ++ADI; 2091 } 2092 } 2093 ActualDefs.clear(); 2094 } 2095 } 2096 2097 /// Verify the def-use lists in MemorySSA, by verifying that \p Use 2098 /// appears in the use list of \p Def. 2099 void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const { 2100 // The live on entry use may cause us to get a NULL def here 2101 if (!Def) 2102 assert(isLiveOnEntryDef(Use) && 2103 "Null def but use not point to live on entry def"); 2104 else 2105 assert(is_contained(Def->users(), Use) && 2106 "Did not find use in def's use list"); 2107 } 2108 2109 /// Perform a local numbering on blocks so that instruction ordering can be 2110 /// determined in constant time. 2111 /// TODO: We currently just number in order. If we numbered by N, we could 2112 /// allow at least N-1 sequences of insertBefore or insertAfter (and at least 2113 /// log2(N) sequences of mixed before and after) without needing to invalidate 2114 /// the numbering. 2115 void MemorySSA::renumberBlock(const BasicBlock *B) const { 2116 // The pre-increment ensures the numbers really start at 1. 2117 unsigned long CurrentNumber = 0; 2118 const AccessList *AL = getBlockAccesses(B); 2119 assert(AL != nullptr && "Asking to renumber an empty block"); 2120 for (const auto &I : *AL) 2121 BlockNumbering[&I] = ++CurrentNumber; 2122 BlockNumberingValid.insert(B); 2123 } 2124 2125 /// Determine, for two memory accesses in the same block, 2126 /// whether \p Dominator dominates \p Dominatee. 2127 /// \returns True if \p Dominator dominates \p Dominatee. 2128 bool MemorySSA::locallyDominates(const MemoryAccess *Dominator, 2129 const MemoryAccess *Dominatee) const { 2130 const BasicBlock *DominatorBlock = Dominator->getBlock(); 2131 2132 assert((DominatorBlock == Dominatee->getBlock()) && 2133 "Asking for local domination when accesses are in different blocks!"); 2134 // A node dominates itself. 2135 if (Dominatee == Dominator) 2136 return true; 2137 2138 // When Dominatee is defined on function entry, it is not dominated by another 2139 // memory access. 2140 if (isLiveOnEntryDef(Dominatee)) 2141 return false; 2142 2143 // When Dominator is defined on function entry, it dominates the other memory 2144 // access. 2145 if (isLiveOnEntryDef(Dominator)) 2146 return true; 2147 2148 if (!BlockNumberingValid.count(DominatorBlock)) 2149 renumberBlock(DominatorBlock); 2150 2151 unsigned long DominatorNum = BlockNumbering.lookup(Dominator); 2152 // All numbers start with 1 2153 assert(DominatorNum != 0 && "Block was not numbered properly"); 2154 unsigned long DominateeNum = BlockNumbering.lookup(Dominatee); 2155 assert(DominateeNum != 0 && "Block was not numbered properly"); 2156 return DominatorNum < DominateeNum; 2157 } 2158 2159 bool MemorySSA::dominates(const MemoryAccess *Dominator, 2160 const MemoryAccess *Dominatee) const { 2161 if (Dominator == Dominatee) 2162 return true; 2163 2164 if (isLiveOnEntryDef(Dominatee)) 2165 return false; 2166 2167 if (Dominator->getBlock() != Dominatee->getBlock()) 2168 return DT->dominates(Dominator->getBlock(), Dominatee->getBlock()); 2169 return locallyDominates(Dominator, Dominatee); 2170 } 2171 2172 bool MemorySSA::dominates(const MemoryAccess *Dominator, 2173 const Use &Dominatee) const { 2174 if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) { 2175 BasicBlock *UseBB = MP->getIncomingBlock(Dominatee); 2176 // The def must dominate the incoming block of the phi. 2177 if (UseBB != Dominator->getBlock()) 2178 return DT->dominates(Dominator->getBlock(), UseBB); 2179 // If the UseBB and the DefBB are the same, compare locally. 