| File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/Analysis/MemorySSA.cpp |
| Warning: | line 892, column 37 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 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/Intrinsics.h" | ||||||
| 40 | #include "llvm/IR/LLVMContext.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/raw_ostream.h" | ||||||
| 53 | #include <algorithm> | ||||||
| 54 | #include <cassert> | ||||||
| 55 | #include <cstdlib> | ||||||
| 56 | #include <iterator> | ||||||
| 57 | #include <memory> | ||||||
| 58 | #include <utility> | ||||||
| 59 | |||||||
| 60 | using namespace llvm; | ||||||
| 61 | |||||||
| 62 | #define DEBUG_TYPE"memoryssa" "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,static void *initializeMemorySSAWrapperPassPassOnce(PassRegistry &Registry) { | ||||||
| 70 | true)static void *initializeMemorySSAWrapperPassPassOnce(PassRegistry &Registry) { | ||||||
| 71 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry); | ||||||
| 72 | INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry); | ||||||
| 73 | INITIALIZE_PASS_END(MemorySSAWrapperPass, "memoryssa", "Memory SSA", false,PassInfo *PI = new PassInfo( "Memory SSA", "memoryssa", & MemorySSAWrapperPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <MemorySSAWrapperPass>), false, true); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeMemorySSAWrapperPassPassFlag ; void llvm::initializeMemorySSAWrapperPassPass(PassRegistry & Registry) { llvm::call_once(InitializeMemorySSAWrapperPassPassFlag , initializeMemorySSAWrapperPassPassOnce, std::ref(Registry)) ; } | ||||||
| 74 | true)PassInfo *PI = new PassInfo( "Memory SSA", "memoryssa", & MemorySSAWrapperPass::ID, PassInfo::NormalCtor_t(callDefaultCtor <MemorySSAWrapperPass>), false, true); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeMemorySSAWrapperPassPassFlag ; void llvm::initializeMemorySSAWrapperPassPass(PassRegistry & Registry) { llvm::call_once(InitializeMemorySSAWrapperPassPassFlag , initializeMemorySSAWrapperPassPassOnce, std::ref(Registry)) ; } | ||||||
| 75 | |||||||
| 76 | INITIALIZE_PASS_BEGIN(MemorySSAPrinterLegacyPass, "print-memoryssa",static void *initializeMemorySSAPrinterLegacyPassPassOnce(PassRegistry &Registry) { | ||||||
| 77 | "Memory SSA Printer", false, false)static void *initializeMemorySSAPrinterLegacyPassPassOnce(PassRegistry &Registry) { | ||||||
| 78 | INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry); | ||||||
| 79 | INITIALIZE_PASS_END(MemorySSAPrinterLegacyPass, "print-memoryssa",PassInfo *PI = new PassInfo( "Memory SSA Printer", "print-memoryssa" , &MemorySSAPrinterLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<MemorySSAPrinterLegacyPass>), false, false ); Registry.registerPass(*PI, true); return PI; } static llvm ::once_flag InitializeMemorySSAPrinterLegacyPassPassFlag; void llvm::initializeMemorySSAPrinterLegacyPassPass(PassRegistry & Registry) { llvm::call_once(InitializeMemorySSAPrinterLegacyPassPassFlag , initializeMemorySSAPrinterLegacyPassPassOnce, std::ref(Registry )); } | ||||||
| 80 | "Memory SSA Printer", false, false)PassInfo *PI = new PassInfo( "Memory SSA Printer", "print-memoryssa" , &MemorySSAPrinterLegacyPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<MemorySSAPrinterLegacyPass>), false, false ); Registry.registerPass(*PI, true); return PI; } static llvm ::once_flag InitializeMemorySSAPrinterLegacyPassPassFlag; void llvm::initializeMemorySSAPrinterLegacyPassPass(PassRegistry & Registry) { llvm::call_once(InitializeMemorySSAPrinterLegacyPassPassFlag , initializeMemorySSAPrinterLegacyPassPassOnce, std::ref(Registry )); } | ||||||
| 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 | /// Enables memory ssa as a dependency for loop passes in legacy pass manager. | ||||||
| 94 | cl::opt<bool> llvm::EnableMSSALoopDependency( | ||||||
| 95 | "enable-mssa-loop-dependency", cl::Hidden, cl::init(true), | ||||||
| 96 | cl::desc("Enable MemorySSA dependency for loop pass manager")); | ||||||
| 97 | |||||||
| 98 | static cl::opt<bool, true> | ||||||
| 99 | VerifyMemorySSAX("verify-memoryssa", cl::location(VerifyMemorySSA), | ||||||
| 100 | cl::Hidden, cl::desc("Enable verification of MemorySSA.")); | ||||||
| 101 | |||||||
| 102 | namespace llvm { | ||||||
| 103 | |||||||
| 104 | /// An assembly annotator class to print Memory SSA information in | ||||||
| 105 | /// comments. | ||||||
| 106 | class MemorySSAAnnotatedWriter : public AssemblyAnnotationWriter { | ||||||
| 107 | friend class MemorySSA; | ||||||
| 108 | |||||||
| 109 | const MemorySSA *MSSA; | ||||||
| 110 | |||||||
| 111 | public: | ||||||
| 112 | MemorySSAAnnotatedWriter(const MemorySSA *M) : MSSA(M) {} | ||||||
| 113 | |||||||
| 114 | void emitBasicBlockStartAnnot(const BasicBlock *BB, | ||||||
| 115 | formatted_raw_ostream &OS) override { | ||||||
| 116 | if (MemoryAccess *MA = MSSA->getMemoryAccess(BB)) | ||||||
| 117 | OS << "; " << *MA << "\n"; | ||||||
| 118 | } | ||||||
| 119 | |||||||
| 120 | void emitInstructionAnnot(const Instruction *I, | ||||||
| 121 | formatted_raw_ostream &OS) override { | ||||||
| 122 | if (MemoryAccess *MA = MSSA->getMemoryAccess(I)) | ||||||
| 123 | OS << "; " << *MA << "\n"; | ||||||
| 124 | } | ||||||
| 125 | }; | ||||||
| 126 | |||||||
| 127 | } // end namespace llvm | ||||||
| 128 | |||||||
| 129 | namespace { | ||||||
| 130 | |||||||
| 131 | /// Our current alias analysis API differentiates heavily between calls and | ||||||
| 132 | /// non-calls, and functions called on one usually assert on the other. | ||||||
| 133 | /// This class encapsulates the distinction to simplify other code that wants | ||||||
| 134 | /// "Memory affecting instructions and related data" to use as a key. | ||||||
| 135 | /// For example, this class is used as a densemap key in the use optimizer. | ||||||
| 136 | class MemoryLocOrCall { | ||||||
| 137 | public: | ||||||
| 138 | bool IsCall = false; | ||||||
| 139 | |||||||
| 140 | MemoryLocOrCall(MemoryUseOrDef *MUD) | ||||||
| 141 | : MemoryLocOrCall(MUD->getMemoryInst()) {} | ||||||
| 142 | MemoryLocOrCall(const MemoryUseOrDef *MUD) | ||||||
| 143 | : MemoryLocOrCall(MUD->getMemoryInst()) {} | ||||||
| 144 | |||||||
| 145 | MemoryLocOrCall(Instruction *Inst) { | ||||||
| 146 | if (auto *C = dyn_cast<CallBase>(Inst)) { | ||||||
| 147 | IsCall = true; | ||||||
| 148 | Call = C; | ||||||
| 149 | } else { | ||||||
| 150 | IsCall = false; | ||||||
| 151 | // There is no such thing as a memorylocation for a fence inst, and it is | ||||||
| 152 | // unique in that regard. | ||||||
| 153 | if (!isa<FenceInst>(Inst)) | ||||||
| 154 | Loc = MemoryLocation::get(Inst); | ||||||
| 155 | } | ||||||
| 156 | } | ||||||
| 157 | |||||||
| 158 | explicit MemoryLocOrCall(const MemoryLocation &Loc) : Loc(Loc) {} | ||||||
| 159 | |||||||
| 160 | const CallBase *getCall() const { | ||||||
| 161 | assert(IsCall)((void)0); | ||||||
| 162 | return Call; | ||||||
| 163 | } | ||||||
| 164 | |||||||
| 165 | MemoryLocation getLoc() const { | ||||||
| 166 | assert(!IsCall)((void)0); | ||||||
| 167 | return Loc; | ||||||
| 168 | } | ||||||
| 169 | |||||||
| 170 | bool operator==(const MemoryLocOrCall &Other) const { | ||||||
| 171 | if (IsCall != Other.IsCall) | ||||||
| 172 | return false; | ||||||
| 173 | |||||||
| 174 | if (!IsCall) | ||||||
| 175 | return Loc == Other.Loc; | ||||||
| 176 | |||||||
| 177 | if (Call->getCalledOperand() != Other.Call->getCalledOperand()) | ||||||
| 178 | return false; | ||||||
| 179 | |||||||
| 180 | return Call->arg_size() == Other.Call->arg_size() && | ||||||
| 181 | std::equal(Call->arg_begin(), Call->arg_end(), | ||||||
| 182 | Other.Call->arg_begin()); | ||||||
| 183 | } | ||||||
| 184 | |||||||
| 185 | private: | ||||||
| 186 | union { | ||||||
| 187 | const CallBase *Call; | ||||||
| 188 | MemoryLocation Loc; | ||||||
| 189 | }; | ||||||
| 190 | }; | ||||||
| 191 | |||||||
| 192 | } // end anonymous namespace | ||||||
| 193 | |||||||
| 194 | namespace llvm { | ||||||
| 195 | |||||||
| 196 | template <> struct DenseMapInfo<MemoryLocOrCall> { | ||||||
| 197 | static inline MemoryLocOrCall getEmptyKey() { | ||||||
| 198 | return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getEmptyKey()); | ||||||
| 199 | } | ||||||
| 200 | |||||||
| 201 | static inline MemoryLocOrCall getTombstoneKey() { | ||||||
| 202 | return MemoryLocOrCall(DenseMapInfo<MemoryLocation>::getTombstoneKey()); | ||||||
| 203 | } | ||||||
| 204 | |||||||
| 205 | static unsigned getHashValue(const MemoryLocOrCall &MLOC) { | ||||||
| 206 | if (!MLOC.IsCall) | ||||||
| 207 | return hash_combine( | ||||||
| 208 | MLOC.IsCall, | ||||||
| 209 | DenseMapInfo<MemoryLocation>::getHashValue(MLOC.getLoc())); | ||||||
| 210 | |||||||
| 211 | hash_code hash = | ||||||
| 212 | hash_combine(MLOC.IsCall, DenseMapInfo<const Value *>::getHashValue( | ||||||
| 213 | MLOC.getCall()->getCalledOperand())); | ||||||
| 214 | |||||||
| 215 | for (const Value *Arg : MLOC.getCall()->args()) | ||||||
| 216 | hash = hash_combine(hash, DenseMapInfo<const Value *>::getHashValue(Arg)); | ||||||
| 217 | return hash; | ||||||
| 218 | } | ||||||
| 219 | |||||||
| 220 | static bool isEqual(const MemoryLocOrCall &LHS, const MemoryLocOrCall &RHS) { | ||||||
| 221 | return LHS == RHS; | ||||||
| 222 | } | ||||||
| 223 | }; | ||||||
| 224 | |||||||
| 225 | } // end namespace llvm | ||||||
| 226 | |||||||
| 227 | /// This does one-way checks to see if Use could theoretically be hoisted above | ||||||
| 228 | /// MayClobber. This will not check the other way around. | ||||||
| 229 | /// | ||||||
| 230 | /// This assumes that, for the purposes of MemorySSA, Use comes directly after | ||||||
| 231 | /// MayClobber, with no potentially clobbering operations in between them. | ||||||
| 232 | /// (Where potentially clobbering ops are memory barriers, aliased stores, etc.) | ||||||
| 233 | static bool areLoadsReorderable(const LoadInst *Use, | ||||||
| 234 | const LoadInst *MayClobber) { | ||||||
| 235 | bool VolatileUse = Use->isVolatile(); | ||||||
| 236 | bool VolatileClobber = MayClobber->isVolatile(); | ||||||
| 237 | // Volatile operations may never be reordered with other volatile operations. | ||||||
| 238 | if (VolatileUse && VolatileClobber) | ||||||
| 239 | return false; | ||||||
| 240 | // Otherwise, volatile doesn't matter here. From the language reference: | ||||||
| 241 | // 'optimizers may change the order of volatile operations relative to | ||||||
| 242 | // non-volatile operations.'" | ||||||
| 243 | |||||||
| 244 | // If a load is seq_cst, it cannot be moved above other loads. If its ordering | ||||||
| 245 | // is weaker, it can be moved above other loads. We just need to be sure that | ||||||
| 246 | // MayClobber isn't an acquire load, because loads can't be moved above | ||||||
| 247 | // acquire loads. | ||||||
| 248 | // | ||||||
| 249 | // Note that this explicitly *does* allow the free reordering of monotonic (or | ||||||
| 250 | // weaker) loads of the same address. | ||||||
| 251 | bool SeqCstUse = Use->getOrdering() == AtomicOrdering::SequentiallyConsistent; | ||||||
| 252 | bool MayClobberIsAcquire = isAtLeastOrStrongerThan(MayClobber->getOrdering(), | ||||||
| 253 | AtomicOrdering::Acquire); | ||||||
| 254 | return !(SeqCstUse || MayClobberIsAcquire); | ||||||
| 255 | } | ||||||
| 256 | |||||||
| 257 | namespace { | ||||||
| 258 | |||||||
| 259 | struct ClobberAlias { | ||||||
| 260 | bool IsClobber; | ||||||
| 261 | Optional<AliasResult> AR; | ||||||
| 262 | }; | ||||||
| 263 | |||||||
| 264 | } // end anonymous namespace | ||||||
| 265 | |||||||
| 266 | // Return a pair of {IsClobber (bool), AR (AliasResult)}. It relies on AR being | ||||||
| 267 | // ignored if IsClobber = false. | ||||||
| 268 | template <typename AliasAnalysisType> | ||||||
| 269 | static ClobberAlias | ||||||
| 270 | instructionClobbersQuery(const MemoryDef *MD, const MemoryLocation &UseLoc, | ||||||
| 271 | const Instruction *UseInst, AliasAnalysisType &AA) { | ||||||
| 272 | Instruction *DefInst = MD->getMemoryInst(); | ||||||
| 273 | assert(DefInst && "Defining instruction not actually an instruction")((void)0); | ||||||
| 274 | Optional<AliasResult> AR; | ||||||
| 275 | |||||||
| 276 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) { | ||||||
| 277 | // These intrinsics will show up as affecting memory, but they are just | ||||||
| 278 | // markers, mostly. | ||||||
| 279 | // | ||||||
| 280 | // FIXME: We probably don't actually want MemorySSA to model these at all | ||||||
| 281 | // (including creating MemoryAccesses for them): we just end up inventing | ||||||
| 282 | // clobbers where they don't really exist at all. Please see D43269 for | ||||||
| 283 | // context. | ||||||
| 284 | switch (II->getIntrinsicID()) { | ||||||
| 285 | case Intrinsic::invariant_start: | ||||||
| 286 | case Intrinsic::invariant_end: | ||||||
| 287 | case Intrinsic::assume: | ||||||
| 288 | case Intrinsic::experimental_noalias_scope_decl: | ||||||
| 289 | return {false, AliasResult(AliasResult::NoAlias)}; | ||||||
| 290 | case Intrinsic::dbg_addr: | ||||||
| 291 | case Intrinsic::dbg_declare: | ||||||
| 292 | case Intrinsic::dbg_label: | ||||||
| 293 | case Intrinsic::dbg_value: | ||||||
| 294 | llvm_unreachable("debuginfo shouldn't have associated defs!")__builtin_unreachable(); | ||||||
| 295 | default: | ||||||
| 296 | break; | ||||||
| 297 | } | ||||||
| 298 | } | ||||||
| 299 | |||||||
| 300 | if (auto *CB = dyn_cast_or_null<CallBase>(UseInst)) { | ||||||
| 301 | ModRefInfo I = AA.getModRefInfo(DefInst, CB); | ||||||
| 302 | AR = isMustSet(I) ? AliasResult::MustAlias : AliasResult::MayAlias; | ||||||
| 303 | return {isModOrRefSet(I), AR}; | ||||||
| 304 | } | ||||||
| 305 | |||||||
| 306 | if (auto *DefLoad = dyn_cast<LoadInst>(DefInst)) | ||||||
| 307 | if (auto *UseLoad = dyn_cast_or_null<LoadInst>(UseInst)) | ||||||
| 308 | return {!areLoadsReorderable(UseLoad, DefLoad), | ||||||
| 309 | AliasResult(AliasResult::MayAlias)}; | ||||||
| 310 | |||||||
| 311 | ModRefInfo I = AA.getModRefInfo(DefInst, UseLoc); | ||||||
| 312 | AR = isMustSet(I) ? AliasResult::MustAlias : AliasResult::MayAlias; | ||||||
| 313 | return {isModSet(I), AR}; | ||||||
| 314 | } | ||||||
| 315 | |||||||
| 316 | template <typename AliasAnalysisType> | ||||||
| 317 | static ClobberAlias instructionClobbersQuery(MemoryDef *MD, | ||||||
| 318 | const MemoryUseOrDef *MU, | ||||||
| 319 | const MemoryLocOrCall &UseMLOC, | ||||||
| 320 | AliasAnalysisType &AA) { | ||||||
| 321 | // FIXME: This is a temporary hack to allow a single instructionClobbersQuery | ||||||
| 322 | // to exist while MemoryLocOrCall is pushed through places. | ||||||
| 323 | if (UseMLOC.IsCall) | ||||||
| 324 | return instructionClobbersQuery(MD, MemoryLocation(), MU->getMemoryInst(), | ||||||
| 325 | AA); | ||||||
| 326 | return instructionClobbersQuery(MD, UseMLOC.getLoc(), MU->getMemoryInst(), | ||||||
| 327 | AA); | ||||||
| 328 | } | ||||||
| 329 | |||||||
| 330 | // Return true when MD may alias MU, return false otherwise. | ||||||
| 331 | bool MemorySSAUtil::defClobbersUseOrDef(MemoryDef *MD, const MemoryUseOrDef *MU, | ||||||
| 332 | AliasAnalysis &AA) { | ||||||
| 333 | return instructionClobbersQuery(MD, MU, MemoryLocOrCall(MU), AA).IsClobber; | ||||||
| 334 | } | ||||||
| 335 | |||||||
| 336 | namespace { | ||||||
| 337 | |||||||
| 338 | struct UpwardsMemoryQuery { | ||||||
| 339 | // True if our original query started off as a call | ||||||
| 340 | bool IsCall = false; | ||||||
| 341 | // The pointer location we started the query with. This will be empty if | ||||||
| 342 | // IsCall is true. | ||||||
| 343 | MemoryLocation StartingLoc; | ||||||
| 344 | // This is the instruction we were querying about. | ||||||
| 345 | const Instruction *Inst = nullptr; | ||||||
| 346 | // The MemoryAccess we actually got called with, used to test local domination | ||||||
| 347 | const MemoryAccess *OriginalAccess = nullptr; | ||||||
| 348 | Optional<AliasResult> AR = AliasResult(AliasResult::MayAlias); | ||||||
| 349 | bool SkipSelfAccess = false; | ||||||
| 350 | |||||||
| 351 | UpwardsMemoryQuery() = default; | ||||||
| 352 | |||||||
| 353 | UpwardsMemoryQuery(const Instruction *Inst, const MemoryAccess *Access) | ||||||
| 354 | : IsCall(isa<CallBase>(Inst)), Inst(Inst), OriginalAccess(Access) { | ||||||
| 355 | if (!IsCall) | ||||||
| 356 | StartingLoc = MemoryLocation::get(Inst); | ||||||
| 357 | } | ||||||
| 358 | }; | ||||||
| 359 | |||||||
| 360 | } // end anonymous namespace | ||||||
| 361 | |||||||
| 362 | template <typename AliasAnalysisType> | ||||||
| 363 | static bool isUseTriviallyOptimizableToLiveOnEntry(AliasAnalysisType &AA, | ||||||
| 364 | const Instruction *I) { | ||||||
| 365 | // If the memory can't be changed, then loads of the memory can't be | ||||||
| 366 | // clobbered. | ||||||
| 367 | if (auto *LI = dyn_cast<LoadInst>(I)) | ||||||
| 368 | return I->hasMetadata(LLVMContext::MD_invariant_load) || | ||||||
| 369 | AA.pointsToConstantMemory(MemoryLocation::get(LI)); | ||||||
| 370 | return false; | ||||||
| 371 | } | ||||||
| 372 | |||||||
| 373 | /// Verifies that `Start` is clobbered by `ClobberAt`, and that nothing | ||||||
| 374 | /// inbetween `Start` and `ClobberAt` can clobbers `Start`. | ||||||
| 375 | /// | ||||||
| 376 | /// This is meant to be as simple and self-contained as possible. Because it | ||||||
| 377 | /// uses no cache, etc., it can be relatively expensive. | ||||||
| 378 | /// | ||||||
| 379 | /// \param Start The MemoryAccess that we want to walk from. | ||||||
| 380 | /// \param ClobberAt A clobber for Start. | ||||||
| 381 | /// \param StartLoc The MemoryLocation for Start. | ||||||
| 382 | /// \param MSSA The MemorySSA instance that Start and ClobberAt belong to. | ||||||
| 383 | /// \param Query The UpwardsMemoryQuery we used for our search. | ||||||
| 384 | /// \param AA The AliasAnalysis we used for our search. | ||||||
| 385 | /// \param AllowImpreciseClobber Always false, unless we do relaxed verify. | ||||||
| 386 | |||||||
| 387 | template <typename AliasAnalysisType> | ||||||
| 388 | LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) static void | ||||||
| 389 | checkClobberSanity(const MemoryAccess *Start, MemoryAccess *ClobberAt, | ||||||
| 390 | const MemoryLocation &StartLoc, const MemorySSA &MSSA, | ||||||
| 391 | const UpwardsMemoryQuery &Query, AliasAnalysisType &AA, | ||||||
| 392 | bool AllowImpreciseClobber = false) { | ||||||
| 393 | assert(MSSA.dominates(ClobberAt, Start) && "Clobber doesn't dominate start?")((void)0); | ||||||
| 394 | |||||||
| 395 | if (MSSA.isLiveOnEntryDef(Start)) { | ||||||
| 396 | assert(MSSA.isLiveOnEntryDef(ClobberAt) &&((void)0) | ||||||
| 397 | "liveOnEntry must clobber itself")((void)0); | ||||||
| 398 | return; | ||||||
| 399 | } | ||||||
| 400 | |||||||
| 401 | bool FoundClobber = false; | ||||||
| 402 | DenseSet<ConstMemoryAccessPair> VisitedPhis; | ||||||
| 403 | SmallVector<ConstMemoryAccessPair, 8> Worklist; | ||||||
| 404 | Worklist.emplace_back(Start, StartLoc); | ||||||
| 405 | // Walk all paths from Start to ClobberAt, while looking for clobbers. If one | ||||||
| 406 | // is found, complain. | ||||||
| 407 | while (!Worklist.empty()) { | ||||||
| 408 | auto MAP = Worklist.pop_back_val(); | ||||||
| 409 | // All we care about is that nothing from Start to ClobberAt clobbers Start. | ||||||
| 410 | // We learn nothing from revisiting nodes. | ||||||
| 411 | if (!VisitedPhis.insert(MAP).second) | ||||||
| 412 | continue; | ||||||
| 413 | |||||||
| 414 | for (const auto *MA : def_chain(MAP.first)) { | ||||||
| 415 | if (MA == ClobberAt) { | ||||||
| 416 | if (const auto *MD = dyn_cast<MemoryDef>(MA)) { | ||||||
| 417 | // instructionClobbersQuery isn't essentially free, so don't use `|=`, | ||||||
| 418 | // since it won't let us short-circuit. | ||||||
| 419 | // | ||||||
| 420 | // Also, note that this can't be hoisted out of the `Worklist` loop, | ||||||
| 421 | // since MD may only act as a clobber for 1 of N MemoryLocations. | ||||||
| 422 | FoundClobber = FoundClobber || MSSA.isLiveOnEntryDef(MD); | ||||||
| 423 | if (!FoundClobber) { | ||||||
| 424 | ClobberAlias CA = | ||||||
| 425 | instructionClobbersQuery(MD, MAP.second, Query.Inst, AA); | ||||||
| 426 | if (CA.IsClobber) { | ||||||
| 427 | FoundClobber = true; | ||||||
| 428 | // Not used: CA.AR; | ||||||
| 429 | } | ||||||
| 430 | } | ||||||
| 431 | } | ||||||
| 432 | break; | ||||||
| 433 | } | ||||||
| 434 | |||||||
| 435 | // We should never hit liveOnEntry, unless it's the clobber. | ||||||
| 436 | assert(!MSSA.isLiveOnEntryDef(MA) && "Hit liveOnEntry before clobber?")((void)0); | ||||||
| 437 | |||||||
| 438 | if (const auto *MD = dyn_cast<MemoryDef>(MA)) { | ||||||
| 439 | // If Start is a Def, skip self. | ||||||
| 440 | if (MD == Start) | ||||||
| 441 | continue; | ||||||
| 442 | |||||||
| 443 | assert(!instructionClobbersQuery(MD, MAP.second, Query.Inst, AA)((void)0) | ||||||
| 444 | .IsClobber &&((void)0) | ||||||
| 445 | "Found clobber before reaching ClobberAt!")((void)0); | ||||||
| 446 | continue; | ||||||
| 447 | } | ||||||
| 448 | |||||||
| 449 | if (const auto *MU = dyn_cast<MemoryUse>(MA)) { | ||||||
| 450 | (void)MU; | ||||||
| 451 | assert (MU == Start &&((void)0) | ||||||
| 452 | "Can only find use in def chain if Start is a use")((void)0); | ||||||
| 453 | continue; | ||||||
| 454 | } | ||||||
| 455 | |||||||
| 456 | assert(isa<MemoryPhi>(MA))((void)0); | ||||||
| 457 | |||||||
| 458 | // Add reachable phi predecessors | ||||||
| 459 | for (auto ItB = upward_defs_begin( | ||||||
| 460 | {const_cast<MemoryAccess *>(MA), MAP.second}, | ||||||
| 461 | MSSA.getDomTree()), | ||||||
| 462 | ItE = upward_defs_end(); | ||||||
| 463 | ItB != ItE; ++ItB) | ||||||
| 464 | if (MSSA.getDomTree().isReachableFromEntry(ItB.getPhiArgBlock())) | ||||||
| 465 | Worklist.emplace_back(*ItB); | ||||||
| 466 | } | ||||||
| 467 | } | ||||||
| 468 | |||||||
| 469 | // If the verify is done following an optimization, it's possible that | ||||||
| 470 | // ClobberAt was a conservative clobbering, that we can now infer is not a | ||||||
| 471 | // true clobbering access. Don't fail the verify if that's the case. | ||||||
| 472 | // We do have accesses that claim they're optimized, but could be optimized | ||||||
| 473 | // further. Updating all these can be expensive, so allow it for now (FIXME). | ||||||
| 474 | if (AllowImpreciseClobber) | ||||||
| 475 | return; | ||||||
| 476 | |||||||
| 477 | // If ClobberAt is a MemoryPhi, we can assume something above it acted as a | ||||||
| 478 | // clobber. Otherwise, `ClobberAt` should've acted as a clobber at some point. | ||||||
| 479 | assert((isa<MemoryPhi>(ClobberAt) || FoundClobber) &&((void)0) | ||||||
| 480 | "ClobberAt never acted as a clobber")((void)0); | ||||||
| 481 | } | ||||||
| 482 | |||||||
| 483 | namespace { | ||||||
| 484 | |||||||
| 485 | /// Our algorithm for walking (and trying to optimize) clobbers, all wrapped up | ||||||
| 486 | /// in one class. | ||||||
| 487 | template <class AliasAnalysisType> class ClobberWalker { | ||||||
| 488 | /// Save a few bytes by using unsigned instead of size_t. | ||||||
| 489 | using ListIndex = unsigned; | ||||||
| 490 | |||||||
| 491 | /// Represents a span of contiguous MemoryDefs, potentially ending in a | ||||||
| 492 | /// MemoryPhi. | ||||||
| 493 | struct DefPath { | ||||||
| 494 | MemoryLocation Loc; | ||||||
| 495 | // Note that, because we always walk in reverse, Last will always dominate | ||||||
| 496 | // First. Also note that First and Last are inclusive. | ||||||
| 497 | MemoryAccess *First; | ||||||
| 498 | MemoryAccess *Last; | ||||||
| 499 | Optional<ListIndex> Previous; | ||||||
| 500 | |||||||
| 501 | DefPath(const MemoryLocation &Loc, MemoryAccess *First, MemoryAccess *Last, | ||||||
| 502 | Optional<ListIndex> Previous) | ||||||
| 503 | : Loc(Loc), First(First), Last(Last), Previous(Previous) {} | ||||||
| 504 | |||||||
| 505 | DefPath(const MemoryLocation &Loc, MemoryAccess *Init, | ||||||
| 506 | Optional<ListIndex> Previous) | ||||||
| 507 | : DefPath(Loc, Init, Init, Previous) {} | ||||||
| 508 | }; | ||||||
| 509 | |||||||
| 510 | const MemorySSA &MSSA; | ||||||
| 511 | AliasAnalysisType &AA; | ||||||
| 512 | DominatorTree &DT; | ||||||
| 513 | UpwardsMemoryQuery *Query; | ||||||
| 514 | unsigned *UpwardWalkLimit; | ||||||
| 515 | |||||||
| 516 | // Phi optimization bookkeeping: | ||||||
| 517 | // List of DefPath to process during the current phi optimization walk. | ||||||
| 518 | SmallVector<DefPath, 32> Paths; | ||||||
| 519 | // List of visited <Access, Location> pairs; we can skip paths already | ||||||
| 520 | // visited with the same memory location. | ||||||
| 521 | DenseSet<ConstMemoryAccessPair> VisitedPhis; | ||||||
| 522 | // Record if phi translation has been performed during the current phi | ||||||
| 523 | // optimization walk, as merging alias results after phi translation can | ||||||
| 524 | // yield incorrect results. Context in PR46156. | ||||||
| 525 | bool PerformedPhiTranslation = false; | ||||||
| 526 | |||||||
| 527 | /// Find the nearest def or phi that `From` can legally be optimized to. | ||||||
| 528 | const MemoryAccess *getWalkTarget(const MemoryPhi *From) const { | ||||||
| 529 | assert(From->getNumOperands() && "Phi with no operands?")((void)0); | ||||||
| 530 | |||||||
| 531 | BasicBlock *BB = From->getBlock(); | ||||||
| 532 | MemoryAccess *Result = MSSA.getLiveOnEntryDef(); | ||||||
| 533 | DomTreeNode *Node = DT.getNode(BB); | ||||||
| 534 | while ((Node = Node->getIDom())) { | ||||||
| 535 | auto *Defs = MSSA.getBlockDefs(Node->getBlock()); | ||||||
| 536 | if (Defs) | ||||||
| 537 | return &*Defs->rbegin(); | ||||||
| 538 | } | ||||||
| 539 | return Result; | ||||||
| 540 | } | ||||||
| 541 | |||||||
| 542 | /// Result of calling walkToPhiOrClobber. | ||||||
| 543 | struct UpwardsWalkResult { | ||||||
| 544 | /// The "Result" of the walk. Either a clobber, the last thing we walked, or | ||||||
| 545 | /// both. Include alias info when clobber found. | ||||||
| 546 | MemoryAccess *Result; | ||||||
| 547 | bool IsKnownClobber; | ||||||
| 548 | Optional<AliasResult> AR; | ||||||
| 549 | }; | ||||||
| 550 | |||||||
| 551 | /// Walk to the next Phi or Clobber in the def chain starting at Desc.Last. | ||||||
| 552 | /// This will update Desc.Last as it walks. It will (optionally) also stop at | ||||||
| 553 | /// StopAt. | ||||||
| 554 | /// | ||||||
| 555 | /// This does not test for whether StopAt is a clobber | ||||||
| 556 | UpwardsWalkResult | ||||||
| 557 | walkToPhiOrClobber(DefPath &Desc, const MemoryAccess *StopAt = nullptr, | ||||||
| 558 | const MemoryAccess *SkipStopAt = nullptr) const { | ||||||
| 559 | assert(!isa<MemoryUse>(Desc.Last) && "Uses don't exist in my world")((void)0); | ||||||
| 560 | assert(UpwardWalkLimit && "Need a valid walk limit")((void)0); | ||||||
| 561 | bool LimitAlreadyReached = false; | ||||||
| 562 | // (*UpwardWalkLimit) may be 0 here, due to the loop in tryOptimizePhi. Set | ||||||
| 563 | // it to 1. This will not do any alias() calls. It either returns in the | ||||||
| 564 | // first iteration in the loop below, or is set back to 0 if all def chains | ||||||
| 565 | // are free of MemoryDefs. | ||||||
| 566 | if (!*UpwardWalkLimit) { | ||||||
| 567 | *UpwardWalkLimit = 1; | ||||||
| 568 | LimitAlreadyReached = true; | ||||||
| 569 | } | ||||||
| 570 | |||||||
| 571 | for (MemoryAccess *Current : def_chain(Desc.Last)) { | ||||||
| 572 | Desc.Last = Current; | ||||||
| 573 | if (Current == StopAt || Current == SkipStopAt) | ||||||
| 574 | return {Current, false, AliasResult(AliasResult::MayAlias)}; | ||||||
| 575 | |||||||
| 576 | if (auto *MD = dyn_cast<MemoryDef>(Current)) { | ||||||
| 577 | if (MSSA.isLiveOnEntryDef(MD)) | ||||||
| 578 | return {MD, true, AliasResult(AliasResult::MustAlias)}; | ||||||
| 579 | |||||||
| 580 | if (!--*UpwardWalkLimit) | ||||||
| 581 | return {Current, true, AliasResult(AliasResult::MayAlias)}; | ||||||
| 582 | |||||||
| 583 | ClobberAlias CA = | ||||||
| 584 | instructionClobbersQuery(MD, Desc.Loc, Query->Inst, AA); | ||||||
| 585 | if (CA.IsClobber) | ||||||
| 586 | return {MD, true, CA.AR}; | ||||||
| 587 | } | ||||||
| 588 | } | ||||||
| 589 | |||||||
| 590 | if (LimitAlreadyReached) | ||||||
| 591 | *UpwardWalkLimit = 0; | ||||||
| 592 | |||||||
| 593 | assert(isa<MemoryPhi>(Desc.Last) &&((void)0) | ||||||
| 594 | "Ended at a non-clobber that's not a phi?")((void)0); | ||||||
| 595 | return {Desc.Last, false, AliasResult(AliasResult::MayAlias)}; | ||||||
| 596 | } | ||||||
| 597 | |||||||
| 598 | void addSearches(MemoryPhi *Phi, SmallVectorImpl<ListIndex> &PausedSearches, | ||||||
| 599 | ListIndex PriorNode) { | ||||||
| 600 | auto UpwardDefsBegin = upward_defs_begin({Phi, Paths[PriorNode].Loc}, DT, | ||||||
| 601 | &PerformedPhiTranslation); | ||||||
| 602 | auto UpwardDefs = make_range(UpwardDefsBegin, upward_defs_end()); | ||||||
| 603 | for (const MemoryAccessPair &P : UpwardDefs) { | ||||||
| 604 | PausedSearches.push_back(Paths.size()); | ||||||
| 605 | Paths.emplace_back(P.second, P.first, PriorNode); | ||||||
| 606 | } | ||||||
| 607 | } | ||||||
| 608 | |||||||
| 609 | /// Represents a search that terminated after finding a clobber. This clobber | ||||||
| 610 | /// may or may not be present in the path of defs from LastNode..SearchStart, | ||||||
| 611 | /// since it may have been retrieved from cache. | ||||||
| 612 | struct TerminatedPath { | ||||||
| 613 | MemoryAccess *Clobber; | ||||||
| 614 | ListIndex LastNode; | ||||||
| 615 | }; | ||||||
| 616 | |||||||
| 617 | /// Get an access that keeps us from optimizing to the given phi. | ||||||
| 618 | /// | ||||||
| 619 | /// PausedSearches is an array of indices into the Paths array. Its incoming | ||||||
| 620 | /// value is the indices of searches that stopped at the last phi optimization | ||||||
| 621 | /// target. It's left in an unspecified state. | ||||||
| 622 | /// | ||||||
| 623 | /// If this returns None, NewPaused is a vector of searches that terminated | ||||||
| 624 | /// at StopWhere. Otherwise, NewPaused is left in an unspecified state. | ||||||
| 625 | Optional<TerminatedPath> | ||||||
| 626 | getBlockingAccess(const MemoryAccess *StopWhere, | ||||||
| 627 | SmallVectorImpl<ListIndex> &PausedSearches, | ||||||
| 628 | SmallVectorImpl<ListIndex> &NewPaused, | ||||||
| 629 | SmallVectorImpl<TerminatedPath> &Terminated) { | ||||||
| 630 | assert(!PausedSearches.empty() && "No searches to continue?")((void)0); | ||||||
| 631 | |||||||
| 632 | // BFS vs DFS really doesn't make a difference here, so just do a DFS with | ||||||
| 633 | // PausedSearches as our stack. | ||||||
| 634 | while (!PausedSearches.empty()) { | ||||||
| 635 | ListIndex PathIndex = PausedSearches.pop_back_val(); | ||||||
| 636 | DefPath &Node = Paths[PathIndex]; | ||||||
| 637 | |||||||
| 638 | // If we've already visited this path with this MemoryLocation, we don't | ||||||
| 639 | // need to do so again. | ||||||
| 640 | // | ||||||
| 641 | // NOTE: That we just drop these paths on the ground makes caching | ||||||
| 642 | // behavior sporadic. e.g. given a diamond: | ||||||
| 643 | // A | ||||||
| 644 | // B C | ||||||
| 645 | // D | ||||||
| 646 | // | ||||||
| 647 | // ...If we walk D, B, A, C, we'll only cache the result of phi | ||||||
| 648 | // optimization for A, B, and D; C will be skipped because it dies here. | ||||||
| 649 | // This arguably isn't the worst thing ever, since: | ||||||
| 650 | // - We generally query things in a top-down order, so if we got below D | ||||||
| 651 | // without needing cache entries for {C, MemLoc}, then chances are | ||||||
| 652 | // that those cache entries would end up ultimately unused. | ||||||
| 653 | // - We still cache things for A, so C only needs to walk up a bit. | ||||||
| 654 | // If this behavior becomes problematic, we can fix without a ton of extra | ||||||
| 655 | // work. | ||||||
| 656 | if (!VisitedPhis.insert({Node.Last, Node.Loc}).second) { | ||||||
| 657 | if (PerformedPhiTranslation) { | ||||||
| 658 | // If visiting this path performed Phi translation, don't continue, | ||||||
| 659 | // since it may not be correct to merge results from two paths if one | ||||||
| 660 | // relies on the phi translation. | ||||||
| 661 | TerminatedPath Term{Node.Last, PathIndex}; | ||||||
| 662 | return Term; | ||||||
| 663 | } | ||||||
| 664 | continue; | ||||||
| 665 | } | ||||||
| 666 | |||||||
| 667 | const MemoryAccess *SkipStopWhere = nullptr; | ||||||
| 668 | if (Query->SkipSelfAccess && Node.Loc == Query->StartingLoc) { | ||||||
| 669 | assert(isa<MemoryDef>(Query->OriginalAccess))((void)0); | ||||||
| 670 | SkipStopWhere = Query->OriginalAccess; | ||||||
| 671 | } | ||||||
| 672 | |||||||
| 673 | UpwardsWalkResult Res = walkToPhiOrClobber(Node, | ||||||
| 674 | /*StopAt=*/StopWhere, | ||||||