2180 return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee)); 2181 } 2182 // If it's not a PHI node use, the normal dominates can already handle it. 2183 return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser())); 2184 } 2185 2186 void MemorySSA::ensureOptimizedUses() { 2187 if (IsOptimized) 2188 return; 2189 2190 BatchAAResults BatchAA(*AA); 2191 ClobberWalkerBase<BatchAAResults> WalkerBase(this, &BatchAA, DT); 2192 CachingWalker<BatchAAResults> WalkerLocal(this, &WalkerBase); 2193 OptimizeUses(this, &WalkerLocal, &BatchAA, DT).optimizeUses(); 2194 IsOptimized = true; 2195 } 2196 2197 void MemoryAccess::print(raw_ostream &OS) const { 2198 switch (getValueID()) { 2199 case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS); 2200 case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS); 2201 case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS); 2202 } 2203 llvm_unreachable("invalid value id"); 2204 } 2205 2206 void MemoryDef::print(raw_ostream &OS) const { 2207 MemoryAccess *UO = getDefiningAccess(); 2208 2209 auto printID = [&OS](MemoryAccess *A) { 2210 if (A && A->getID()) 2211 OS << A->getID(); 2212 else 2213 OS << LiveOnEntryStr; 2214 }; 2215 2216 OS << getID() << " = MemoryDef("; 2217 printID(UO); 2218 OS << ")"; 2219 2220 if (isOptimized()) { 2221 OS << "->"; 2222 printID(getOptimized()); 2223 2224 if (Optional<AliasResult> AR = getOptimizedAccessType()) 2225 OS << " " << *AR; 2226 } 2227 } 2228 2229 void MemoryPhi::print(raw_ostream &OS) const { 2230 ListSeparator LS(","); 2231 OS << getID() << " = MemoryPhi("; 2232 for (const auto &Op : operands()) { 2233 BasicBlock *BB = getIncomingBlock(Op); 2234 MemoryAccess *MA = cast<MemoryAccess>(Op); 2235 2236 OS << LS << '{'; 2237 if (BB->hasName()) 2238 OS << BB->getName(); 2239 else 2240 BB->printAsOperand(OS, false); 2241 OS << ','; 2242 if (unsigned ID = MA->getID()) 2243 OS << ID; 2244 else 2245 OS << LiveOnEntryStr; 2246 OS << '}'; 2247 } 2248 OS << ')'; 2249 } 2250 2251 void MemoryUse::print(raw_ostream &OS) const { 2252 MemoryAccess *UO = getDefiningAccess(); 2253 OS << "MemoryUse("; 2254 if (UO && UO->getID()) 2255 OS << UO->getID(); 2256 else 2257 OS << LiveOnEntryStr; 2258 OS << ')'; 2259 2260 if (Optional<AliasResult> AR = getOptimizedAccessType()) 2261 OS << " " << *AR; 2262 } 2263 2264 void MemoryAccess::dump() const { 2265 // Cannot completely remove virtual function even in release mode. 2266 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 2267 print(dbgs()); 2268 dbgs() << "\n"; 2269 #endif 2270 } 2271 2272 char MemorySSAPrinterLegacyPass::ID = 0; 2273 2274 MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) { 2275 initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry()); 2276 } 2277 2278 void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const { 2279 AU.setPreservesAll(); 2280 AU.addRequired<MemorySSAWrapperPass>(); 2281 } 2282 2283 class DOTFuncMSSAInfo { 2284 private: 2285 const Function &F; 2286 MemorySSAAnnotatedWriter MSSAWriter; 2287 2288 public: 2289 DOTFuncMSSAInfo(const Function &F, MemorySSA &MSSA) 2290 : F(F), MSSAWriter(&MSSA) {} 2291 2292 const Function *getFunction() { return &F; } 2293 MemorySSAAnnotatedWriter &getWriter() { return MSSAWriter; } 2294 }; 2295 2296 namespace llvm { 2297 2298 template <> 2299 struct GraphTraits<DOTFuncMSSAInfo *> : public GraphTraits<const BasicBlock *> { 2300 static NodeRef getEntryNode(DOTFuncMSSAInfo *CFGInfo) { 2301 return &(CFGInfo->getFunction()->getEntryBlock()); 2302 } 2303 2304 // nodes_iterator/begin/end - Allow iteration over all nodes in the graph 2305 using nodes_iterator = pointer_iterator<Function::const_iterator>; 