| 675 | /*SkipStopAt=*/SkipStopWhere); | ||||||
| 676 | if (Res.IsKnownClobber) { | ||||||
| 677 | assert(Res.Result != StopWhere && Res.Result != SkipStopWhere)((void)0); | ||||||
| 678 | |||||||
| 679 | // If this wasn't a cache hit, we hit a clobber when walking. That's a | ||||||
| 680 | // failure. | ||||||
| 681 | TerminatedPath Term{Res.Result, PathIndex}; | ||||||
| 682 | if (!MSSA.dominates(Res.Result, StopWhere)) | ||||||
| 683 | return Term; | ||||||
| 684 | |||||||
| 685 | // Otherwise, it's a valid thing to potentially optimize to. | ||||||
| 686 | Terminated.push_back(Term); | ||||||
| 687 | continue; | ||||||
| 688 | } | ||||||
| 689 | |||||||
| 690 | if (Res.Result == StopWhere || Res.Result == SkipStopWhere) { | ||||||
| 691 | // We've hit our target. Save this path off for if we want to continue | ||||||
| 692 | // walking. If we are in the mode of skipping the OriginalAccess, and | ||||||
| 693 | // we've reached back to the OriginalAccess, do not save path, we've | ||||||
| 694 | // just looped back to self. | ||||||
| 695 | if (Res.Result != SkipStopWhere) | ||||||
| 696 | NewPaused.push_back(PathIndex); | ||||||
| 697 | continue; | ||||||
| 698 | } | ||||||
| 699 | |||||||
| 700 | assert(!MSSA.isLiveOnEntryDef(Res.Result) && "liveOnEntry is a clobber")((void)0); | ||||||
| 701 | addSearches(cast<MemoryPhi>(Res.Result), PausedSearches, PathIndex); | ||||||
| 702 | } | ||||||
| 703 | |||||||
| 704 | return None; | ||||||
| 705 | } | ||||||
| 706 | |||||||
| 707 | template <typename T, typename Walker> | ||||||
| 708 | struct generic_def_path_iterator | ||||||
| 709 | : public iterator_facade_base<generic_def_path_iterator<T, Walker>, | ||||||
| 710 | std::forward_iterator_tag, T *> { | ||||||
| 711 | generic_def_path_iterator() {} | ||||||
| 712 | generic_def_path_iterator(Walker *W, ListIndex N) : W(W), N(N) {} | ||||||
| 713 | |||||||
| 714 | T &operator*() const { return curNode(); } | ||||||
| 715 | |||||||
| 716 | generic_def_path_iterator &operator++() { | ||||||
| 717 | N = curNode().Previous; | ||||||
| 718 | return *this; | ||||||
| 719 | } | ||||||
| 720 | |||||||
| 721 | bool operator==(const generic_def_path_iterator &O) const { | ||||||
| 722 | if (N.hasValue() != O.N.hasValue()) | ||||||
| 723 | return false; | ||||||
| 724 | return !N.hasValue() || *N == *O.N; | ||||||
| 725 | } | ||||||
| 726 | |||||||
| 727 | private: | ||||||
| 728 | T &curNode() const { return W->Paths[*N]; } | ||||||
| 729 | |||||||
| 730 | Walker *W = nullptr; | ||||||
| 731 | Optional<ListIndex> N = None; | ||||||
| 732 | }; | ||||||
| 733 | |||||||
| 734 | using def_path_iterator = generic_def_path_iterator<DefPath, ClobberWalker>; | ||||||
| 735 | using const_def_path_iterator = | ||||||
| 736 | generic_def_path_iterator<const DefPath, const ClobberWalker>; | ||||||
| 737 | |||||||
| 738 | iterator_range<def_path_iterator> def_path(ListIndex From) { | ||||||
| 739 | return make_range(def_path_iterator(this, From), def_path_iterator()); | ||||||
| 740 | } | ||||||
| 741 | |||||||
| 742 | iterator_range<const_def_path_iterator> const_def_path(ListIndex From) const { | ||||||
| 743 | return make_range(const_def_path_iterator(this, From), | ||||||
| 744 | const_def_path_iterator()); | ||||||
| 745 | } | ||||||
| 746 | |||||||
| 747 | struct OptznResult { | ||||||
| 748 | /// The path that contains our result. | ||||||
| 749 | TerminatedPath PrimaryClobber; | ||||||
| 750 | /// The paths that we can legally cache back from, but that aren't | ||||||
| 751 | /// necessarily the result of the Phi optimization. | ||||||
| 752 | SmallVector<TerminatedPath, 4> OtherClobbers; | ||||||
| 753 | }; | ||||||
| 754 | |||||||
| 755 | ListIndex defPathIndex(const DefPath &N) const { | ||||||
| 756 | // The assert looks nicer if we don't need to do &N | ||||||
| 757 | const DefPath *NP = &N; | ||||||
| 758 | assert(!Paths.empty() && NP >= &Paths.front() && NP <= &Paths.back() &&((void)0) | ||||||
| 759 | "Out of bounds DefPath!")((void)0); | ||||||
| 760 | return NP - &Paths.front(); | ||||||
| 761 | } | ||||||
| 762 | |||||||
| 763 | /// Try to optimize a phi as best as we can. Returns a SmallVector of Paths | ||||||
| 764 | /// that act as legal clobbers. Note that this won't return *all* clobbers. | ||||||
| 765 | /// | ||||||
| 766 | /// Phi optimization algorithm tl;dr: | ||||||
| 767 | /// - Find the earliest def/phi, A, we can optimize to | ||||||
| 768 | /// - Find if all paths from the starting memory access ultimately reach A | ||||||
| 769 | /// - If not, optimization isn't possible. | ||||||
| 770 | /// - Otherwise, walk from A to another clobber or phi, A'. | ||||||
| 771 | /// - If A' is a def, we're done. | ||||||
| 772 | /// - If A' is a phi, try to optimize it. | ||||||
| 773 | /// | ||||||
| 774 | /// A path is a series of {MemoryAccess, MemoryLocation} pairs. A path | ||||||
| 775 | /// terminates when a MemoryAccess that clobbers said MemoryLocation is found. | ||||||
| 776 | OptznResult tryOptimizePhi(MemoryPhi *Phi, MemoryAccess *Start, | ||||||
| 777 | const MemoryLocation &Loc) { | ||||||
| 778 | assert(Paths.empty() && VisitedPhis.empty() && !PerformedPhiTranslation &&((void)0) | ||||||
| 779 | "Reset the optimization state.")((void)0); | ||||||
| 780 | |||||||
| 781 | Paths.emplace_back(Loc, Start, Phi, None); | ||||||
| 782 | // Stores how many "valid" optimization nodes we had prior to calling | ||||||
| 783 | // addSearches/getBlockingAccess. Necessary for caching if we had a blocker. | ||||||
| 784 | auto PriorPathsSize = Paths.size(); | ||||||
| 785 | |||||||
| 786 | SmallVector<ListIndex, 16> PausedSearches; | ||||||
| 787 | SmallVector<ListIndex, 8> NewPaused; | ||||||
| 788 | SmallVector<TerminatedPath, 4> TerminatedPaths; | ||||||
| 789 | |||||||
| 790 | addSearches(Phi, PausedSearches, 0); | ||||||
| 791 | |||||||
| 792 | // Moves the TerminatedPath with the "most dominated" Clobber to the end of | ||||||
| 793 | // Paths. | ||||||
| 794 | auto MoveDominatedPathToEnd = [&](SmallVectorImpl<TerminatedPath> &Paths) { | ||||||
| 795 | assert(!Paths.empty() && "Need a path to move")((void)0); | ||||||
| 796 | auto Dom = Paths.begin(); | ||||||
| 797 | for (auto I = std::next(Dom), E = Paths.end(); I != E; ++I) | ||||||
| 798 | if (!MSSA.dominates(I->Clobber, Dom->Clobber)) | ||||||
| 799 | Dom = I; | ||||||
| 800 | auto Last = Paths.end() - 1; | ||||||
| 801 | if (Last != Dom) | ||||||
| 802 | std::iter_swap(Last, Dom); | ||||||
| 803 | }; | ||||||
| 804 | |||||||
| 805 | MemoryPhi *Current = Phi; | ||||||
| 806 | while (true) { | ||||||
| 807 | assert(!MSSA.isLiveOnEntryDef(Current) &&((void)0) | ||||||
| 808 | "liveOnEntry wasn't treated as a clobber?")((void)0); | ||||||
| 809 | |||||||
| 810 | const auto *Target = getWalkTarget(Current); | ||||||
| 811 | // If a TerminatedPath doesn't dominate Target, then it wasn't a legal | ||||||
| 812 | // optimization for the prior phi. | ||||||
| 813 | assert(all_of(TerminatedPaths, [&](const TerminatedPath &P) {((void)0) | ||||||
| 814 | return MSSA.dominates(P.Clobber, Target);((void)0) | ||||||
| 815 | }))((void)0); | ||||||
| 816 | |||||||
| 817 | // FIXME: This is broken, because the Blocker may be reported to be | ||||||
| 818 | // liveOnEntry, and we'll happily wait for that to disappear (read: never) | ||||||
| 819 | // For the moment, this is fine, since we do nothing with blocker info. | ||||||
| 820 | if (Optional<TerminatedPath> Blocker = getBlockingAccess( | ||||||
| 821 | Target, PausedSearches, NewPaused, TerminatedPaths)) { | ||||||
| 822 | |||||||
| 823 | // Find the node we started at. We can't search based on N->Last, since | ||||||
| 824 | // we may have gone around a loop with a different MemoryLocation. | ||||||
| 825 | auto Iter = find_if(def_path(Blocker->LastNode), [&](const DefPath &N) { | ||||||
| 826 | return defPathIndex(N) < PriorPathsSize; | ||||||
| 827 | }); | ||||||
| 828 | assert(Iter != def_path_iterator())((void)0); | ||||||
| 829 | |||||||
| 830 | DefPath &CurNode = *Iter; | ||||||
| 831 | assert(CurNode.Last == Current)((void)0); | ||||||
| 832 | |||||||
| 833 | // Two things: | ||||||
| 834 | // A. We can't reliably cache all of NewPaused back. Consider a case | ||||||
| 835 | // where we have two paths in NewPaused; one of which can't optimize | ||||||
| 836 | // above this phi, whereas the other can. If we cache the second path | ||||||
| 837 | // back, we'll end up with suboptimal cache entries. We can handle | ||||||
| 838 | // cases like this a bit better when we either try to find all | ||||||
| 839 | // clobbers that block phi optimization, or when our cache starts | ||||||
| 840 | // supporting unfinished searches. | ||||||
| 841 | // B. We can't reliably cache TerminatedPaths back here without doing | ||||||
| 842 | // extra checks; consider a case like: | ||||||
| 843 | // T | ||||||
| 844 | // / \ | ||||||
| 845 | // D C | ||||||
| 846 | // \ / | ||||||
| 847 | // S | ||||||
| 848 | // Where T is our target, C is a node with a clobber on it, D is a | ||||||
| 849 | // diamond (with a clobber *only* on the left or right node, N), and | ||||||
| 850 | // S is our start. Say we walk to D, through the node opposite N | ||||||
| 851 | // (read: ignoring the clobber), and see a cache entry in the top | ||||||
| 852 | // node of D. That cache entry gets put into TerminatedPaths. We then | ||||||
| 853 | // walk up to C (N is later in our worklist), find the clobber, and | ||||||
| 854 | // quit. If we append TerminatedPaths to OtherClobbers, we'll cache | ||||||
| 855 | // the bottom part of D to the cached clobber, ignoring the clobber | ||||||
| 856 | // in N. Again, this problem goes away if we start tracking all | ||||||
| 857 | // blockers for a given phi optimization. | ||||||
| 858 | TerminatedPath Result{CurNode.Last, defPathIndex(CurNode)}; | ||||||
| 859 | return {Result, {}}; | ||||||
| 860 | } | ||||||
| 861 | |||||||
| 862 | // If there's nothing left to search, then all paths led to valid clobbers | ||||||
| 863 | // that we got from our cache; pick the nearest to the start, and allow | ||||||
| 864 | // the rest to be cached back. | ||||||
| 865 | if (NewPaused.empty()) { | ||||||
| 866 | MoveDominatedPathToEnd(TerminatedPaths); | ||||||
| 867 | TerminatedPath Result = TerminatedPaths.pop_back_val(); | ||||||
| 868 | return {Result, std::move(TerminatedPaths)}; | ||||||
| 869 | } | ||||||
| 870 | |||||||
| 871 | MemoryAccess *DefChainEnd = nullptr; | ||||||
| 872 | SmallVector<TerminatedPath, 4> Clobbers; | ||||||
| 873 | for (ListIndex Paused : NewPaused) { | ||||||
| 874 | UpwardsWalkResult WR = walkToPhiOrClobber(Paths[Paused]); | ||||||
| 875 | if (WR.IsKnownClobber) | ||||||
| 876 | Clobbers.push_back({WR.Result, Paused}); | ||||||
| 877 | else | ||||||
| 878 | // Micro-opt: If we hit the end of the chain, save it. | ||||||
| 879 | DefChainEnd = WR.Result; | ||||||
| 880 | } | ||||||
| 881 | |||||||
| 882 | if (!TerminatedPaths.empty()) { | ||||||
| 883 | // If we couldn't find the dominating phi/liveOnEntry in the above loop, | ||||||
| 884 | // do it now. | ||||||
| 885 | if (!DefChainEnd
| ||||||
| 886 | for (auto *MA : def_chain(const_cast<MemoryAccess *>(Target))) | ||||||
| 887 | DefChainEnd = MA; | ||||||
| 888 | assert(DefChainEnd && "Failed to find dominating phi/liveOnEntry")((void)0); | ||||||
| 889 | |||||||
| 890 | // If any of the terminated paths don't dominate the phi we'll try to | ||||||
| 891 | // optimize, we need to figure out what they are and quit. | ||||||
| 892 | const BasicBlock *ChainBB = DefChainEnd->getBlock(); | ||||||
| |||||||
| 893 | for (const TerminatedPath &TP : TerminatedPaths) { | ||||||
| 894 | // Because we know that DefChainEnd is as "high" as we can go, we | ||||||
| 895 | // don't need local dominance checks; BB dominance is sufficient. | ||||||
| 896 | if (DT.dominates(ChainBB, TP.Clobber->getBlock())) | ||||||
| 897 | Clobbers.push_back(TP); | ||||||
| 898 | } | ||||||
| 899 | } | ||||||
| 900 | |||||||
| 901 | // If we have clobbers in the def chain, find the one closest to Current | ||||||
| 902 | // and quit. | ||||||
| 903 | if (!Clobbers.empty()) { | ||||||
| 904 | MoveDominatedPathToEnd(Clobbers); | ||||||
| 905 | TerminatedPath Result = Clobbers.pop_back_val(); | ||||||
| 906 | return {Result, std::move(Clobbers)}; | ||||||
| 907 | } | ||||||
| 908 | |||||||
| 909 | assert(all_of(NewPaused,((void)0) | ||||||
| 910 | [&](ListIndex I) { return Paths[I].Last == DefChainEnd; }))((void)0); | ||||||
| 911 | |||||||
| 912 | // Because liveOnEntry is a clobber, this must be a phi. | ||||||
| 913 | auto *DefChainPhi = cast<MemoryPhi>(DefChainEnd); | ||||||
| 914 | |||||||
| 915 | PriorPathsSize = Paths.size(); | ||||||
| 916 | PausedSearches.clear(); | ||||||
| 917 | for (ListIndex I : NewPaused) | ||||||
| 918 | addSearches(DefChainPhi, PausedSearches, I); | ||||||
| 919 | NewPaused.clear(); | ||||||
| 920 | |||||||
| 921 | Current = DefChainPhi; | ||||||
| 922 | } | ||||||
| 923 | } | ||||||
| 924 | |||||||
| 925 | void verifyOptResult(const OptznResult &R) const { | ||||||
| 926 | assert(all_of(R.OtherClobbers, [&](const TerminatedPath &P) {((void)0) | ||||||
| 927 | return MSSA.dominates(P.Clobber, R.PrimaryClobber.Clobber);((void)0) | ||||||
| 928 | }))((void)0); | ||||||
| 929 | } | ||||||
| 930 | |||||||
| 931 | void resetPhiOptznState() { | ||||||
| 932 | Paths.clear(); | ||||||
| 933 | VisitedPhis.clear(); | ||||||
| 934 | PerformedPhiTranslation = false; | ||||||
| 935 | } | ||||||
| 936 | |||||||
| 937 | public: | ||||||
| 938 | ClobberWalker(const MemorySSA &MSSA, AliasAnalysisType &AA, DominatorTree &DT) | ||||||
| 939 | : MSSA(MSSA), AA(AA), DT(DT) {} | ||||||
| 940 | |||||||
| 941 | AliasAnalysisType *getAA() { return &AA; } | ||||||
| 942 | /// Finds the nearest clobber for the given query, optimizing phis if | ||||||
| 943 | /// possible. | ||||||
| 944 | MemoryAccess *findClobber(MemoryAccess *Start, UpwardsMemoryQuery &Q, | ||||||
| 945 | unsigned &UpWalkLimit) { | ||||||
| 946 | Query = &Q; | ||||||
| 947 | UpwardWalkLimit = &UpWalkLimit; | ||||||
| 948 | // Starting limit must be > 0. | ||||||
| 949 | if (!UpWalkLimit) | ||||||
| 950 | UpWalkLimit++; | ||||||
| 951 | |||||||
| 952 | MemoryAccess *Current = Start; | ||||||
| 953 | // This walker pretends uses don't exist. If we're handed one, silently grab | ||||||
| 954 | // its def. (This has the nice side-effect of ensuring we never cache uses) | ||||||
| 955 | if (auto *MU
| ||||||
| 956 | Current = MU->getDefiningAccess(); | ||||||
| 957 | |||||||
| 958 | DefPath FirstDesc(Q.StartingLoc, Current, Current, None); | ||||||
| 959 | // Fast path for the overly-common case (no crazy phi optimization | ||||||
| 960 | // necessary) | ||||||
| 961 | UpwardsWalkResult WalkResult = walkToPhiOrClobber(FirstDesc); | ||||||
| 962 | MemoryAccess *Result; | ||||||
| 963 | if (WalkResult.IsKnownClobber
| ||||||
| 964 | Result = WalkResult.Result; | ||||||
| 965 | Q.AR = WalkResult.AR; | ||||||
| 966 | } else { | ||||||
| 967 | OptznResult OptRes = tryOptimizePhi(cast<MemoryPhi>(FirstDesc.Last), | ||||||
| 968 | Current, Q.StartingLoc); | ||||||
| 969 | verifyOptResult(OptRes); | ||||||
| 970 | resetPhiOptznState(); | ||||||
| 971 | Result = OptRes.PrimaryClobber.Clobber; | ||||||
| 972 | } | ||||||
| 973 | |||||||
| 974 | #ifdef EXPENSIVE_CHECKS | ||||||
| 975 | if (!Q.SkipSelfAccess && *UpwardWalkLimit > 0) | ||||||
| 976 | checkClobberSanity(Current, Result, Q.StartingLoc, MSSA, Q, AA); | ||||||
| 977 | #endif | ||||||
| 978 | return Result; | ||||||
| 979 | } | ||||||
| 980 | }; | ||||||
| 981 | |||||||
| 982 | struct RenamePassData { | ||||||
| 983 | DomTreeNode *DTN; | ||||||
| 984 | DomTreeNode::const_iterator ChildIt; | ||||||
| 985 | MemoryAccess *IncomingVal; | ||||||
| 986 | |||||||
| 987 | RenamePassData(DomTreeNode *D, DomTreeNode::const_iterator It, | ||||||
| 988 | MemoryAccess *M) | ||||||
| 989 | : DTN(D), ChildIt(It), IncomingVal(M) {} | ||||||
| 990 | |||||||
| 991 | void swap(RenamePassData &RHS) { | ||||||
| 992 | std::swap(DTN, RHS.DTN); | ||||||
| 993 | std::swap(ChildIt, RHS.ChildIt); | ||||||
| 994 | std::swap(IncomingVal, RHS.IncomingVal); | ||||||
| 995 | } | ||||||
| 996 | }; | ||||||
| 997 | |||||||
| 998 | } // end anonymous namespace | ||||||
| 999 | |||||||
| 1000 | namespace llvm { | ||||||
| 1001 | |||||||
| 1002 | template <class AliasAnalysisType> class MemorySSA::ClobberWalkerBase { | ||||||
| 1003 | ClobberWalker<AliasAnalysisType> Walker; | ||||||
| 1004 | MemorySSA *MSSA; | ||||||
| 1005 | |||||||
| 1006 | public: | ||||||
| 1007 | ClobberWalkerBase(MemorySSA *M, AliasAnalysisType *A, DominatorTree *D) | ||||||
| 1008 | : Walker(*M, *A, *D), MSSA(M) {} | ||||||
| 1009 | |||||||
| 1010 | MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *, | ||||||
| 1011 | const MemoryLocation &, | ||||||
| 1012 | unsigned &); | ||||||
| 1013 | // Third argument (bool), defines whether the clobber search should skip the | ||||||
| 1014 | // original queried access. If true, there will be a follow-up query searching | ||||||
| 1015 | // for a clobber access past "self". Note that the Optimized access is not | ||||||
| 1016 | // updated if a new clobber is found by this SkipSelf search. If this | ||||||
| 1017 | // additional query becomes heavily used we may decide to cache the result. | ||||||
| 1018 | // Walker instantiations will decide how to set the SkipSelf bool. | ||||||
| 1019 | MemoryAccess *getClobberingMemoryAccessBase(MemoryAccess *, unsigned &, bool); | ||||||
| 1020 | }; | ||||||
| 1021 | |||||||
| 1022 | /// A MemorySSAWalker that does AA walks to disambiguate accesses. It no | ||||||
| 1023 | /// longer does caching on its own, but the name has been retained for the | ||||||
| 1024 | /// moment. | ||||||
| 1025 | template <class AliasAnalysisType> | ||||||
| 1026 | class MemorySSA::CachingWalker final : public MemorySSAWalker { | ||||||
| 1027 | ClobberWalkerBase<AliasAnalysisType> *Walker; | ||||||
| 1028 | |||||||
| 1029 | public: | ||||||
| 1030 | CachingWalker(MemorySSA *M, ClobberWalkerBase<AliasAnalysisType> *W) | ||||||
| 1031 | : MemorySSAWalker(M), Walker(W) {} | ||||||
| 1032 | ~CachingWalker() override = default; | ||||||
| 1033 | |||||||
| 1034 | using MemorySSAWalker::getClobberingMemoryAccess; | ||||||
| 1035 | |||||||
| 1036 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, unsigned &UWL) { | ||||||
| 1037 | return Walker->getClobberingMemoryAccessBase(MA, UWL, false); | ||||||
| 1038 | } | ||||||
| 1039 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, | ||||||
| 1040 | const MemoryLocation &Loc, | ||||||
| 1041 | unsigned &UWL) { | ||||||
| 1042 | return Walker->getClobberingMemoryAccessBase(MA, Loc, UWL); | ||||||
| 1043 | } | ||||||
| 1044 | |||||||
| 1045 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA) override { | ||||||
| 1046 | unsigned UpwardWalkLimit = MaxCheckLimit; | ||||||
| 1047 | return getClobberingMemoryAccess(MA, UpwardWalkLimit); | ||||||
| 1048 | } | ||||||
| 1049 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, | ||||||
| 1050 | const MemoryLocation &Loc) override { | ||||||
| 1051 | unsigned UpwardWalkLimit = MaxCheckLimit; | ||||||
| 1052 | return getClobberingMemoryAccess(MA, Loc, UpwardWalkLimit); | ||||||
| 1053 | } | ||||||
| 1054 | |||||||
| 1055 | void invalidateInfo(MemoryAccess *MA) override { | ||||||
| 1056 | if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) | ||||||
| 1057 | MUD->resetOptimized(); | ||||||
| 1058 | } | ||||||
| 1059 | }; | ||||||
| 1060 | |||||||
| 1061 | template <class AliasAnalysisType> | ||||||
| 1062 | class MemorySSA::SkipSelfWalker final : public MemorySSAWalker { | ||||||
| 1063 | ClobberWalkerBase<AliasAnalysisType> *Walker; | ||||||
| 1064 | |||||||
| 1065 | public: | ||||||
| 1066 | SkipSelfWalker(MemorySSA *M, ClobberWalkerBase<AliasAnalysisType> *W) | ||||||
| 1067 | : MemorySSAWalker(M), Walker(W) {} | ||||||
| 1068 | ~SkipSelfWalker() override = default; | ||||||
| 1069 | |||||||
| 1070 | using MemorySSAWalker::getClobberingMemoryAccess; | ||||||
| 1071 | |||||||
| 1072 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, unsigned &UWL) { | ||||||
| 1073 | return Walker->getClobberingMemoryAccessBase(MA, UWL, true); | ||||||
| 1074 | } | ||||||
| 1075 | MemoryAccess *getClobberingMemoryAccess(MemoryAccess *MA, | ||||||
| 1076 | const MemoryLocation &Loc, | ||||||
| 1077 | unsigned &UWL) { | ||||||
| 1078 | return Walker->getClobberingMemoryAccessBase(MA, Loc, UWL); | ||||||
| 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 | } // end namespace llvm | ||||||
| 1098 | |||||||
| 1099 | void MemorySSA::renameSuccessorPhis(BasicBlock *BB, MemoryAccess *IncomingVal, | ||||||
| 1100 | bool RenameAllUses) { | ||||||
| 1101 | // Pass through values to our successors | ||||||
| 1102 | for (const BasicBlock *S : successors(BB)) { | ||||||
| 1103 | auto It = PerBlockAccesses.find(S); | ||||||
| 1104 | // Rename the phi nodes in our successor block | ||||||
| 1105 | if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front())) | ||||||
| 1106 | continue; | ||||||
| 1107 | AccessList *Accesses = It->second.get(); | ||||||
| 1108 | auto *Phi = cast<MemoryPhi>(&Accesses->front()); | ||||||
| 1109 | if (RenameAllUses) { | ||||||
| 1110 | bool ReplacementDone = false; | ||||||
| 1111 | for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) | ||||||
| 1112 | if (Phi->getIncomingBlock(I) == BB) { | ||||||
| 1113 | Phi->setIncomingValue(I, IncomingVal); | ||||||
| 1114 | ReplacementDone = true; | ||||||
| 1115 | } | ||||||
| 1116 | (void) ReplacementDone; | ||||||
| 1117 | assert(ReplacementDone && "Incomplete phi during partial rename")((void)0); | ||||||
| 1118 | } else | ||||||
| 1119 | Phi->addIncoming(IncomingVal, BB); | ||||||
| 1120 | } | ||||||
| 1121 | } | ||||||
| 1122 | |||||||
| 1123 | /// Rename a single basic block into MemorySSA form. | ||||||
| 1124 | /// Uses the standard SSA renaming algorithm. | ||||||
| 1125 | /// \returns The new incoming value. | ||||||
| 1126 | MemoryAccess *MemorySSA::renameBlock(BasicBlock *BB, MemoryAccess *IncomingVal, | ||||||
| 1127 | bool RenameAllUses) { | ||||||
| 1128 | auto It = PerBlockAccesses.find(BB); | ||||||
| 1129 | // Skip most processing if the list is empty. | ||||||
| 1130 | if (It != PerBlockAccesses.end()) { | ||||||
| 1131 | AccessList *Accesses = It->second.get(); | ||||||
| 1132 | for (MemoryAccess &L : *Accesses) { | ||||||
| 1133 | if (MemoryUseOrDef *MUD = dyn_cast<MemoryUseOrDef>(&L)) { | ||||||
| 1134 | if (MUD->getDefiningAccess() == nullptr || RenameAllUses) | ||||||
| 1135 | MUD->setDefiningAccess(IncomingVal); | ||||||
| 1136 | if (isa<MemoryDef>(&L)) | ||||||
| 1137 | IncomingVal = &L; | ||||||
| 1138 | } else { | ||||||
| 1139 | IncomingVal = &L; | ||||||
| 1140 | } | ||||||
| 1141 | } | ||||||
| 1142 | } | ||||||
| 1143 | return IncomingVal; | ||||||
| 1144 | } | ||||||
| 1145 | |||||||
| 1146 | /// This is the standard SSA renaming algorithm. | ||||||
| 1147 | /// | ||||||
| 1148 | /// We walk the dominator tree in preorder, renaming accesses, and then filling | ||||||
| 1149 | /// in phi nodes in our successors. | ||||||
| 1150 | void MemorySSA::renamePass(DomTreeNode *Root, MemoryAccess *IncomingVal, | ||||||
| 1151 | SmallPtrSetImpl<BasicBlock *> &Visited, | ||||||
| 1152 | bool SkipVisited, bool RenameAllUses) { | ||||||
| 1153 | assert(Root && "Trying to rename accesses in an unreachable block")((void)0); | ||||||
| 1154 | |||||||
| 1155 | SmallVector<RenamePassData, 32> WorkStack; | ||||||
| 1156 | // Skip everything if we already renamed this block and we are skipping. | ||||||
| 1157 | // Note: You can't sink this into the if, because we need it to occur | ||||||
| 1158 | // regardless of whether we skip blocks or not. | ||||||
| 1159 | bool AlreadyVisited = !Visited.insert(Root->getBlock()).second; | ||||||
| 1160 | if (SkipVisited && AlreadyVisited) | ||||||
| 1161 | return; | ||||||
| 1162 | |||||||
| 1163 | IncomingVal = renameBlock(Root->getBlock(), IncomingVal, RenameAllUses); | ||||||
| 1164 | renameSuccessorPhis(Root->getBlock(), IncomingVal, RenameAllUses); | ||||||
| 1165 | WorkStack.push_back({Root, Root->begin(), IncomingVal}); | ||||||
| 1166 | |||||||
| 1167 | while (!WorkStack.empty()) { | ||||||
| 1168 | DomTreeNode *Node = WorkStack.back().DTN; | ||||||
| 1169 | DomTreeNode::const_iterator ChildIt = WorkStack.back().ChildIt; | ||||||
| 1170 | IncomingVal = WorkStack.back().IncomingVal; | ||||||
| 1171 | |||||||
| 1172 | if (ChildIt == Node->end()) { | ||||||
| 1173 | WorkStack.pop_back(); | ||||||
| 1174 | } else { | ||||||
| 1175 | DomTreeNode *Child = *ChildIt; | ||||||
| 1176 | ++WorkStack.back().ChildIt; | ||||||
| 1177 | BasicBlock *BB = Child->getBlock(); | ||||||
| 1178 | // Note: You can't sink this into the if, because we need it to occur | ||||||
| 1179 | // regardless of whether we skip blocks or not. | ||||||
| 1180 | AlreadyVisited = !Visited.insert(BB).second; | ||||||
| 1181 | if (SkipVisited && AlreadyVisited) { | ||||||
| 1182 | // We already visited this during our renaming, which can happen when | ||||||
| 1183 | // being asked to rename multiple blocks. Figure out the incoming val, | ||||||
| 1184 | // which is the last def. | ||||||
| 1185 | // Incoming value can only change if there is a block def, and in that | ||||||
| 1186 | // case, it's the last block def in the list. | ||||||
| 1187 | if (auto *BlockDefs = getWritableBlockDefs(BB)) | ||||||
| 1188 | IncomingVal = &*BlockDefs->rbegin(); | ||||||
| 1189 | } else | ||||||
| 1190 | IncomingVal = renameBlock(BB, IncomingVal, RenameAllUses); | ||||||
| 1191 | renameSuccessorPhis(BB, IncomingVal, RenameAllUses); | ||||||
| 1192 | WorkStack.push_back({Child, Child->begin(), IncomingVal}); | ||||||
| 1193 | } | ||||||
| 1194 | } | ||||||
| 1195 | } | ||||||
| 1196 | |||||||
| 1197 | /// This handles unreachable block accesses by deleting phi nodes in | ||||||
| 1198 | /// unreachable blocks, and marking all other unreachable MemoryAccess's as | ||||||
| 1199 | /// being uses of the live on entry definition. | ||||||
| 1200 | void MemorySSA::markUnreachableAsLiveOnEntry(BasicBlock *BB) { | ||||||
| 1201 | assert(!DT->isReachableFromEntry(BB) &&((void)0) | ||||||
| 1202 | "Reachable block found while handling unreachable blocks")((void)0); | ||||||
| 1203 | |||||||
| 1204 | // Make sure phi nodes in our reachable successors end up with a | ||||||
| 1205 | // LiveOnEntryDef for our incoming edge, even though our block is forward | ||||||
| 1206 | // unreachable. We could just disconnect these blocks from the CFG fully, | ||||||
| 1207 | // but we do not right now. | ||||||
| 1208 | for (const BasicBlock *S : successors(BB)) { | ||||||
| 1209 | if (!DT->isReachableFromEntry(S)) | ||||||
| 1210 | continue; | ||||||
| 1211 | auto It = PerBlockAccesses.find(S); | ||||||
| 1212 | // Rename the phi nodes in our successor block | ||||||
| 1213 | if (It == PerBlockAccesses.end() || !isa<MemoryPhi>(It->second->front())) | ||||||
| 1214 | continue; | ||||||
| 1215 | AccessList *Accesses = It->second.get(); | ||||||
| 1216 | auto *Phi = cast<MemoryPhi>(&Accesses->front()); | ||||||
| 1217 | Phi->addIncoming(LiveOnEntryDef.get(), BB); | ||||||
| 1218 | } | ||||||
| 1219 | |||||||
| 1220 | auto It = PerBlockAccesses.find(BB); | ||||||
| 1221 | if (It == PerBlockAccesses.end()) | ||||||
| 1222 | return; | ||||||
| 1223 | |||||||
| 1224 | auto &Accesses = It->second; | ||||||
| 1225 | for (auto AI = Accesses->begin(), AE = Accesses->end(); AI != AE;) { | ||||||
| 1226 | auto Next = std::next(AI); | ||||||
| 1227 | // If we have a phi, just remove it. We are going to replace all | ||||||
| 1228 | // users with live on entry. | ||||||
| 1229 | if (auto *UseOrDef = dyn_cast<MemoryUseOrDef>(AI)) | ||||||
| 1230 | UseOrDef->setDefiningAccess(LiveOnEntryDef.get()); | ||||||
| 1231 | else | ||||||
| 1232 | Accesses->erase(AI); | ||||||
| 1233 | AI = Next; | ||||||
| 1234 | } | ||||||
| 1235 | } | ||||||
| 1236 | |||||||
| 1237 | MemorySSA::MemorySSA(Function &Func, AliasAnalysis *AA, DominatorTree *DT) | ||||||
| 1238 | : AA(nullptr), DT(DT), F(Func), LiveOnEntryDef(nullptr), Walker(nullptr), | ||||||
| 1239 | SkipWalker(nullptr), NextID(0) { | ||||||
| 1240 | // Build MemorySSA using a batch alias analysis. This reuses the internal | ||||||
| 1241 | // state that AA collects during an alias()/getModRefInfo() call. This is | ||||||
| 1242 | // safe because there are no CFG changes while building MemorySSA and can | ||||||
| 1243 | // significantly reduce the time spent by the compiler in AA, because we will | ||||||
| 1244 | // make queries about all the instructions in the Function. | ||||||
| 1245 | assert(AA && "No alias analysis?")((void)0); | ||||||
| 1246 | BatchAAResults BatchAA(*AA); | ||||||
| 1247 | buildMemorySSA(BatchAA); | ||||||
| 1248 | // Intentionally leave AA to nullptr while building so we don't accidently | ||||||
| 1249 | // use non-batch AliasAnalysis. | ||||||
| 1250 | this->AA = AA; | ||||||
| 1251 | // Also create the walker here. | ||||||
| 1252 | getWalker(); | ||||||
| 1253 | } | ||||||
| 1254 | |||||||
| 1255 | MemorySSA::~MemorySSA() { | ||||||
| 1256 | // Drop all our references | ||||||
| 1257 | for (const auto &Pair : PerBlockAccesses) | ||||||
| 1258 | for (MemoryAccess &MA : *Pair.second) | ||||||
| 1259 | MA.dropAllReferences(); | ||||||
| 1260 | } | ||||||
| 1261 | |||||||
| 1262 | MemorySSA::AccessList *MemorySSA::getOrCreateAccessList(const BasicBlock *BB) { | ||||||
| 1263 | auto Res = PerBlockAccesses.insert(std::make_pair(BB, nullptr)); | ||||||
| 1264 | |||||||
| 1265 | if (Res.second) | ||||||
| 1266 | Res.first->second = std::make_unique<AccessList>(); | ||||||
| 1267 | return Res.first->second.get(); | ||||||
| 1268 | } | ||||||
| 1269 | |||||||
| 1270 | MemorySSA::DefsList *MemorySSA::getOrCreateDefsList(const BasicBlock *BB) { | ||||||
| 1271 | auto Res = PerBlockDefs.insert(std::make_pair(BB, nullptr)); | ||||||
| 1272 | |||||||
| 1273 | if (Res.second) | ||||||
| 1274 | Res.first->second = std::make_unique<DefsList>(); | ||||||
| 1275 | return Res.first->second.get(); | ||||||
| 1276 | } | ||||||
| 1277 | |||||||
| 1278 | namespace llvm { | ||||||
| 1279 | |||||||
| 1280 | /// This class is a batch walker of all MemoryUse's in the program, and points | ||||||
| 1281 | /// their defining access at the thing that actually clobbers them. Because it | ||||||
| 1282 | /// is a batch walker that touches everything, it does not operate like the | ||||||
| 1283 | /// other walkers. This walker is basically performing a top-down SSA renaming | ||||||
| 1284 | /// pass, where the version stack is used as the cache. This enables it to be | ||||||
| 1285 | /// significantly more time and memory efficient than using the regular walker, | ||||||
| 1286 | /// which is walking bottom-up. | ||||||
| 1287 | class MemorySSA::OptimizeUses { | ||||||
| 1288 | public: | ||||||
| 1289 | OptimizeUses(MemorySSA *MSSA, CachingWalker<BatchAAResults> *Walker, | ||||||
| 1290 | BatchAAResults *BAA, DominatorTree *DT) | ||||||
| 1291 | : MSSA(MSSA), Walker(Walker), AA(BAA), DT(DT) {} | ||||||
| 1292 | |||||||
| 1293 | void optimizeUses(); | ||||||
| 1294 | |||||||
| 1295 | private: | ||||||
| 1296 | /// This represents where a given memorylocation is in the stack. | ||||||
| 1297 | struct MemlocStackInfo { | ||||||
| 1298 | // This essentially is keeping track of versions of the stack. Whenever | ||||||
| 1299 | // the stack changes due to pushes or pops, these versions increase. | ||||||
| 1300 | unsigned long StackEpoch; | ||||||
| 1301 | unsigned long PopEpoch; | ||||||
| 1302 | // This is the lower bound of places on the stack to check. It is equal to | ||||||
| 1303 | // the place the last stack walk ended. | ||||||
| 1304 | // Note: Correctness depends on this being initialized to 0, which densemap | ||||||
| 1305 | // does | ||||||
| 1306 | unsigned long LowerBound; | ||||||
| 1307 | const BasicBlock *LowerBoundBlock; | ||||||
| 1308 | // This is where the last walk for this memory location ended. | ||||||
| 1309 | unsigned long LastKill; | ||||||
| 1310 | bool LastKillValid; | ||||||
| 1311 | Optional<AliasResult> AR; | ||||||
| 1312 | }; | ||||||
| 1313 | |||||||
| 1314 | void optimizeUsesInBlock(const BasicBlock *, unsigned long &, unsigned long &, | ||||||
| 1315 | SmallVectorImpl<MemoryAccess *> &, | ||||||
| 1316 | DenseMap<MemoryLocOrCall, MemlocStackInfo> &); | ||||||
| 1317 | |||||||
| 1318 | MemorySSA *MSSA; | ||||||
| 1319 | CachingWalker<BatchAAResults> *Walker; | ||||||
| 1320 | BatchAAResults *AA; | ||||||
| 1321 | DominatorTree *DT; | ||||||
| 1322 | }; | ||||||
| 1323 | |||||||
| 1324 | } // end namespace llvm | ||||||
| 1325 | |||||||
| 1326 | /// Optimize the uses in a given block This is basically the SSA renaming | ||||||