2306 2307 static nodes_iterator nodes_begin(DOTFuncMSSAInfo *CFGInfo) { 2308 return nodes_iterator(CFGInfo->getFunction()->begin()); 2309 } 2310 2311 static nodes_iterator nodes_end(DOTFuncMSSAInfo *CFGInfo) { 2312 return nodes_iterator(CFGInfo->getFunction()->end()); 2313 } 2314 2315 static size_t size(DOTFuncMSSAInfo *CFGInfo) { 2316 return CFGInfo->getFunction()->size(); 2317 } 2318 }; 2319 2320 template <> 2321 struct DOTGraphTraits<DOTFuncMSSAInfo *> : public DefaultDOTGraphTraits { 2322 2323 DOTGraphTraits(bool IsSimple = false) : DefaultDOTGraphTraits(IsSimple) {} 2324 2325 static std::string getGraphName(DOTFuncMSSAInfo *CFGInfo) { 2326 return "MSSA CFG for '" + CFGInfo->getFunction()->getName().str() + 2327 "' function"; 2328 } 2329 2330 std::string getNodeLabel(const BasicBlock *Node, DOTFuncMSSAInfo *CFGInfo) { 2331 return DOTGraphTraits<DOTFuncInfo *>::getCompleteNodeLabel( 2332 Node, nullptr, 2333 [CFGInfo](raw_string_ostream &OS, const BasicBlock &BB) -> void { 2334 BB.print(OS, &CFGInfo->getWriter(), true, true); 2335 }, 2336 [](std::string &S, unsigned &I, unsigned Idx) -> void { 2337 std::string Str = S.substr(I, Idx - I); 2338 StringRef SR = Str; 2339 if (SR.count(" = MemoryDef(") || SR.count(" = MemoryPhi(") || 2340 SR.count("MemoryUse(")) 2341 return; 2342 DOTGraphTraits<DOTFuncInfo *>::eraseComment(S, I, Idx); 2343 }); 2344 } 2345 2346 static std::string getEdgeSourceLabel(const BasicBlock *Node, 2347 const_succ_iterator I) { 2348 return DOTGraphTraits<DOTFuncInfo *>::getEdgeSourceLabel(Node, I); 2349 } 2350 2351 /// Display the raw branch weights from PGO. 2352 std::string getEdgeAttributes(const BasicBlock *Node, const_succ_iterator I, 2353 DOTFuncMSSAInfo *CFGInfo) { 2354 return ""; 2355 } 2356 2357 std::string getNodeAttributes(const BasicBlock *Node, 2358 DOTFuncMSSAInfo *CFGInfo) { 2359 return getNodeLabel(Node, CFGInfo).find(';') != std::string::npos 2360 ? "style=filled, fillcolor=lightpink" 2361 : ""; 2362 } 2363 }; 2364 2365 } // namespace llvm 2366 2367 bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) { 2368 auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA(); 2369 MSSA.ensureOptimizedUses(); 2370 if (DotCFGMSSA != "") { 2371 DOTFuncMSSAInfo CFGInfo(F, MSSA); 2372 WriteGraph(&CFGInfo, "", false, "MSSA", DotCFGMSSA); 2373 } else 2374 MSSA.print(dbgs()); 2375 2376 if (VerifyMemorySSA) 2377 MSSA.verifyMemorySSA(); 2378 return false; 2379 } 2380 2381 AnalysisKey MemorySSAAnalysis::Key; 2382 2383 MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F, 2384 FunctionAnalysisManager &AM) { 2385 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 2386 auto &AA = AM.getResult<AAManager>(F); 2387 return MemorySSAAnalysis::Result(std::make_unique<MemorySSA>(F, &AA, &DT)); 2388 } 2389 2390 bool MemorySSAAnalysis::Result::invalidate( 2391 Function &F, const PreservedAnalyses &PA, 2392 FunctionAnalysisManager::Invalidator &Inv) { 2393 auto PAC = PA.getChecker<MemorySSAAnalysis>(); 2394 return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>()) || 2395 Inv.invalidate<AAManager>(F, PA) || 2396 Inv.invalidate<DominatorTreeAnalysis>(F, PA); 2397 } 2398 2399 PreservedAnalyses MemorySSAPrinterPass::run(Function &F, 2400 FunctionAnalysisManager &AM) { 2401 auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA(); 2402 MSSA.ensureOptimizedUses(); 2403 if (DotCFGMSSA != "") { 2404 DOTFuncMSSAInfo CFGInfo(F, MSSA); 2405 WriteGraph(&CFGInfo, "", false, "MSSA", DotCFGMSSA); 2406 } else { 2407 OS << "MemorySSA for function: " << F.getName() << "\n"; 2408 MSSA.print(OS); 2409 } 2410 2411 return