| 1327 | /// algorithm, with one caveat: We are able to use a single stack for all | ||||||
| 1328 | /// MemoryUses. This is because the set of *possible* reaching MemoryDefs is | ||||||
| 1329 | /// the same for every MemoryUse. The *actual* clobbering MemoryDef is just | ||||||
| 1330 | /// going to be some position in that stack of possible ones. | ||||||
| 1331 | /// | ||||||
| 1332 | /// We track the stack positions that each MemoryLocation needs | ||||||
| 1333 | /// to check, and last ended at. This is because we only want to check the | ||||||
| 1334 | /// things that changed since last time. The same MemoryLocation should | ||||||
| 1335 | /// get clobbered by the same store (getModRefInfo does not use invariantness or | ||||||
| 1336 | /// things like this, and if they start, we can modify MemoryLocOrCall to | ||||||
| 1337 | /// include relevant data) | ||||||
| 1338 | void MemorySSA::OptimizeUses::optimizeUsesInBlock( | ||||||
| 1339 | const BasicBlock *BB, unsigned long &StackEpoch, unsigned long &PopEpoch, | ||||||
| 1340 | SmallVectorImpl<MemoryAccess *> &VersionStack, | ||||||
| 1341 | DenseMap<MemoryLocOrCall, MemlocStackInfo> &LocStackInfo) { | ||||||
| 1342 | |||||||
| 1343 | /// If no accesses, nothing to do. | ||||||
| 1344 | MemorySSA::AccessList *Accesses = MSSA->getWritableBlockAccesses(BB); | ||||||
| 1345 | if (Accesses == nullptr) | ||||||
| 1346 | return; | ||||||
| 1347 | |||||||
| 1348 | // Pop everything that doesn't dominate the current block off the stack, | ||||||
| 1349 | // increment the PopEpoch to account for this. | ||||||
| 1350 | while (true) { | ||||||
| 1351 | assert(((void)0) | ||||||
| 1352 | !VersionStack.empty() &&((void)0) | ||||||
| 1353 | "Version stack should have liveOnEntry sentinel dominating everything")((void)0); | ||||||
| 1354 | BasicBlock *BackBlock = VersionStack.back()->getBlock(); | ||||||
| 1355 | if (DT->dominates(BackBlock, BB)) | ||||||
| 1356 | break; | ||||||
| 1357 | while (VersionStack.back()->getBlock() == BackBlock) | ||||||
| 1358 | VersionStack.pop_back(); | ||||||
| 1359 | ++PopEpoch; | ||||||
| 1360 | } | ||||||
| 1361 | |||||||
| 1362 | for (MemoryAccess &MA : *Accesses) { | ||||||
| 1363 | auto *MU = dyn_cast<MemoryUse>(&MA); | ||||||
| 1364 | if (!MU) { | ||||||
| 1365 | VersionStack.push_back(&MA); | ||||||
| 1366 | ++StackEpoch; | ||||||
| 1367 | continue; | ||||||
| 1368 | } | ||||||
| 1369 | |||||||
| 1370 | if (isUseTriviallyOptimizableToLiveOnEntry(*AA, MU->getMemoryInst())) { | ||||||
| 1371 | MU->setDefiningAccess(MSSA->getLiveOnEntryDef(), true, None); | ||||||
| 1372 | continue; | ||||||
| 1373 | } | ||||||
| 1374 | |||||||
| 1375 | MemoryLocOrCall UseMLOC(MU); | ||||||
| 1376 | auto &LocInfo = LocStackInfo[UseMLOC]; | ||||||
| 1377 | // If the pop epoch changed, it means we've removed stuff from top of | ||||||
| 1378 | // stack due to changing blocks. We may have to reset the lower bound or | ||||||
| 1379 | // last kill info. | ||||||
| 1380 | if (LocInfo.PopEpoch != PopEpoch) { | ||||||
| 1381 | LocInfo.PopEpoch = PopEpoch; | ||||||
| 1382 | LocInfo.StackEpoch = StackEpoch; | ||||||
| 1383 | // If the lower bound was in something that no longer dominates us, we | ||||||
| 1384 | // have to reset it. | ||||||
| 1385 | // We can't simply track stack size, because the stack may have had | ||||||
| 1386 | // pushes/pops in the meantime. | ||||||
| 1387 | // XXX: This is non-optimal, but only is slower cases with heavily | ||||||
| 1388 | // branching dominator trees. To get the optimal number of queries would | ||||||
| 1389 | // be to make lowerbound and lastkill a per-loc stack, and pop it until | ||||||
| 1390 | // the top of that stack dominates us. This does not seem worth it ATM. | ||||||
| 1391 | // A much cheaper optimization would be to always explore the deepest | ||||||
| 1392 | // branch of the dominator tree first. This will guarantee this resets on | ||||||
| 1393 | // the smallest set of blocks. | ||||||
| 1394 | if (LocInfo.LowerBoundBlock && LocInfo.LowerBoundBlock != BB && | ||||||
| 1395 | !DT->dominates(LocInfo.LowerBoundBlock, BB)) { | ||||||
| 1396 | // Reset the lower bound of things to check. | ||||||
| 1397 | // TODO: Some day we should be able to reset to last kill, rather than | ||||||
| 1398 | // 0. | ||||||
| 1399 | LocInfo.LowerBound = 0; | ||||||
| 1400 | LocInfo.LowerBoundBlock = VersionStack[0]->getBlock(); | ||||||
| 1401 | LocInfo.LastKillValid = false; | ||||||
| 1402 | } | ||||||
| 1403 | } else if (LocInfo.StackEpoch != StackEpoch) { | ||||||
| 1404 | // If all that has changed is the StackEpoch, we only have to check the | ||||||
| 1405 | // new things on the stack, because we've checked everything before. In | ||||||
| 1406 | // this case, the lower bound of things to check remains the same. | ||||||
| 1407 | LocInfo.PopEpoch = PopEpoch; | ||||||
| 1408 | LocInfo.StackEpoch = StackEpoch; | ||||||
| 1409 | } | ||||||
| 1410 | if (!LocInfo.LastKillValid) { | ||||||
| 1411 | LocInfo.LastKill = VersionStack.size() - 1; | ||||||
| 1412 | LocInfo.LastKillValid = true; | ||||||
| 1413 | LocInfo.AR = AliasResult::MayAlias; | ||||||
| 1414 | } | ||||||
| 1415 | |||||||
| 1416 | // At this point, we should have corrected last kill and LowerBound to be | ||||||
| 1417 | // in bounds. | ||||||
| 1418 | assert(LocInfo.LowerBound < VersionStack.size() &&((void)0) | ||||||
| 1419 | "Lower bound out of range")((void)0); | ||||||
| 1420 | assert(LocInfo.LastKill < VersionStack.size() &&((void)0) | ||||||
| 1421 | "Last kill info out of range")((void)0); | ||||||
| 1422 | // In any case, the new upper bound is the top of the stack. | ||||||
| 1423 | unsigned long UpperBound = VersionStack.size() - 1; | ||||||
| 1424 | |||||||
| 1425 | if (UpperBound - LocInfo.LowerBound > MaxCheckLimit) { | ||||||
| 1426 | LLVM_DEBUG(dbgs() << "MemorySSA skipping optimization of " << *MU << " ("do { } while (false) | ||||||
| 1427 | << *(MU->getMemoryInst()) << ")"do { } while (false) | ||||||
| 1428 | << " because there are "do { } while (false) | ||||||
| 1429 | << UpperBound - LocInfo.LowerBounddo { } while (false) | ||||||
| 1430 | << " stores to disambiguate\n")do { } while (false); | ||||||
| 1431 | // Because we did not walk, LastKill is no longer valid, as this may | ||||||
| 1432 | // have been a kill. | ||||||
| 1433 | LocInfo.LastKillValid = false; | ||||||
| 1434 | continue; | ||||||
| 1435 | } | ||||||
| 1436 | bool FoundClobberResult = false; | ||||||
| 1437 | unsigned UpwardWalkLimit = MaxCheckLimit; | ||||||
| 1438 | while (UpperBound > LocInfo.LowerBound) { | ||||||
| 1439 | if (isa<MemoryPhi>(VersionStack[UpperBound])) { | ||||||
| 1440 | // For phis, use the walker, see where we ended up, go there | ||||||
| 1441 | MemoryAccess *Result = | ||||||
| 1442 | Walker->getClobberingMemoryAccess(MU, UpwardWalkLimit); | ||||||
| 1443 | // We are guaranteed to find it or something is wrong | ||||||
| 1444 | while (VersionStack[UpperBound] != Result) { | ||||||
| 1445 | assert(UpperBound != 0)((void)0); | ||||||
| 1446 | --UpperBound; | ||||||
| 1447 | } | ||||||
| 1448 | FoundClobberResult = true; | ||||||
| 1449 | break; | ||||||
| 1450 | } | ||||||
| 1451 | |||||||
| 1452 | MemoryDef *MD = cast<MemoryDef>(VersionStack[UpperBound]); | ||||||
| 1453 | ClobberAlias CA = instructionClobbersQuery(MD, MU, UseMLOC, *AA); | ||||||
| 1454 | if (CA.IsClobber) { | ||||||
| 1455 | FoundClobberResult = true; | ||||||
| 1456 | LocInfo.AR = CA.AR; | ||||||
| 1457 | break; | ||||||
| 1458 | } | ||||||
| 1459 | --UpperBound; | ||||||
| 1460 | } | ||||||
| 1461 | |||||||
| 1462 | // Note: Phis always have AliasResult AR set to MayAlias ATM. | ||||||
| 1463 | |||||||
| 1464 | // At the end of this loop, UpperBound is either a clobber, or lower bound | ||||||
| 1465 | // PHI walking may cause it to be < LowerBound, and in fact, < LastKill. | ||||||
| 1466 | if (FoundClobberResult || UpperBound < LocInfo.LastKill) { | ||||||
| 1467 | // We were last killed now by where we got to | ||||||
| 1468 | if (MSSA->isLiveOnEntryDef(VersionStack[UpperBound])) | ||||||
| 1469 | LocInfo.AR = None; | ||||||
| 1470 | MU->setDefiningAccess(VersionStack[UpperBound], true, LocInfo.AR); | ||||||
| 1471 | LocInfo.LastKill = UpperBound; | ||||||
| 1472 | } else { | ||||||
| 1473 | // Otherwise, we checked all the new ones, and now we know we can get to | ||||||
| 1474 | // LastKill. | ||||||
| 1475 | MU->setDefiningAccess(VersionStack[LocInfo.LastKill], true, LocInfo.AR); | ||||||
| 1476 | } | ||||||
| 1477 | LocInfo.LowerBound = VersionStack.size() - 1; | ||||||
| 1478 | LocInfo.LowerBoundBlock = BB; | ||||||
| 1479 | } | ||||||
| 1480 | } | ||||||
| 1481 | |||||||
| 1482 | /// Optimize uses to point to their actual clobbering definitions. | ||||||
| 1483 | void MemorySSA::OptimizeUses::optimizeUses() { | ||||||
| 1484 | SmallVector<MemoryAccess *, 16> VersionStack; | ||||||
| 1485 | DenseMap<MemoryLocOrCall, MemlocStackInfo> LocStackInfo; | ||||||
| 1486 | VersionStack.push_back(MSSA->getLiveOnEntryDef()); | ||||||
| 1487 | |||||||
| 1488 | unsigned long StackEpoch = 1; | ||||||
| 1489 | unsigned long PopEpoch = 1; | ||||||
| 1490 | // We perform a non-recursive top-down dominator tree walk. | ||||||
| 1491 | for (const auto *DomNode : depth_first(DT->getRootNode())) | ||||||
| 1492 | optimizeUsesInBlock(DomNode->getBlock(), StackEpoch, PopEpoch, VersionStack, | ||||||
| 1493 | LocStackInfo); | ||||||
| 1494 | } | ||||||
| 1495 | |||||||
| 1496 | void MemorySSA::placePHINodes( | ||||||
| 1497 | const SmallPtrSetImpl<BasicBlock *> &DefiningBlocks) { | ||||||
| 1498 | // Determine where our MemoryPhi's should go | ||||||
| 1499 | ForwardIDFCalculator IDFs(*DT); | ||||||
| 1500 | IDFs.setDefiningBlocks(DefiningBlocks); | ||||||
| 1501 | SmallVector<BasicBlock *, 32> IDFBlocks; | ||||||
| 1502 | IDFs.calculate(IDFBlocks); | ||||||
| 1503 | |||||||
| 1504 | // Now place MemoryPhi nodes. | ||||||
| 1505 | for (auto &BB : IDFBlocks) | ||||||
| 1506 | createMemoryPhi(BB); | ||||||
| 1507 | } | ||||||
| 1508 | |||||||
| 1509 | void MemorySSA::buildMemorySSA(BatchAAResults &BAA) { | ||||||
| 1510 | // We create an access to represent "live on entry", for things like | ||||||
| 1511 | // arguments or users of globals, where the memory they use is defined before | ||||||
| 1512 | // the beginning of the function. We do not actually insert it into the IR. | ||||||
| 1513 | // We do not define a live on exit for the immediate uses, and thus our | ||||||
| 1514 | // semantics do *not* imply that something with no immediate uses can simply | ||||||
| 1515 | // be removed. | ||||||
| 1516 | BasicBlock &StartingPoint = F.getEntryBlock(); | ||||||
| 1517 | LiveOnEntryDef.reset(new MemoryDef(F.getContext(), nullptr, nullptr, | ||||||
| 1518 | &StartingPoint, NextID++)); | ||||||
| 1519 | |||||||
| 1520 | // We maintain lists of memory accesses per-block, trading memory for time. We | ||||||
| 1521 | // could just look up the memory access for every possible instruction in the | ||||||
| 1522 | // stream. | ||||||
| 1523 | SmallPtrSet<BasicBlock *, 32> DefiningBlocks; | ||||||
| 1524 | // Go through each block, figure out where defs occur, and chain together all | ||||||
| 1525 | // the accesses. | ||||||
| 1526 | for (BasicBlock &B : F) { | ||||||
| 1527 | bool InsertIntoDef = false; | ||||||
| 1528 | AccessList *Accesses = nullptr; | ||||||
| 1529 | DefsList *Defs = nullptr; | ||||||
| 1530 | for (Instruction &I : B) { | ||||||
| 1531 | MemoryUseOrDef *MUD = createNewAccess(&I, &BAA); | ||||||
| 1532 | if (!MUD) | ||||||
| 1533 | continue; | ||||||
| 1534 | |||||||
| 1535 | if (!Accesses) | ||||||
| 1536 | Accesses = getOrCreateAccessList(&B); | ||||||
| 1537 | Accesses->push_back(MUD); | ||||||
| 1538 | if (isa<MemoryDef>(MUD)) { | ||||||
| 1539 | InsertIntoDef = true; | ||||||
| 1540 | if (!Defs) | ||||||
| 1541 | Defs = getOrCreateDefsList(&B); | ||||||
| 1542 | Defs->push_back(*MUD); | ||||||
| 1543 | } | ||||||
| 1544 | } | ||||||
| 1545 | if (InsertIntoDef) | ||||||
| 1546 | DefiningBlocks.insert(&B); | ||||||
| 1547 | } | ||||||
| 1548 | placePHINodes(DefiningBlocks); | ||||||
| 1549 | |||||||
| 1550 | // Now do regular SSA renaming on the MemoryDef/MemoryUse. Visited will get | ||||||
| 1551 | // filled in with all blocks. | ||||||
| 1552 | SmallPtrSet<BasicBlock *, 16> Visited; | ||||||
| 1553 | renamePass(DT->getRootNode(), LiveOnEntryDef.get(), Visited); | ||||||
| 1554 | |||||||
| 1555 | ClobberWalkerBase<BatchAAResults> WalkerBase(this, &BAA, DT); | ||||||
| 1556 | CachingWalker<BatchAAResults> WalkerLocal(this, &WalkerBase); | ||||||
| 1557 | OptimizeUses(this, &WalkerLocal, &BAA, DT).optimizeUses(); | ||||||
| 1558 | |||||||
| 1559 | // Mark the uses in unreachable blocks as live on entry, so that they go | ||||||
| 1560 | // somewhere. | ||||||
| 1561 | for (auto &BB : F) | ||||||
| 1562 | if (!Visited.count(&BB)) | ||||||
| 1563 | markUnreachableAsLiveOnEntry(&BB); | ||||||
| 1564 | } | ||||||
| 1565 | |||||||
| 1566 | MemorySSAWalker *MemorySSA::getWalker() { return getWalkerImpl(); } | ||||||
| 1567 | |||||||
| 1568 | MemorySSA::CachingWalker<AliasAnalysis> *MemorySSA::getWalkerImpl() { | ||||||
| 1569 | if (Walker) | ||||||
| 1570 | return Walker.get(); | ||||||
| 1571 | |||||||
| 1572 | if (!WalkerBase) | ||||||
| 1573 | WalkerBase = | ||||||
| 1574 | std::make_unique<ClobberWalkerBase<AliasAnalysis>>(this, AA, DT); | ||||||
| 1575 | |||||||
| 1576 | Walker = | ||||||
| 1577 | std::make_unique<CachingWalker<AliasAnalysis>>(this, WalkerBase.get()); | ||||||
| 1578 | return Walker.get(); | ||||||
| 1579 | } | ||||||
| 1580 | |||||||
| 1581 | MemorySSAWalker *MemorySSA::getSkipSelfWalker() { | ||||||
| 1582 | if (SkipWalker) | ||||||
| 1583 | return SkipWalker.get(); | ||||||
| 1584 | |||||||
| 1585 | if (!WalkerBase) | ||||||
| 1586 | WalkerBase = | ||||||
| 1587 | std::make_unique<ClobberWalkerBase<AliasAnalysis>>(this, AA, DT); | ||||||
| 1588 | |||||||
| 1589 | SkipWalker = | ||||||
| 1590 | std::make_unique<SkipSelfWalker<AliasAnalysis>>(this, WalkerBase.get()); | ||||||
| 1591 | return SkipWalker.get(); | ||||||
| 1592 | } | ||||||
| 1593 | |||||||
| 1594 | |||||||
| 1595 | // This is a helper function used by the creation routines. It places NewAccess | ||||||
| 1596 | // into the access and defs lists for a given basic block, at the given | ||||||
| 1597 | // insertion point. | ||||||
| 1598 | void MemorySSA::insertIntoListsForBlock(MemoryAccess *NewAccess, | ||||||
| 1599 | const BasicBlock *BB, | ||||||
| 1600 | InsertionPlace Point) { | ||||||
| 1601 | auto *Accesses = getOrCreateAccessList(BB); | ||||||
| 1602 | if (Point == Beginning) { | ||||||
| 1603 | // If it's a phi node, it goes first, otherwise, it goes after any phi | ||||||
| 1604 | // nodes. | ||||||
| 1605 | if (isa<MemoryPhi>(NewAccess)) { | ||||||
| 1606 | Accesses->push_front(NewAccess); | ||||||
| 1607 | auto *Defs = getOrCreateDefsList(BB); | ||||||
| 1608 | Defs->push_front(*NewAccess); | ||||||
| 1609 | } else { | ||||||
| 1610 | auto AI = find_if_not( | ||||||
| 1611 | *Accesses, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); }); | ||||||
| 1612 | Accesses->insert(AI, NewAccess); | ||||||
| 1613 | if (!isa<MemoryUse>(NewAccess)) { | ||||||
| 1614 | auto *Defs = getOrCreateDefsList(BB); | ||||||
| 1615 | auto DI = find_if_not( | ||||||
| 1616 | *Defs, [](const MemoryAccess &MA) { return isa<MemoryPhi>(MA); }); | ||||||
| 1617 | Defs->insert(DI, *NewAccess); | ||||||
| 1618 | } | ||||||
| 1619 | } | ||||||
| 1620 | } else { | ||||||
| 1621 | Accesses->push_back(NewAccess); | ||||||
| 1622 | if (!isa<MemoryUse>(NewAccess)) { | ||||||
| 1623 | auto *Defs = getOrCreateDefsList(BB); | ||||||
| 1624 | Defs->push_back(*NewAccess); | ||||||
| 1625 | } | ||||||
| 1626 | } | ||||||
| 1627 | BlockNumberingValid.erase(BB); | ||||||
| 1628 | } | ||||||
| 1629 | |||||||
| 1630 | void MemorySSA::insertIntoListsBefore(MemoryAccess *What, const BasicBlock *BB, | ||||||
| 1631 | AccessList::iterator InsertPt) { | ||||||
| 1632 | auto *Accesses = getWritableBlockAccesses(BB); | ||||||
| 1633 | bool WasEnd = InsertPt == Accesses->end(); | ||||||
| 1634 | Accesses->insert(AccessList::iterator(InsertPt), What); | ||||||
| 1635 | if (!isa<MemoryUse>(What)) { | ||||||
| 1636 | auto *Defs = getOrCreateDefsList(BB); | ||||||
| 1637 | // If we got asked to insert at the end, we have an easy job, just shove it | ||||||
| 1638 | // at the end. If we got asked to insert before an existing def, we also get | ||||||
| 1639 | // an iterator. If we got asked to insert before a use, we have to hunt for | ||||||
| 1640 | // the next def. | ||||||
| 1641 | if (WasEnd) { | ||||||
| 1642 | Defs->push_back(*What); | ||||||
| 1643 | } else if (isa<MemoryDef>(InsertPt)) { | ||||||
| 1644 | Defs->insert(InsertPt->getDefsIterator(), *What); | ||||||
| 1645 | } else { | ||||||
| 1646 | while (InsertPt != Accesses->end() && !isa<MemoryDef>(InsertPt)) | ||||||
| 1647 | ++InsertPt; | ||||||
| 1648 | // Either we found a def, or we are inserting at the end | ||||||
| 1649 | if (InsertPt == Accesses->end()) | ||||||
| 1650 | Defs->push_back(*What); | ||||||
| 1651 | else | ||||||
| 1652 | Defs->insert(InsertPt->getDefsIterator(), *What); | ||||||
| 1653 | } | ||||||
| 1654 | } | ||||||
| 1655 | BlockNumberingValid.erase(BB); | ||||||
| 1656 | } | ||||||
| 1657 | |||||||
| 1658 | void MemorySSA::prepareForMoveTo(MemoryAccess *What, BasicBlock *BB) { | ||||||
| 1659 | // Keep it in the lookup tables, remove from the lists | ||||||
| 1660 | removeFromLists(What, false); | ||||||
| 1661 | |||||||
| 1662 | // Note that moving should implicitly invalidate the optimized state of a | ||||||
| 1663 | // MemoryUse (and Phis can't be optimized). However, it doesn't do so for a | ||||||
| 1664 | // MemoryDef. | ||||||
| 1665 | if (auto *MD = dyn_cast<MemoryDef>(What)) | ||||||
| 1666 | MD->resetOptimized(); | ||||||
| 1667 | What->setBlock(BB); | ||||||
| 1668 | } | ||||||
| 1669 | |||||||
| 1670 | // Move What before Where in the IR. The end result is that What will belong to | ||||||
| 1671 | // the right lists and have the right Block set, but will not otherwise be | ||||||
| 1672 | // correct. It will not have the right defining access, and if it is a def, | ||||||
| 1673 | // things below it will not properly be updated. | ||||||
| 1674 | void MemorySSA::moveTo(MemoryUseOrDef *What, BasicBlock *BB, | ||||||
| 1675 | AccessList::iterator Where) { | ||||||
| 1676 | prepareForMoveTo(What, BB); | ||||||
| 1677 | insertIntoListsBefore(What, BB, Where); | ||||||
| 1678 | } | ||||||
| 1679 | |||||||
| 1680 | void MemorySSA::moveTo(MemoryAccess *What, BasicBlock *BB, | ||||||
| 1681 | InsertionPlace Point) { | ||||||
| 1682 | if (isa<MemoryPhi>(What)) { | ||||||
| 1683 | assert(Point == Beginning &&((void)0) | ||||||
| 1684 | "Can only move a Phi at the beginning of the block")((void)0); | ||||||
| 1685 | // Update lookup table entry | ||||||
| 1686 | ValueToMemoryAccess.erase(What->getBlock()); | ||||||
| 1687 | bool Inserted = ValueToMemoryAccess.insert({BB, What}).second; | ||||||
| 1688 | (void)Inserted; | ||||||
| 1689 | assert(Inserted && "Cannot move a Phi to a block that already has one")((void)0); | ||||||
| 1690 | } | ||||||
| 1691 | |||||||
| 1692 | prepareForMoveTo(What, BB); | ||||||
| 1693 | insertIntoListsForBlock(What, BB, Point); | ||||||
| 1694 | } | ||||||
| 1695 | |||||||
| 1696 | MemoryPhi *MemorySSA::createMemoryPhi(BasicBlock *BB) { | ||||||
| 1697 | assert(!getMemoryAccess(BB) && "MemoryPhi already exists for this BB")((void)0); | ||||||
| 1698 | MemoryPhi *Phi = new MemoryPhi(BB->getContext(), BB, NextID++); | ||||||
| 1699 | // Phi's always are placed at the front of the block. | ||||||
| 1700 | insertIntoListsForBlock(Phi, BB, Beginning); | ||||||
| 1701 | ValueToMemoryAccess[BB] = Phi; | ||||||
| 1702 | return Phi; | ||||||
| 1703 | } | ||||||
| 1704 | |||||||
| 1705 | MemoryUseOrDef *MemorySSA::createDefinedAccess(Instruction *I, | ||||||
| 1706 | MemoryAccess *Definition, | ||||||
| 1707 | const MemoryUseOrDef *Template, | ||||||
| 1708 | bool CreationMustSucceed) { | ||||||
| 1709 | assert(!isa<PHINode>(I) && "Cannot create a defined access for a PHI")((void)0); | ||||||
| 1710 | MemoryUseOrDef *NewAccess = createNewAccess(I, AA, Template); | ||||||
| 1711 | if (CreationMustSucceed) | ||||||
| 1712 | assert(NewAccess != nullptr && "Tried to create a memory access for a "((void)0) | ||||||
| 1713 | "non-memory touching instruction")((void)0); | ||||||
| 1714 | if (NewAccess) { | ||||||
| 1715 | assert((!Definition || !isa<MemoryUse>(Definition)) &&((void)0) | ||||||
| 1716 | "A use cannot be a defining access")((void)0); | ||||||
| 1717 | NewAccess->setDefiningAccess(Definition); | ||||||
| 1718 | } | ||||||
| 1719 | return NewAccess; | ||||||
| 1720 | } | ||||||
| 1721 | |||||||
| 1722 | // Return true if the instruction has ordering constraints. | ||||||
| 1723 | // Note specifically that this only considers stores and loads | ||||||
| 1724 | // because others are still considered ModRef by getModRefInfo. | ||||||
| 1725 | static inline bool isOrdered(const Instruction *I) { | ||||||
| 1726 | if (auto *SI = dyn_cast<StoreInst>(I)) { | ||||||
| 1727 | if (!SI->isUnordered()) | ||||||
| 1728 | return true; | ||||||
| 1729 | } else if (auto *LI = dyn_cast<LoadInst>(I)) { | ||||||
| 1730 | if (!LI->isUnordered()) | ||||||
| 1731 | return true; | ||||||
| 1732 | } | ||||||
| 1733 | return false; | ||||||
| 1734 | } | ||||||
| 1735 | |||||||
| 1736 | /// Helper function to create new memory accesses | ||||||
| 1737 | template <typename AliasAnalysisType> | ||||||
| 1738 | MemoryUseOrDef *MemorySSA::createNewAccess(Instruction *I, | ||||||
| 1739 | AliasAnalysisType *AAP, | ||||||
| 1740 | const MemoryUseOrDef *Template) { | ||||||
| 1741 | // The assume intrinsic has a control dependency which we model by claiming | ||||||
| 1742 | // that it writes arbitrarily. Debuginfo intrinsics may be considered | ||||||
| 1743 | // clobbers when we have a nonstandard AA pipeline. Ignore these fake memory | ||||||
| 1744 | // dependencies here. | ||||||
| 1745 | // FIXME: Replace this special casing with a more accurate modelling of | ||||||
| 1746 | // assume's control dependency. | ||||||
| 1747 | if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) { | ||||||
| 1748 | switch (II->getIntrinsicID()) { | ||||||
| 1749 | default: | ||||||
| 1750 | break; | ||||||
| 1751 | case Intrinsic::assume: | ||||||
| 1752 | case Intrinsic::experimental_noalias_scope_decl: | ||||||
| 1753 | return nullptr; | ||||||
| 1754 | } | ||||||
| 1755 | } | ||||||
| 1756 | |||||||
| 1757 | // Using a nonstandard AA pipelines might leave us with unexpected modref | ||||||
| 1758 | // results for I, so add a check to not model instructions that may not read | ||||||
| 1759 | // from or write to memory. This is necessary for correctness. | ||||||
| 1760 | if (!I->mayReadFromMemory() && !I->mayWriteToMemory()) | ||||||
| 1761 | return nullptr; | ||||||
| 1762 | |||||||
| 1763 | bool Def, Use; | ||||||
| 1764 | if (Template) { | ||||||
| 1765 | Def = isa<MemoryDef>(Template); | ||||||
| 1766 | Use = isa<MemoryUse>(Template); | ||||||
| 1767 | #if !defined(NDEBUG1) | ||||||
| 1768 | ModRefInfo ModRef = AAP->getModRefInfo(I, None); | ||||||
| 1769 | bool DefCheck, UseCheck; | ||||||
| 1770 | DefCheck = isModSet(ModRef) || isOrdered(I); | ||||||
| 1771 | UseCheck = isRefSet(ModRef); | ||||||
| 1772 | assert(Def == DefCheck && (Def || Use == UseCheck) && "Invalid template")((void)0); | ||||||
| 1773 | #endif | ||||||
| 1774 | } else { | ||||||
| 1775 | // Find out what affect this instruction has on memory. | ||||||
| 1776 | ModRefInfo ModRef = AAP->getModRefInfo(I, None); | ||||||
| 1777 | // The isOrdered check is used to ensure that volatiles end up as defs | ||||||
| 1778 | // (atomics end up as ModRef right now anyway). Until we separate the | ||||||
| 1779 | // ordering chain from the memory chain, this enables people to see at least | ||||||
| 1780 | // some relative ordering to volatiles. Note that getClobberingMemoryAccess | ||||||
| 1781 | // will still give an answer that bypasses other volatile loads. TODO: | ||||||
| 1782 | // Separate memory aliasing and ordering into two different chains so that | ||||||
| 1783 | // we can precisely represent both "what memory will this read/write/is | ||||||
| 1784 | // clobbered by" and "what instructions can I move this past". | ||||||
| 1785 | Def = isModSet(ModRef) || isOrdered(I); | ||||||
| 1786 | Use = isRefSet(ModRef); | ||||||
| 1787 | } | ||||||
| 1788 | |||||||
| 1789 | // It's possible for an instruction to not modify memory at all. During | ||||||
| 1790 | // construction, we ignore them. | ||||||
| 1791 | if (!Def && !Use) | ||||||
| 1792 | return nullptr; | ||||||
| 1793 | |||||||
| 1794 | MemoryUseOrDef *MUD; | ||||||
| 1795 | if (Def) | ||||||
| 1796 | MUD = new MemoryDef(I->getContext(), nullptr, I, I->getParent(), NextID++); | ||||||
| 1797 | else | ||||||
| 1798 | MUD = new MemoryUse(I->getContext(), nullptr, I, I->getParent()); | ||||||
| 1799 | ValueToMemoryAccess[I] = MUD; | ||||||
| 1800 | return MUD; | ||||||
| 1801 | } | ||||||
| 1802 | |||||||
| 1803 | /// Properly remove \p MA from all of MemorySSA's lookup tables. | ||||||
| 1804 | void MemorySSA::removeFromLookups(MemoryAccess *MA) { | ||||||
| 1805 | assert(MA->use_empty() &&((void)0) | ||||||
| 1806 | "Trying to remove memory access that still has uses")((void)0); | ||||||
| 1807 | BlockNumbering.erase(MA); | ||||||
| 1808 | if (auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) | ||||||
| 1809 | MUD->setDefiningAccess(nullptr); | ||||||
| 1810 | // Invalidate our walker's cache if necessary | ||||||
| 1811 | if (!isa<MemoryUse>(MA)) | ||||||
| 1812 | getWalker()->invalidateInfo(MA); | ||||||
| 1813 | |||||||
| 1814 | Value *MemoryInst; | ||||||
| 1815 | if (const auto *MUD = dyn_cast<MemoryUseOrDef>(MA)) | ||||||
| 1816 | MemoryInst = MUD->getMemoryInst(); | ||||||
| 1817 | else | ||||||
| 1818 | MemoryInst = MA->getBlock(); | ||||||
| 1819 | |||||||
| 1820 | auto VMA = ValueToMemoryAccess.find(MemoryInst); | ||||||
| 1821 | if (VMA->second == MA) | ||||||
| 1822 | ValueToMemoryAccess.erase(VMA); | ||||||
| 1823 | } | ||||||
| 1824 | |||||||
| 1825 | /// Properly remove \p MA from all of MemorySSA's lists. | ||||||
| 1826 | /// | ||||||
| 1827 | /// Because of the way the intrusive list and use lists work, it is important to | ||||||
| 1828 | /// do removal in the right order. | ||||||
| 1829 | /// ShouldDelete defaults to true, and will cause the memory access to also be | ||||||
| 1830 | /// deleted, not just removed. | ||||||
| 1831 | void MemorySSA::removeFromLists(MemoryAccess *MA, bool ShouldDelete) { | ||||||
| 1832 | BasicBlock *BB = MA->getBlock(); | ||||||
| 1833 | // The access list owns the reference, so we erase it from the non-owning list | ||||||
| 1834 | // first. | ||||||
| 1835 | if (!isa<MemoryUse>(MA)) { | ||||||
| 1836 | auto DefsIt = PerBlockDefs.find(BB); | ||||||
| 1837 | std::unique_ptr<DefsList> &Defs = DefsIt->second; | ||||||
| 1838 | Defs->remove(*MA); | ||||||
| 1839 | if (Defs->empty()) | ||||||
| 1840 | PerBlockDefs.erase(DefsIt); | ||||||
| 1841 | } | ||||||
| 1842 | |||||||
| 1843 | // The erase call here will delete it. If we don't want it deleted, we call | ||||||
| 1844 | // remove instead. | ||||||
| 1845 | auto AccessIt = PerBlockAccesses.find(BB); | ||||||
| 1846 | std::unique_ptr<AccessList> &Accesses = AccessIt->second; | ||||||
| 1847 | if (ShouldDelete) | ||||||
| 1848 | Accesses->erase(MA); | ||||||
| 1849 | else | ||||||
| 1850 | Accesses->remove(MA); | ||||||
| 1851 | |||||||
| 1852 | if (Accesses->empty()) { | ||||||
| 1853 | PerBlockAccesses.erase(AccessIt); | ||||||
| 1854 | BlockNumberingValid.erase(BB); | ||||||
| 1855 | } | ||||||
| 1856 | } | ||||||
| 1857 | |||||||
| 1858 | void MemorySSA::print(raw_ostream &OS) const { | ||||||
| 1859 | MemorySSAAnnotatedWriter Writer(this); | ||||||
| 1860 | F.print(OS, &Writer); | ||||||
| 1861 | } | ||||||
| 1862 | |||||||
| 1863 | #if !defined(NDEBUG1) || defined(LLVM_ENABLE_DUMP) | ||||||
| 1864 | LLVM_DUMP_METHOD__attribute__((noinline)) void MemorySSA::dump() const { print(dbgs()); } | ||||||
| 1865 | #endif | ||||||
| 1866 | |||||||
| 1867 | void MemorySSA::verifyMemorySSA() const { | ||||||
| 1868 | verifyOrderingDominationAndDefUses(F); | ||||||
| 1869 | verifyDominationNumbers(F); | ||||||
| 1870 | verifyPrevDefInPhis(F); | ||||||
| 1871 | // Previously, the verification used to also verify that the clobberingAccess | ||||||
| 1872 | // cached by MemorySSA is the same as the clobberingAccess found at a later | ||||||
| 1873 | // query to AA. This does not hold true in general due to the current fragility | ||||||
| 1874 | // of BasicAA which has arbitrary caps on the things it analyzes before giving | ||||||
| 1875 | // up. As a result, transformations that are correct, will lead to BasicAA | ||||||
| 1876 | // returning different Alias answers before and after that transformation. | ||||||
| 1877 | // Invalidating MemorySSA is not an option, as the results in BasicAA can be so | ||||||
| 1878 | // random, in the worst case we'd need to rebuild MemorySSA from scratch after | ||||||
| 1879 | // every transformation, which defeats the purpose of using it. For such an | ||||||
| 1880 | // example, see test4 added in D51960. | ||||||
| 1881 | } | ||||||
| 1882 | |||||||
| 1883 | void MemorySSA::verifyPrevDefInPhis(Function &F) const { | ||||||
| 1884 | #if !defined(NDEBUG1) && defined(EXPENSIVE_CHECKS) | ||||||
| 1885 | for (const BasicBlock &BB : F) { | ||||||
| 1886 | if (MemoryPhi *Phi = getMemoryAccess(&BB)) { | ||||||
| 1887 | for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) { | ||||||
| 1888 | auto *Pred = Phi->getIncomingBlock(I); | ||||||
| 1889 | auto *IncAcc = Phi->getIncomingValue(I); | ||||||
| 1890 | // If Pred has no unreachable predecessors, get last def looking at | ||||||
| 1891 | // IDoms. If, while walkings IDoms, any of these has an unreachable | ||||||
| 1892 | // predecessor, then the incoming def can be any access. | ||||||
| 1893 | if (auto *DTNode = DT->getNode(Pred)) { | ||||||
| 1894 | while (DTNode) { | ||||||
| 1895 | if (auto *DefList = getBlockDefs(DTNode->getBlock())) { | ||||||
| 1896 | auto *LastAcc = &*(--DefList->end()); | ||||||
| 1897 | assert(LastAcc == IncAcc &&((void)0) | ||||||
| 1898 | "Incorrect incoming access into phi.")((void)0); | ||||||
| 1899 | break; | ||||||
| 1900 | } | ||||||
| 1901 | DTNode = DTNode->getIDom(); | ||||||
| 1902 | } | ||||||
| 1903 | } else { | ||||||
| 1904 | // If Pred has unreachable predecessors, but has at least a Def, the | ||||||
| 1905 | // incoming access can be the last Def in Pred, or it could have been | ||||||
| 1906 | // optimized to LoE. After an update, though, the LoE may have been | ||||||
| 1907 | // replaced by another access, so IncAcc may be any access. | ||||||
| 1908 | // If Pred has unreachable predecessors and no Defs, incoming access | ||||||
| 1909 | // should be LoE; However, after an update, it may be any access. | ||||||
| 1910 | } | ||||||
| 1911 | } | ||||||
| 1912 | } | ||||||
| 1913 | } | ||||||
| 1914 | #endif | ||||||
| 1915 | } | ||||||
| 1916 | |||||||
| 1917 | /// Verify that all of the blocks we believe to have valid domination numbers | ||||||
| 1918 | /// actually have valid domination numbers. | ||||||
| 1919 | void MemorySSA::verifyDominationNumbers(const Function &F) const { | ||||||
| 1920 | #ifndef NDEBUG1 | ||||||
| 1921 | if (BlockNumberingValid.empty()) | ||||||
| 1922 | return; | ||||||
| 1923 | |||||||
| 1924 | SmallPtrSet<const BasicBlock *, 16> ValidBlocks = BlockNumberingValid; | ||||||
| 1925 | for (const BasicBlock &BB : F) { | ||||||
| 1926 | if (!ValidBlocks.count(&BB)) | ||||||
| 1927 | continue; | ||||||
| 1928 | |||||||
| 1929 | ValidBlocks.erase(&BB); | ||||||
| 1930 | |||||||
| 1931 | const AccessList *Accesses = getBlockAccesses(&BB); | ||||||
| 1932 | // It's correct to say an empty block has valid numbering. | ||||||
| 1933 | if (!Accesses) | ||||||
| 1934 | continue; | ||||||
| 1935 | |||||||
| 1936 | // Block numbering starts at 1. | ||||||
| 1937 | unsigned long LastNumber = 0; | ||||||
| 1938 | for (const MemoryAccess &MA : *Accesses) { | ||||||
| 1939 | auto ThisNumberIter = BlockNumbering.find(&MA); | ||||||
| 1940 | assert(ThisNumberIter != BlockNumbering.end() &&((void)0) | ||||||
| 1941 | "MemoryAccess has no domination number in a valid block!")((void)0); | ||||||
| 1942 | |||||||
| 1943 | unsigned long ThisNumber = ThisNumberIter->second; | ||||||
| 1944 | assert(ThisNumber > LastNumber &&((void)0) | ||||||
| 1945 | "Domination numbers should be strictly increasing!")((void)0); | ||||||
| 1946 | LastNumber = ThisNumber; | ||||||
| 1947 | } | ||||||
| 1948 | } | ||||||
| 1949 | |||||||