PreservedAnalyses::all(); 2412 } 2413 2414 PreservedAnalyses MemorySSAWalkerPrinterPass::run(Function &F, 2415 FunctionAnalysisManager &AM) { 2416 auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA(); 2417 OS << "MemorySSA (walker) for function: " << F.getName() << "\n"; 2418 MemorySSAWalkerAnnotatedWriter Writer(&MSSA); 2419 F.print(OS, &Writer); 2420 2421 return PreservedAnalyses::all(); 2422 } 2423 2424 PreservedAnalyses MemorySSAVerifierPass::run(Function &F, 2425 FunctionAnalysisManager &AM) { 2426 AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA(); 2427 2428 return PreservedAnalyses::all(); 2429 } 2430 2431 char MemorySSAWrapperPass::ID = 0; 2432 2433 MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) { 2434 initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry()); 2435 } 2436 2437 void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); } 2438 2439 void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { 2440 AU.setPreservesAll(); 2441 AU.addRequiredTransitive<DominatorTreeWrapperPass>(); 2442 AU.addRequiredTransitive<AAResultsWrapperPass>(); 2443 } 2444 2445 bool MemorySSAWrapperPass::runOnFunction(Function &F) { 2446 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2447 auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); 2448 MSSA.reset(new MemorySSA(F, &AA, &DT)); 2449 return false; 2450 } 2451 2452 void MemorySSAWrapperPass::verifyAnalysis() const { 2453 if (VerifyMemorySSA) 2454 MSSA->verifyMemorySSA(); 2455 } 2456 2457 void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const { 2458 MSSA->print(OS); 2459 } 2460 2461 MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {} 2462 2463 /// Walk the use-def chains starting at \p StartingAccess and find 2464 /// the MemoryAccess that actually clobbers Loc. 2465 /// 2466 /// \returns our clobbering memory access 2467 template <typename AliasAnalysisType> 2468 MemoryAccess * 2469 MemorySSA::ClobberWalkerBase<AliasAnalysisType>::getClobberingMemoryAccessBase( 2470 MemoryAccess *StartingAccess, const MemoryLocation &Loc, 2471 unsigned &UpwardWalkLimit) { 2472 assert(!isa<MemoryUse>(StartingAccess) && "Use cannot be defining access"); 2473 2474 Instruction *I = nullptr; 2475 if (auto *StartingUseOrDef = dyn_cast<MemoryUseOrDef>(StartingAccess)) { 2476 if (MSSA->isLiveOnEntryDef(StartingUseOrDef)) 2477 return StartingUseOrDef; 2478 2479 I = StartingUseOrDef->getMemoryInst(); 2480 2481 // Conservatively, fences are always clobbers, so don't perform the walk if 2482 // we hit a fence. 2483 if (!isa<CallBase>(I) && I->isFenceLike()) 2484 return StartingUseOrDef; 2485 } 2486 2487 UpwardsMemoryQuery Q; 2488 Q.OriginalAccess = StartingAccess; 2489 Q.StartingLoc = Loc; 2490 Q.Inst = nullptr; 2491 Q.IsCall = false; 2492 2493 // Unlike the other function, do not walk to the def of a def, because we are 2494 // handed something we already believe is the clobbering access. 2495 // We never set SkipSelf to true in Q in this method. 2496 MemoryAccess *Clobber = 2497 Walker.findClobber(StartingAccess, Q, UpwardWalkLimit); 2498 LLVM_DEBUG({ 2499 dbgs() << "Clobber starting at access " << *StartingAccess << "\n"; 2500 if (I) 2501 dbgs() << " for instruction " << *I << "\n"; 2502 dbgs() << " is " << *Clobber << "\n"; 2503 }); 2504 return Clobber; 2505 } 2506 2507 static const Instruction * 2508 getInvariantGroupClobberingInstruction(Instruction &I, DominatorTree &DT) { 2509 if (!I.hasMetadata(LLVMContext::MD_invariant_group) || I.isVolatile()) 2510 return nullptr; 2511 2512 // We consider bitcasts and zero GEPs to be the same pointer value. Start by 2513 // stripping bitcasts and zero GEPs, then we will recursively look at loads 2514 // and stores through bitcasts and zero GEPs. 2515 Value *PointerOperand = getLoadStorePointerOperand(&I)->stripPointerCasts(); 2516 2517 // It's not safe to walk the use list of a global value because function 2518 // passes aren't allowed to look outside their functions. 