| 1950 | assert(ValidBlocks.empty() &&((void)0) | ||||||
| 1951 | "All valid BasicBlocks should exist in F -- dangling pointers?")((void)0); | ||||||
| 1952 | #endif | ||||||
| 1953 | } | ||||||
| 1954 | |||||||
| 1955 | /// Verify ordering: the order and existence of MemoryAccesses matches the | ||||||
| 1956 | /// order and existence of memory affecting instructions. | ||||||
| 1957 | /// Verify domination: each definition dominates all of its uses. | ||||||
| 1958 | /// Verify def-uses: the immediate use information - walk all the memory | ||||||
| 1959 | /// accesses and verifying that, for each use, it appears in the appropriate | ||||||
| 1960 | /// def's use list | ||||||
| 1961 | void MemorySSA::verifyOrderingDominationAndDefUses(Function &F) const { | ||||||
| 1962 | #if !defined(NDEBUG1) | ||||||
| 1963 | // Walk all the blocks, comparing what the lookups think and what the access | ||||||
| 1964 | // lists think, as well as the order in the blocks vs the order in the access | ||||||
| 1965 | // lists. | ||||||
| 1966 | SmallVector<MemoryAccess *, 32> ActualAccesses; | ||||||
| 1967 | SmallVector<MemoryAccess *, 32> ActualDefs; | ||||||
| 1968 | for (BasicBlock &B : F) { | ||||||
| 1969 | const AccessList *AL = getBlockAccesses(&B); | ||||||
| 1970 | const auto *DL = getBlockDefs(&B); | ||||||
| 1971 | MemoryPhi *Phi = getMemoryAccess(&B); | ||||||
| 1972 | if (Phi) { | ||||||
| 1973 | // Verify ordering. | ||||||
| 1974 | ActualAccesses.push_back(Phi); | ||||||
| 1975 | ActualDefs.push_back(Phi); | ||||||
| 1976 | // Verify domination | ||||||
| 1977 | for (const Use &U : Phi->uses()) | ||||||
| 1978 | assert(dominates(Phi, U) && "Memory PHI does not dominate it's uses")((void)0); | ||||||
| 1979 | #if defined(EXPENSIVE_CHECKS) | ||||||
| 1980 | // Verify def-uses. | ||||||
| 1981 | assert(Phi->getNumOperands() == static_cast<unsigned>(std::distance(((void)0) | ||||||
| 1982 | pred_begin(&B), pred_end(&B))) &&((void)0) | ||||||
| 1983 | "Incomplete MemoryPhi Node")((void)0); | ||||||
| 1984 | for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) { | ||||||
| 1985 | verifyUseInDefs(Phi->getIncomingValue(I), Phi); | ||||||
| 1986 | assert(is_contained(predecessors(&B), Phi->getIncomingBlock(I)) &&((void)0) | ||||||
| 1987 | "Incoming phi block not a block predecessor")((void)0); | ||||||
| 1988 | } | ||||||
| 1989 | #endif | ||||||
| 1990 | } | ||||||
| 1991 | |||||||
| 1992 | for (Instruction &I : B) { | ||||||
| 1993 | MemoryUseOrDef *MA = getMemoryAccess(&I); | ||||||
| 1994 | assert((!MA || (AL && (isa<MemoryUse>(MA) || DL))) &&((void)0) | ||||||
| 1995 | "We have memory affecting instructions "((void)0) | ||||||
| 1996 | "in this block but they are not in the "((void)0) | ||||||
| 1997 | "access list or defs list")((void)0); | ||||||
| 1998 | if (MA) { | ||||||
| 1999 | // Verify ordering. | ||||||
| 2000 | ActualAccesses.push_back(MA); | ||||||
| 2001 | if (MemoryAccess *MD = dyn_cast<MemoryDef>(MA)) { | ||||||
| 2002 | // Verify ordering. | ||||||
| 2003 | ActualDefs.push_back(MA); | ||||||
| 2004 | // Verify domination. | ||||||
| 2005 | for (const Use &U : MD->uses()) | ||||||
| 2006 | assert(dominates(MD, U) &&((void)0) | ||||||
| 2007 | "Memory Def does not dominate it's uses")((void)0); | ||||||
| 2008 | } | ||||||
| 2009 | #if defined(EXPENSIVE_CHECKS) | ||||||
| 2010 | // Verify def-uses. | ||||||
| 2011 | verifyUseInDefs(MA->getDefiningAccess(), MA); | ||||||
| 2012 | #endif | ||||||
| 2013 | } | ||||||
| 2014 | } | ||||||
| 2015 | // Either we hit the assert, really have no accesses, or we have both | ||||||
| 2016 | // accesses and an access list. Same with defs. | ||||||
| 2017 | if (!AL && !DL) | ||||||
| 2018 | continue; | ||||||
| 2019 | // Verify ordering. | ||||||
| 2020 | assert(AL->size() == ActualAccesses.size() &&((void)0) | ||||||
| 2021 | "We don't have the same number of accesses in the block as on the "((void)0) | ||||||
| 2022 | "access list")((void)0); | ||||||
| 2023 | assert((DL || ActualDefs.size() == 0) &&((void)0) | ||||||
| 2024 | "Either we should have a defs list, or we should have no defs")((void)0); | ||||||
| 2025 | assert((!DL || DL->size() == ActualDefs.size()) &&((void)0) | ||||||
| 2026 | "We don't have the same number of defs in the block as on the "((void)0) | ||||||
| 2027 | "def list")((void)0); | ||||||
| 2028 | auto ALI = AL->begin(); | ||||||
| 2029 | auto AAI = ActualAccesses.begin(); | ||||||
| 2030 | while (ALI != AL->end() && AAI != ActualAccesses.end()) { | ||||||
| 2031 | assert(&*ALI == *AAI && "Not the same accesses in the same order")((void)0); | ||||||
| 2032 | ++ALI; | ||||||
| 2033 | ++AAI; | ||||||
| 2034 | } | ||||||
| 2035 | ActualAccesses.clear(); | ||||||
| 2036 | if (DL) { | ||||||
| 2037 | auto DLI = DL->begin(); | ||||||
| 2038 | auto ADI = ActualDefs.begin(); | ||||||
| 2039 | while (DLI != DL->end() && ADI != ActualDefs.end()) { | ||||||
| 2040 | assert(&*DLI == *ADI && "Not the same defs in the same order")((void)0); | ||||||
| 2041 | ++DLI; | ||||||
| 2042 | ++ADI; | ||||||
| 2043 | } | ||||||
| 2044 | } | ||||||
| 2045 | ActualDefs.clear(); | ||||||
| 2046 | } | ||||||
| 2047 | #endif | ||||||
| 2048 | } | ||||||
| 2049 | |||||||
| 2050 | /// Verify the def-use lists in MemorySSA, by verifying that \p Use | ||||||
| 2051 | /// appears in the use list of \p Def. | ||||||
| 2052 | void MemorySSA::verifyUseInDefs(MemoryAccess *Def, MemoryAccess *Use) const { | ||||||
| 2053 | #ifndef NDEBUG1 | ||||||
| 2054 | // The live on entry use may cause us to get a NULL def here | ||||||
| 2055 | if (!Def) | ||||||
| 2056 | assert(isLiveOnEntryDef(Use) &&((void)0) | ||||||
| 2057 | "Null def but use not point to live on entry def")((void)0); | ||||||
| 2058 | else | ||||||
| 2059 | assert(is_contained(Def->users(), Use) &&((void)0) | ||||||
| 2060 | "Did not find use in def's use list")((void)0); | ||||||
| 2061 | #endif | ||||||
| 2062 | } | ||||||
| 2063 | |||||||
| 2064 | /// Perform a local numbering on blocks so that instruction ordering can be | ||||||
| 2065 | /// determined in constant time. | ||||||
| 2066 | /// TODO: We currently just number in order. If we numbered by N, we could | ||||||
| 2067 | /// allow at least N-1 sequences of insertBefore or insertAfter (and at least | ||||||
| 2068 | /// log2(N) sequences of mixed before and after) without needing to invalidate | ||||||
| 2069 | /// the numbering. | ||||||
| 2070 | void MemorySSA::renumberBlock(const BasicBlock *B) const { | ||||||
| 2071 | // The pre-increment ensures the numbers really start at 1. | ||||||
| 2072 | unsigned long CurrentNumber = 0; | ||||||
| 2073 | const AccessList *AL = getBlockAccesses(B); | ||||||
| 2074 | assert(AL != nullptr && "Asking to renumber an empty block")((void)0); | ||||||
| 2075 | for (const auto &I : *AL) | ||||||
| 2076 | BlockNumbering[&I] = ++CurrentNumber; | ||||||
| 2077 | BlockNumberingValid.insert(B); | ||||||
| 2078 | } | ||||||
| 2079 | |||||||
| 2080 | /// Determine, for two memory accesses in the same block, | ||||||
| 2081 | /// whether \p Dominator dominates \p Dominatee. | ||||||
| 2082 | /// \returns True if \p Dominator dominates \p Dominatee. | ||||||
| 2083 | bool MemorySSA::locallyDominates(const MemoryAccess *Dominator, | ||||||
| 2084 | const MemoryAccess *Dominatee) const { | ||||||
| 2085 | const BasicBlock *DominatorBlock = Dominator->getBlock(); | ||||||
| 2086 | |||||||
| 2087 | assert((DominatorBlock == Dominatee->getBlock()) &&((void)0) | ||||||
| 2088 | "Asking for local domination when accesses are in different blocks!")((void)0); | ||||||
| 2089 | // A node dominates itself. | ||||||
| 2090 | if (Dominatee == Dominator) | ||||||
| 2091 | return true; | ||||||
| 2092 | |||||||
| 2093 | // When Dominatee is defined on function entry, it is not dominated by another | ||||||
| 2094 | // memory access. | ||||||
| 2095 | if (isLiveOnEntryDef(Dominatee)) | ||||||
| 2096 | return false; | ||||||
| 2097 | |||||||
| 2098 | // When Dominator is defined on function entry, it dominates the other memory | ||||||
| 2099 | // access. | ||||||
| 2100 | if (isLiveOnEntryDef(Dominator)) | ||||||
| 2101 | return true; | ||||||
| 2102 | |||||||
| 2103 | if (!BlockNumberingValid.count(DominatorBlock)) | ||||||
| 2104 | renumberBlock(DominatorBlock); | ||||||
| 2105 | |||||||
| 2106 | unsigned long DominatorNum = BlockNumbering.lookup(Dominator); | ||||||
| 2107 | // All numbers start with 1 | ||||||
| 2108 | assert(DominatorNum != 0 && "Block was not numbered properly")((void)0); | ||||||
| 2109 | unsigned long DominateeNum = BlockNumbering.lookup(Dominatee); | ||||||
| 2110 | assert(DominateeNum != 0 && "Block was not numbered properly")((void)0); | ||||||
| 2111 | return DominatorNum < DominateeNum; | ||||||
| 2112 | } | ||||||
| 2113 | |||||||
| 2114 | bool MemorySSA::dominates(const MemoryAccess *Dominator, | ||||||
| 2115 | const MemoryAccess *Dominatee) const { | ||||||
| 2116 | if (Dominator == Dominatee) | ||||||
| 2117 | return true; | ||||||
| 2118 | |||||||
| 2119 | if (isLiveOnEntryDef(Dominatee)) | ||||||
| 2120 | return false; | ||||||
| 2121 | |||||||
| 2122 | if (Dominator->getBlock() != Dominatee->getBlock()) | ||||||
| 2123 | return DT->dominates(Dominator->getBlock(), Dominatee->getBlock()); | ||||||
| 2124 | return locallyDominates(Dominator, Dominatee); | ||||||
| 2125 | } | ||||||
| 2126 | |||||||
| 2127 | bool MemorySSA::dominates(const MemoryAccess *Dominator, | ||||||
| 2128 | const Use &Dominatee) const { | ||||||
| 2129 | if (MemoryPhi *MP = dyn_cast<MemoryPhi>(Dominatee.getUser())) { | ||||||
| 2130 | BasicBlock *UseBB = MP->getIncomingBlock(Dominatee); | ||||||
| 2131 | // The def must dominate the incoming block of the phi. | ||||||
| 2132 | if (UseBB != Dominator->getBlock()) | ||||||
| 2133 | return DT->dominates(Dominator->getBlock(), UseBB); | ||||||
| 2134 | // If the UseBB and the DefBB are the same, compare locally. | ||||||
| 2135 | return locallyDominates(Dominator, cast<MemoryAccess>(Dominatee)); | ||||||
| 2136 | } | ||||||
| 2137 | // If it's not a PHI node use, the normal dominates can already handle it. | ||||||
| 2138 | return dominates(Dominator, cast<MemoryAccess>(Dominatee.getUser())); | ||||||
| 2139 | } | ||||||
| 2140 | |||||||
| 2141 | const static char LiveOnEntryStr[] = "liveOnEntry"; | ||||||
| 2142 | |||||||
| 2143 | void MemoryAccess::print(raw_ostream &OS) const { | ||||||
| 2144 | switch (getValueID()) { | ||||||
| 2145 | case MemoryPhiVal: return static_cast<const MemoryPhi *>(this)->print(OS); | ||||||
| 2146 | case MemoryDefVal: return static_cast<const MemoryDef *>(this)->print(OS); | ||||||
| 2147 | case MemoryUseVal: return static_cast<const MemoryUse *>(this)->print(OS); | ||||||
| 2148 | } | ||||||
| 2149 | llvm_unreachable("invalid value id")__builtin_unreachable(); | ||||||
| 2150 | } | ||||||
| 2151 | |||||||
| 2152 | void MemoryDef::print(raw_ostream &OS) const { | ||||||
| 2153 | MemoryAccess *UO = getDefiningAccess(); | ||||||
| 2154 | |||||||
| 2155 | auto printID = [&OS](MemoryAccess *A) { | ||||||
| 2156 | if (A && A->getID()) | ||||||
| 2157 | OS << A->getID(); | ||||||
| 2158 | else | ||||||
| 2159 | OS << LiveOnEntryStr; | ||||||
| 2160 | }; | ||||||
| 2161 | |||||||
| 2162 | OS << getID() << " = MemoryDef("; | ||||||
| 2163 | printID(UO); | ||||||
| 2164 | OS << ")"; | ||||||
| 2165 | |||||||
| 2166 | if (isOptimized()) { | ||||||
| 2167 | OS << "->"; | ||||||
| 2168 | printID(getOptimized()); | ||||||
| 2169 | |||||||
| 2170 | if (Optional<AliasResult> AR = getOptimizedAccessType()) | ||||||
| 2171 | OS << " " << *AR; | ||||||
| 2172 | } | ||||||
| 2173 | } | ||||||
| 2174 | |||||||
| 2175 | void MemoryPhi::print(raw_ostream &OS) const { | ||||||
| 2176 | ListSeparator LS(","); | ||||||
| 2177 | OS << getID() << " = MemoryPhi("; | ||||||
| 2178 | for (const auto &Op : operands()) { | ||||||
| 2179 | BasicBlock *BB = getIncomingBlock(Op); | ||||||
| 2180 | MemoryAccess *MA = cast<MemoryAccess>(Op); | ||||||
| 2181 | |||||||
| 2182 | OS << LS << '{'; | ||||||
| 2183 | if (BB->hasName()) | ||||||
| 2184 | OS << BB->getName(); | ||||||
| 2185 | else | ||||||
| 2186 | BB->printAsOperand(OS, false); | ||||||
| 2187 | OS << ','; | ||||||
| 2188 | if (unsigned ID = MA->getID()) | ||||||
| 2189 | OS << ID; | ||||||
| 2190 | else | ||||||
| 2191 | OS << LiveOnEntryStr; | ||||||
| 2192 | OS << '}'; | ||||||
| 2193 | } | ||||||
| 2194 | OS << ')'; | ||||||
| 2195 | } | ||||||
| 2196 | |||||||
| 2197 | void MemoryUse::print(raw_ostream &OS) const { | ||||||
| 2198 | MemoryAccess *UO = getDefiningAccess(); | ||||||
| 2199 | OS << "MemoryUse("; | ||||||
| 2200 | if (UO && UO->getID()) | ||||||
| 2201 | OS << UO->getID(); | ||||||
| 2202 | else | ||||||
| 2203 | OS << LiveOnEntryStr; | ||||||
| 2204 | OS << ')'; | ||||||
| 2205 | |||||||
| 2206 | if (Optional<AliasResult> AR = getOptimizedAccessType()) | ||||||
| 2207 | OS << " " << *AR; | ||||||
| 2208 | } | ||||||
| 2209 | |||||||
| 2210 | void MemoryAccess::dump() const { | ||||||
| 2211 | // Cannot completely remove virtual function even in release mode. | ||||||
| 2212 | #if !defined(NDEBUG1) || defined(LLVM_ENABLE_DUMP) | ||||||
| 2213 | print(dbgs()); | ||||||
| 2214 | dbgs() << "\n"; | ||||||
| 2215 | #endif | ||||||
| 2216 | } | ||||||
| 2217 | |||||||
| 2218 | char MemorySSAPrinterLegacyPass::ID = 0; | ||||||
| 2219 | |||||||
| 2220 | MemorySSAPrinterLegacyPass::MemorySSAPrinterLegacyPass() : FunctionPass(ID) { | ||||||
| 2221 | initializeMemorySSAPrinterLegacyPassPass(*PassRegistry::getPassRegistry()); | ||||||
| 2222 | } | ||||||
| 2223 | |||||||
| 2224 | void MemorySSAPrinterLegacyPass::getAnalysisUsage(AnalysisUsage &AU) const { | ||||||
| 2225 | AU.setPreservesAll(); | ||||||
| 2226 | AU.addRequired<MemorySSAWrapperPass>(); | ||||||
| 2227 | } | ||||||
| 2228 | |||||||
| 2229 | class DOTFuncMSSAInfo { | ||||||
| 2230 | private: | ||||||
| 2231 | const Function &F; | ||||||
| 2232 | MemorySSAAnnotatedWriter MSSAWriter; | ||||||
| 2233 | |||||||
| 2234 | public: | ||||||
| 2235 | DOTFuncMSSAInfo(const Function &F, MemorySSA &MSSA) | ||||||
| 2236 | : F(F), MSSAWriter(&MSSA) {} | ||||||
| 2237 | |||||||
| 2238 | const Function *getFunction() { return &F; } | ||||||
| 2239 | MemorySSAAnnotatedWriter &getWriter() { return MSSAWriter; } | ||||||
| 2240 | }; | ||||||
| 2241 | |||||||
| 2242 | namespace llvm { | ||||||
| 2243 | |||||||
| 2244 | template <> | ||||||
| 2245 | struct GraphTraits<DOTFuncMSSAInfo *> : public GraphTraits<const BasicBlock *> { | ||||||
| 2246 | static NodeRef getEntryNode(DOTFuncMSSAInfo *CFGInfo) { | ||||||
| 2247 | return &(CFGInfo->getFunction()->getEntryBlock()); | ||||||
| 2248 | } | ||||||
| 2249 | |||||||
| 2250 | // nodes_iterator/begin/end - Allow iteration over all nodes in the graph | ||||||
| 2251 | using nodes_iterator = pointer_iterator<Function::const_iterator>; | ||||||
| 2252 | |||||||
| 2253 | static nodes_iterator nodes_begin(DOTFuncMSSAInfo *CFGInfo) { | ||||||
| 2254 | return nodes_iterator(CFGInfo->getFunction()->begin()); | ||||||
| 2255 | } | ||||||
| 2256 | |||||||
| 2257 | static nodes_iterator nodes_end(DOTFuncMSSAInfo *CFGInfo) { | ||||||
| 2258 | return nodes_iterator(CFGInfo->getFunction()->end()); | ||||||
| 2259 | } | ||||||
| 2260 | |||||||
| 2261 | static size_t size(DOTFuncMSSAInfo *CFGInfo) { | ||||||
| 2262 | return CFGInfo->getFunction()->size(); | ||||||
| 2263 | } | ||||||
| 2264 | }; | ||||||
| 2265 | |||||||
| 2266 | template <> | ||||||
| 2267 | struct DOTGraphTraits<DOTFuncMSSAInfo *> : public DefaultDOTGraphTraits { | ||||||
| 2268 | |||||||
| 2269 | DOTGraphTraits(bool IsSimple = false) : DefaultDOTGraphTraits(IsSimple) {} | ||||||
| 2270 | |||||||
| 2271 | static std::string getGraphName(DOTFuncMSSAInfo *CFGInfo) { | ||||||
| 2272 | return "MSSA CFG for '" + CFGInfo->getFunction()->getName().str() + | ||||||
| 2273 | "' function"; | ||||||
| 2274 | } | ||||||
| 2275 | |||||||
| 2276 | std::string getNodeLabel(const BasicBlock *Node, DOTFuncMSSAInfo *CFGInfo) { | ||||||
| 2277 | return DOTGraphTraits<DOTFuncInfo *>::getCompleteNodeLabel( | ||||||
| 2278 | Node, nullptr, | ||||||
| 2279 | [CFGInfo](raw_string_ostream &OS, const BasicBlock &BB) -> void { | ||||||
| 2280 | BB.print(OS, &CFGInfo->getWriter(), true, true); | ||||||
| 2281 | }, | ||||||
| 2282 | [](std::string &S, unsigned &I, unsigned Idx) -> void { | ||||||
| 2283 | std::string Str = S.substr(I, Idx - I); | ||||||