2519 // FIXME: this could be fixed by filtering instructions from outside of 2520 // current function. 2521 if (isa<Constant>(PointerOperand)) 2522 return nullptr; 2523 2524 // Queue to process all pointers that are equivalent to load operand. 2525 SmallVector<const Value *, 8> PointerUsesQueue; 2526 PointerUsesQueue.push_back(PointerOperand); 2527 2528 const Instruction *MostDominatingInstruction = &I; 2529 2530 // FIXME: This loop is O(n^2) because dominates can be O(n) and in worst case 2531 // we will see all the instructions. It may not matter in practice. If it 2532 // does, we will have to support MemorySSA construction and updates. 2533 while (!PointerUsesQueue.empty()) { 2534 const Value *Ptr = PointerUsesQueue.pop_back_val(); 2535 assert(Ptr && !isa<GlobalValue>(Ptr) && 2536 "Null or GlobalValue should not be inserted"); 2537 2538 for (const User *Us : Ptr->users()) { 2539 auto *U = dyn_cast<Instruction>(Us); 2540 if (!U || U == &I || !DT.dominates(U, MostDominatingInstruction)) 2541 continue; 2542 2543 // Add bitcasts and zero GEPs to queue. 2544 if (isa<BitCastInst>(U)) { 2545 PointerUsesQueue.push_back(U); 2546 continue; 2547 } 2548 if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) { 2549 if (GEP->hasAllZeroIndices()) 2550 PointerUsesQueue.push_back(U); 2551 continue; 2552 } 2553 2554 // If we hit a load/store with an invariant.group metadata and the same 2555 // pointer operand, we can assume that value pointed to by the pointer 2556 // operand didn't change. 2557 if (U->hasMetadata(LLVMContext::MD_invariant_group) && 2558 getLoadStorePointerOperand(U) == Ptr && !U->isVolatile()) { 2559 MostDominatingInstruction = U; 2560 } 2561 } 2562 } 2563 return MostDominatingInstruction == &I ? nullptr : MostDominatingInstruction; 2564 } 2565 2566 template <typename AliasAnalysisType> 2567 MemoryAccess * 2568 MemorySSA::ClobberWalkerBase<AliasAnalysisType>::getClobberingMemoryAccessBase( 2569 MemoryAccess *MA, unsigned &UpwardWalkLimit, bool SkipSelf, 2570 bool UseInvariantGroup) { 2571 auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA); 2572 // If this is a MemoryPhi, we can't do anything. 2573 if (!StartingAccess) 2574 return MA; 2575 2576 if (UseInvariantGroup) { 2577 if (auto *I = getInvariantGroupClobberingInstruction( 2578 *StartingAccess->getMemoryInst(), MSSA->getDomTree())) { 2579 assert(isa<LoadInst>(I) || isa<StoreInst>(I)); 2580 2581 auto *ClobberMA = MSSA->getMemoryAccess(I); 2582 assert(ClobberMA); 2583 if (isa<MemoryUse>(ClobberMA)) 2584 return ClobberMA->getDefiningAccess(); 2585 return ClobberMA; 2586 } 2587 } 2588 2589 bool IsOptimized = false; 2590 2591 // If this is an already optimized use or def, return the optimized result. 2592 // Note: Currently, we store the optimized def result in a separate field, 2593 // since we can't use the defining access. 2594 if (StartingAccess->isOptimized()) { 2595 if (!SkipSelf || !isa<MemoryDef>(StartingAccess)) 2596 return StartingAccess->getOptimized(); 2597 IsOptimized = true; 2598 } 2599 2600 const Instruction *I = StartingAccess->getMemoryInst(); 2601 // We can't sanely do anything with a fence, since they conservatively clobber 2602 // all memory, and have no locations to get pointers from to try to 2603 // disambiguate. 2604 if (!isa<CallBase>(I) && I->isFenceLike()) 2605 return StartingAccess; 2606 2607 UpwardsMemoryQuery Q(I, StartingAccess); 2608 2609 if (isUseTriviallyOptimizableToLiveOnEntry(*Walker.getAA(), I)) { 2610 MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef(); 2611 StartingAccess->setOptimized(LiveOnEntry); 2612 StartingAccess->setOptimizedAccessType(None); 2613 return LiveOnEntry; 2614 } 2615 2616 MemoryAccess *OptimizedAccess; 2617 if (!IsOptimized) { 2618 // Start with the thing we already think clobbers this location 2619 MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess(); 2620 2621 // At this point, DefiningAccess may be the live on entry def. 2622 // If it is, we will not get a better result. 2623 if (MSSA->isLiveOnEntryDef(DefiningAccess)) { 2624 StartingAccess->setOptimized(DefiningAccess); 2625 StartingAccess->setOptimizedAccessType(None); 2626 return DefiningAccess; 2627 } 2628 2629 OptimizedAccess = Walker.findClobber(DefiningAccess, Q, UpwardWalkLimit); 2630 StartingAccess->setOptimized(OptimizedAccess); 2631 if (MSSA->isLiveOnEntryDef(OptimizedAccess)) 2632 StartingAccess->setOptimizedAccessType(None); 2633 else if (Q.AR && *Q.AR == AliasResult::MustAlias) 2634 StartingAccess->setOptimizedAccessType( 2635 AliasResult(AliasResult::MustAlias)); 2636 } else 2637 OptimizedAccess = StartingAccess->getOptimized(); 2638 2639 LLVM_DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is "); 2640 LLVM_DEBUG(dbgs() << *StartingAccess << "\n"); 2641 LLVM_DEBUG(dbgs() << "Optimized Memory SSA clobber for " << *I << " is "); 2642 LLVM_DEBUG(dbgs() << *OptimizedAccess << "\n"); 2643 2644 MemoryAccess *Result; 2645 if (SkipSelf && isa<MemoryPhi>(OptimizedAccess) && 2646 isa<MemoryDef>(StartingAccess) && UpwardWalkLimit) { 2647 assert(isa<MemoryDef>(Q.OriginalAccess)); 2648 Q.SkipSelfAccess = true; 2649 Result = Walker.findClobber(OptimizedAccess, Q, UpwardWalkLimit); 2650 } else 2651 Result = OptimizedAccess; 2652 2653 LLVM_DEBUG(dbgs() << "Result Memory SSA clobber [SkipSelf = " << SkipSelf); 2654 LLVM_DEBUG(dbgs() << "] for " << *I << " is " << *Result << "\n"); 2655 2656 return Result; 2657 } 2658 2659 MemoryAccess * 2660 DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA) { 2661 if (auto *Use = dyn_cast<MemoryUseOrDef>(MA)) 2662 return Use->getDefiningAccess(); 2663 return MA; 2664 } 2665 2666 MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess( 2667 MemoryAccess *StartingAccess, const MemoryLocation &) { 2668 if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess)) 2669 return Use->getDefiningAccess(); 2670 return StartingAccess; 2671 } 2672 2673 void MemoryPhi::deleteMe(DerivedUser *Self) { 2674 delete static_cast<MemoryPhi *>(Self); 2675 } 2676 2677 void MemoryDef::deleteMe(DerivedUser *Self) { 2678 delete static_cast<MemoryDef *>(Self); 2679 } 2680 2681 void MemoryUse::deleteMe(DerivedUser *Self) { 2682 delete static_cast<MemoryUse *>(Self); 2683 } 2684 2685 bool upward_defs_iterator::IsGuaranteedLoopInvariant(Value *Ptr) const { 2686 auto IsGuaranteedLoopInvariantBase = [](Value *Ptr) { 2687 Ptr = Ptr->stripPointerCasts(); 2688 if (!isa<Instruction>(Ptr)) 2689 return true; 2690 return isa<AllocaInst>(Ptr); 2691 }; 2692 2693 Ptr = Ptr->stripPointerCasts(); 2694 if (auto *I = dyn_cast<Instruction>(Ptr)) { 2695 if (I->getParent()->isEntryBlock()) 2696 return true; 2697 } 2698 if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) { 2699 return IsGuaranteedLoopInvariantBase(GEP->getPointerOperand()) && 2700 GEP->hasAllConstantIndices(); 2701 } 2702 return IsGuaranteedLoopInvariantBase(Ptr); 2703 } 2704