| 2284 | StringRef SR = Str; | ||||||
| 2285 | if (SR.count(" = MemoryDef(") || SR.count(" = MemoryPhi(") || | ||||||
| 2286 | SR.count("MemoryUse(")) | ||||||
| 2287 | return; | ||||||
| 2288 | DOTGraphTraits<DOTFuncInfo *>::eraseComment(S, I, Idx); | ||||||
| 2289 | }); | ||||||
| 2290 | } | ||||||
| 2291 | |||||||
| 2292 | static std::string getEdgeSourceLabel(const BasicBlock *Node, | ||||||
| 2293 | const_succ_iterator I) { | ||||||
| 2294 | return DOTGraphTraits<DOTFuncInfo *>::getEdgeSourceLabel(Node, I); | ||||||
| 2295 | } | ||||||
| 2296 | |||||||
| 2297 | /// Display the raw branch weights from PGO. | ||||||
| 2298 | std::string getEdgeAttributes(const BasicBlock *Node, const_succ_iterator I, | ||||||
| 2299 | DOTFuncMSSAInfo *CFGInfo) { | ||||||
| 2300 | return ""; | ||||||
| 2301 | } | ||||||
| 2302 | |||||||
| 2303 | std::string getNodeAttributes(const BasicBlock *Node, | ||||||
| 2304 | DOTFuncMSSAInfo *CFGInfo) { | ||||||
| 2305 | return getNodeLabel(Node, CFGInfo).find(';') != std::string::npos | ||||||
| 2306 | ? "style=filled, fillcolor=lightpink" | ||||||
| 2307 | : ""; | ||||||
| 2308 | } | ||||||
| 2309 | }; | ||||||
| 2310 | |||||||
| 2311 | } // namespace llvm | ||||||
| 2312 | |||||||
| 2313 | bool MemorySSAPrinterLegacyPass::runOnFunction(Function &F) { | ||||||
| 2314 | auto &MSSA = getAnalysis<MemorySSAWrapperPass>().getMSSA(); | ||||||
| 2315 | if (DotCFGMSSA != "") { | ||||||
| 2316 | DOTFuncMSSAInfo CFGInfo(F, MSSA); | ||||||
| 2317 | WriteGraph(&CFGInfo, "", false, "MSSA", DotCFGMSSA); | ||||||
| 2318 | } else | ||||||
| 2319 | MSSA.print(dbgs()); | ||||||
| 2320 | |||||||
| 2321 | if (VerifyMemorySSA) | ||||||
| 2322 | MSSA.verifyMemorySSA(); | ||||||
| 2323 | return false; | ||||||
| 2324 | } | ||||||
| 2325 | |||||||
| 2326 | AnalysisKey MemorySSAAnalysis::Key; | ||||||
| 2327 | |||||||
| 2328 | MemorySSAAnalysis::Result MemorySSAAnalysis::run(Function &F, | ||||||
| 2329 | FunctionAnalysisManager &AM) { | ||||||
| 2330 | auto &DT = AM.getResult<DominatorTreeAnalysis>(F); | ||||||
| 2331 | auto &AA = AM.getResult<AAManager>(F); | ||||||
| 2332 | return MemorySSAAnalysis::Result(std::make_unique<MemorySSA>(F, &AA, &DT)); | ||||||
| 2333 | } | ||||||
| 2334 | |||||||
| 2335 | bool MemorySSAAnalysis::Result::invalidate( | ||||||
| 2336 | Function &F, const PreservedAnalyses &PA, | ||||||
| 2337 | FunctionAnalysisManager::Invalidator &Inv) { | ||||||
| 2338 | auto PAC = PA.getChecker<MemorySSAAnalysis>(); | ||||||
| 2339 | return !(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Function>>()) || | ||||||
| 2340 | Inv.invalidate<AAManager>(F, PA) || | ||||||
| 2341 | Inv.invalidate<DominatorTreeAnalysis>(F, PA); | ||||||
| 2342 | } | ||||||
| 2343 | |||||||
| 2344 | PreservedAnalyses MemorySSAPrinterPass::run(Function &F, | ||||||
| 2345 | FunctionAnalysisManager &AM) { | ||||||
| 2346 | auto &MSSA = AM.getResult<MemorySSAAnalysis>(F).getMSSA(); | ||||||
| 2347 | if (DotCFGMSSA != "") { | ||||||
| 2348 | DOTFuncMSSAInfo CFGInfo(F, MSSA); | ||||||
| 2349 | WriteGraph(&CFGInfo, "", false, "MSSA", DotCFGMSSA); | ||||||
| 2350 | } else { | ||||||
| 2351 | OS << "MemorySSA for function: " << F.getName() << "\n"; | ||||||
| 2352 | MSSA.print(OS); | ||||||
| 2353 | } | ||||||
| 2354 | |||||||
| 2355 | return PreservedAnalyses::all(); | ||||||
| 2356 | } | ||||||
| 2357 | |||||||
| 2358 | PreservedAnalyses MemorySSAVerifierPass::run(Function &F, | ||||||
| 2359 | FunctionAnalysisManager &AM) { | ||||||
| 2360 | AM.getResult<MemorySSAAnalysis>(F).getMSSA().verifyMemorySSA(); | ||||||
| 2361 | |||||||
| 2362 | return PreservedAnalyses::all(); | ||||||
| 2363 | } | ||||||
| 2364 | |||||||
| 2365 | char MemorySSAWrapperPass::ID = 0; | ||||||
| 2366 | |||||||
| 2367 | MemorySSAWrapperPass::MemorySSAWrapperPass() : FunctionPass(ID) { | ||||||
| 2368 | initializeMemorySSAWrapperPassPass(*PassRegistry::getPassRegistry()); | ||||||
| 2369 | } | ||||||
| 2370 | |||||||
| 2371 | void MemorySSAWrapperPass::releaseMemory() { MSSA.reset(); } | ||||||
| 2372 | |||||||
| 2373 | void MemorySSAWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const { | ||||||
| 2374 | AU.setPreservesAll(); | ||||||
| 2375 | AU.addRequiredTransitive<DominatorTreeWrapperPass>(); | ||||||
| 2376 | AU.addRequiredTransitive<AAResultsWrapperPass>(); | ||||||
| 2377 | } | ||||||
| 2378 | |||||||
| 2379 | bool MemorySSAWrapperPass::runOnFunction(Function &F) { | ||||||
| 2380 | auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); | ||||||
| 2381 | auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults(); | ||||||
| 2382 | MSSA.reset(new MemorySSA(F, &AA, &DT)); | ||||||
| 2383 | return false; | ||||||
| 2384 | } | ||||||
| 2385 | |||||||
| 2386 | void MemorySSAWrapperPass::verifyAnalysis() const { | ||||||
| 2387 | if (VerifyMemorySSA) | ||||||
| 2388 | MSSA->verifyMemorySSA(); | ||||||
| 2389 | } | ||||||
| 2390 | |||||||
| 2391 | void MemorySSAWrapperPass::print(raw_ostream &OS, const Module *M) const { | ||||||
| 2392 | MSSA->print(OS); | ||||||
| 2393 | } | ||||||
| 2394 | |||||||
| 2395 | MemorySSAWalker::MemorySSAWalker(MemorySSA *M) : MSSA(M) {} | ||||||
| 2396 | |||||||
| 2397 | /// Walk the use-def chains starting at \p StartingAccess and find | ||||||
| 2398 | /// the MemoryAccess that actually clobbers Loc. | ||||||
| 2399 | /// | ||||||
| 2400 | /// \returns our clobbering memory access | ||||||
| 2401 | template <typename AliasAnalysisType> | ||||||
| 2402 | MemoryAccess * | ||||||
| 2403 | MemorySSA::ClobberWalkerBase<AliasAnalysisType>::getClobberingMemoryAccessBase( | ||||||
| 2404 | MemoryAccess *StartingAccess, const MemoryLocation &Loc, | ||||||
| 2405 | unsigned &UpwardWalkLimit) { | ||||||
| 2406 | assert(!isa<MemoryUse>(StartingAccess) && "Use cannot be defining access")((void)0); | ||||||
| 2407 | |||||||
| 2408 | Instruction *I = nullptr; | ||||||
| 2409 | if (auto *StartingUseOrDef
| ||||||
| 2410 | if (MSSA->isLiveOnEntryDef(StartingUseOrDef)) | ||||||
| 2411 | return StartingUseOrDef; | ||||||
| 2412 | |||||||
| 2413 | I = StartingUseOrDef->getMemoryInst(); | ||||||
| 2414 | |||||||
| 2415 | // Conservatively, fences are always clobbers, so don't perform the walk if | ||||||
| 2416 | // we hit a fence. | ||||||
| 2417 | if (!isa<CallBase>(I) && I->isFenceLike()) | ||||||
| 2418 | return StartingUseOrDef; | ||||||
| 2419 | } | ||||||
| 2420 | |||||||
| 2421 | UpwardsMemoryQuery Q; | ||||||
| 2422 | Q.OriginalAccess = StartingAccess; | ||||||
| 2423 | Q.StartingLoc = Loc; | ||||||
| 2424 | Q.Inst = nullptr; | ||||||
| 2425 | Q.IsCall = false; | ||||||
| 2426 | |||||||
| 2427 | // Unlike the other function, do not walk to the def of a def, because we are | ||||||
| 2428 | // handed something we already believe is the clobbering access. | ||||||
| 2429 | // We never set SkipSelf to true in Q in this method. | ||||||
| 2430 | MemoryAccess *Clobber = | ||||||
| 2431 | Walker.findClobber(StartingAccess, Q, UpwardWalkLimit); | ||||||
| 2432 | LLVM_DEBUG({do { } while (false) | ||||||
| 2433 | dbgs() << "Clobber starting at access " << *StartingAccess << "\n";do { } while (false) | ||||||
| 2434 | if (I)do { } while (false) | ||||||
| 2435 | dbgs() << " for instruction " << *I << "\n";do { } while (false) | ||||||
| 2436 | dbgs() << " is " << *Clobber << "\n";do { } while (false) | ||||||
| 2437 | })do { } while (false); | ||||||
| 2438 | return Clobber; | ||||||
| 2439 | } | ||||||
| 2440 | |||||||
| 2441 | template <typename AliasAnalysisType> | ||||||
| 2442 | MemoryAccess * | ||||||
| 2443 | MemorySSA::ClobberWalkerBase<AliasAnalysisType>::getClobberingMemoryAccessBase( | ||||||
| 2444 | MemoryAccess *MA, unsigned &UpwardWalkLimit, bool SkipSelf) { | ||||||
| 2445 | auto *StartingAccess = dyn_cast<MemoryUseOrDef>(MA); | ||||||
| 2446 | // If this is a MemoryPhi, we can't do anything. | ||||||
| 2447 | if (!StartingAccess) | ||||||
| 2448 | return MA; | ||||||
| 2449 | |||||||
| 2450 | bool IsOptimized = false; | ||||||
| 2451 | |||||||
| 2452 | // If this is an already optimized use or def, return the optimized result. | ||||||
| 2453 | // Note: Currently, we store the optimized def result in a separate field, | ||||||
| 2454 | // since we can't use the defining access. | ||||||
| 2455 | if (StartingAccess->isOptimized()) { | ||||||
| 2456 | if (!SkipSelf || !isa<MemoryDef>(StartingAccess)) | ||||||
| 2457 | return StartingAccess->getOptimized(); | ||||||
| 2458 | IsOptimized = true; | ||||||
| 2459 | } | ||||||
| 2460 | |||||||
| 2461 | const Instruction *I = StartingAccess->getMemoryInst(); | ||||||
| 2462 | // We can't sanely do anything with a fence, since they conservatively clobber | ||||||
| 2463 | // all memory, and have no locations to get pointers from to try to | ||||||
| 2464 | // disambiguate. | ||||||
| 2465 | if (!isa<CallBase>(I) && I->isFenceLike()) | ||||||
| 2466 | return StartingAccess; | ||||||
| 2467 | |||||||
| 2468 | UpwardsMemoryQuery Q(I, StartingAccess); | ||||||
| 2469 | |||||||
| 2470 | if (isUseTriviallyOptimizableToLiveOnEntry(*Walker.getAA(), I)) { | ||||||
| 2471 | MemoryAccess *LiveOnEntry = MSSA->getLiveOnEntryDef(); | ||||||
| 2472 | StartingAccess->setOptimized(LiveOnEntry); | ||||||
| 2473 | StartingAccess->setOptimizedAccessType(None); | ||||||
| 2474 | return LiveOnEntry; | ||||||
| 2475 | } | ||||||
| 2476 | |||||||
| 2477 | MemoryAccess *OptimizedAccess; | ||||||
| 2478 | if (!IsOptimized) { | ||||||
| 2479 | // Start with the thing we already think clobbers this location | ||||||
| 2480 | MemoryAccess *DefiningAccess = StartingAccess->getDefiningAccess(); | ||||||
| 2481 | |||||||
| 2482 | // At this point, DefiningAccess may be the live on entry def. | ||||||
| 2483 | // If it is, we will not get a better result. | ||||||
| 2484 | if (MSSA->isLiveOnEntryDef(DefiningAccess)) { | ||||||
| 2485 | StartingAccess->setOptimized(DefiningAccess); | ||||||
| 2486 | StartingAccess->setOptimizedAccessType(None); | ||||||
| 2487 | return DefiningAccess; | ||||||
| 2488 | } | ||||||
| 2489 | |||||||
| 2490 | OptimizedAccess = Walker.findClobber(DefiningAccess, Q, UpwardWalkLimit); | ||||||
| 2491 | StartingAccess->setOptimized(OptimizedAccess); | ||||||
| 2492 | if (MSSA->isLiveOnEntryDef(OptimizedAccess)) | ||||||
| 2493 | StartingAccess->setOptimizedAccessType(None); | ||||||
| 2494 | else if (Q.AR && *Q.AR == AliasResult::MustAlias) | ||||||
| 2495 | StartingAccess->setOptimizedAccessType( | ||||||
| 2496 | AliasResult(AliasResult::MustAlias)); | ||||||
| 2497 | } else | ||||||
| 2498 | OptimizedAccess = StartingAccess->getOptimized(); | ||||||
| 2499 | |||||||
| 2500 | LLVM_DEBUG(dbgs() << "Starting Memory SSA clobber for " << *I << " is ")do { } while (false); | ||||||
| 2501 | LLVM_DEBUG(dbgs() << *StartingAccess << "\n")do { } while (false); | ||||||
| 2502 | LLVM_DEBUG(dbgs() << "Optimized Memory SSA clobber for " << *I << " is ")do { } while (false); | ||||||
| 2503 | LLVM_DEBUG(dbgs() << *OptimizedAccess << "\n")do { } while (false); | ||||||
| 2504 | |||||||
| 2505 | MemoryAccess *Result; | ||||||
| 2506 | if (SkipSelf && isa<MemoryPhi>(OptimizedAccess) && | ||||||
| 2507 | isa<MemoryDef>(StartingAccess) && UpwardWalkLimit) { | ||||||
| 2508 | assert(isa<MemoryDef>(Q.OriginalAccess))((void)0); | ||||||
| 2509 | Q.SkipSelfAccess = true; | ||||||
| 2510 | Result = Walker.findClobber(OptimizedAccess, Q, UpwardWalkLimit); | ||||||
| 2511 | } else | ||||||
| 2512 | Result = OptimizedAccess; | ||||||
| 2513 | |||||||
| 2514 | LLVM_DEBUG(dbgs() << "Result Memory SSA clobber [SkipSelf = " << SkipSelf)do { } while (false); | ||||||
| 2515 | LLVM_DEBUG(dbgs() << "] for " << *I << " is " << *Result << "\n")do { } while (false); | ||||||
| 2516 | |||||||
| 2517 | return Result; | ||||||
| 2518 | } | ||||||
| 2519 | |||||||
| 2520 | MemoryAccess * | ||||||
| 2521 | DoNothingMemorySSAWalker::getClobberingMemoryAccess(MemoryAccess *MA) { | ||||||
| 2522 | if (auto *Use = dyn_cast<MemoryUseOrDef>(MA)) | ||||||
| 2523 | return Use->getDefiningAccess(); | ||||||
| 2524 | return MA; | ||||||
| 2525 | } | ||||||
| 2526 | |||||||
| 2527 | MemoryAccess *DoNothingMemorySSAWalker::getClobberingMemoryAccess( | ||||||
| 2528 | MemoryAccess *StartingAccess, const MemoryLocation &) { | ||||||
| 2529 | if (auto *Use = dyn_cast<MemoryUseOrDef>(StartingAccess)) | ||||||
| 2530 | return Use->getDefiningAccess(); | ||||||
| 2531 | return StartingAccess; | ||||||
| 2532 | } | ||||||
| 2533 | |||||||
| 2534 | void MemoryPhi::deleteMe(DerivedUser *Self) { | ||||||
| 2535 | delete static_cast<MemoryPhi *>(Self); | ||||||
| 2536 | } | ||||||
| 2537 | |||||||
| 2538 | void MemoryDef::deleteMe(DerivedUser *Self) { | ||||||
| 2539 | delete static_cast<MemoryDef *>(Self); | ||||||
| 2540 | } | ||||||
| 2541 | |||||||
| 2542 | void MemoryUse::deleteMe(DerivedUser *Self) { | ||||||
| 2543 | delete static_cast<MemoryUse *>(Self); | ||||||
| 2544 | } | ||||||
| 2545 | |||||||
| 2546 | bool upward_defs_iterator::IsGuaranteedLoopInvariant(Value *Ptr) const { | ||||||
| 2547 | auto IsGuaranteedLoopInvariantBase = [](Value *Ptr) { | ||||||
| 2548 | Ptr = Ptr->stripPointerCasts(); | ||||||
| 2549 | if (!isa<Instruction>(Ptr)) | ||||||
| 2550 | return true; | ||||||
| 2551 | return isa<AllocaInst>(Ptr); | ||||||
| 2552 | }; | ||||||
| 2553 | |||||||
| 2554 | Ptr = Ptr->stripPointerCasts(); | ||||||
| 2555 | if (auto *I = dyn_cast<Instruction>(Ptr)) { | ||||||
| 2556 | if (I->getParent()->isEntryBlock()) | ||||||
| 2557 | return true; | ||||||
| 2558 | } | ||||||
| 2559 | if (auto *GEP = dyn_cast<GEPOperator>(Ptr)) { | ||||||
| 2560 | return IsGuaranteedLoopInvariantBase(GEP->getPointerOperand()) && | ||||||
| 2561 | GEP->hasAllConstantIndices(); | ||||||
| 2562 | } | ||||||
| 2563 | return IsGuaranteedLoopInvariantBase(Ptr); | ||||||
| 2564 | } |
| 1 | //===- Optional.h - Simple variant for passing optional values --*- C++ -*-===// |
| 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 provides Optional, a template class modeled in the spirit of |
| 10 | // OCaml's 'opt' variant. The idea is to strongly type whether or not |
| 11 | // a value can be optional. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | #ifndef LLVM_ADT_OPTIONAL_H |
| 16 | #define LLVM_ADT_OPTIONAL_H |
| 17 | |
| 18 | #include "llvm/ADT/Hashing.h" |
| 19 | #include "llvm/ADT/None.h" |
| 20 | #include "llvm/ADT/STLForwardCompat.h" |
| 21 | #include "llvm/Support/Compiler.h" |
| 22 | #include "llvm/Support/type_traits.h" |
| 23 | #include <cassert> |
| 24 | #include <memory> |
| 25 | #include <new> |
| 26 | #include <utility> |
| 27 | |
| 28 | namespace llvm { |
| 29 | |
| 30 | class raw_ostream; |
| 31 | |
| 32 | namespace optional_detail { |
| 33 | |
| 34 | /// Storage for any type. |
| 35 | // |
| 36 | // The specialization condition intentionally uses |
| 37 | // llvm::is_trivially_copy_constructible instead of |
| 38 | // std::is_trivially_copy_constructible. GCC versions prior to 7.4 may |
| 39 | // instantiate the copy constructor of `T` when |
| 40 | // std::is_trivially_copy_constructible is instantiated. This causes |
| 41 | // compilation to fail if we query the trivially copy constructible property of |
| 42 | // a class which is not copy constructible. |
| 43 | // |
| 44 | // The current implementation of OptionalStorage insists that in order to use |
| 45 | // the trivial specialization, the value_type must be trivially copy |
| 46 | // constructible and trivially copy assignable due to =default implementations |
| 47 | // of the copy/move constructor/assignment. It does not follow that this is |
| 48 | // necessarily the case std::is_trivially_copyable is true (hence the expanded |
| 49 | // specialization condition). |
| 50 | // |
| 51 | // The move constructible / assignable conditions emulate the remaining behavior |
| 52 | // of std::is_trivially_copyable. |
| 53 | template <typename T, bool = (llvm::is_trivially_copy_constructible<T>::value && |
| 54 | std::is_trivially_copy_assignable<T>::value && |
| 55 | (std::is_trivially_move_constructible<T>::value || |
| 56 | !std::is_move_constructible<T>::value) && |
| 57 | (std::is_trivially_move_assignable<T>::value || |
| 58 | !std::is_move_assignable<T>::value))> |
| 59 | class OptionalStorage { |
| 60 | union { |
| 61 | char empty; |
| 62 | T value; |
| 63 | }; |
| 64 | bool hasVal; |
| 65 | |
| 66 | public: |
| 67 | ~OptionalStorage() { reset(); } |
| 68 | |
| 69 | constexpr OptionalStorage() noexcept : empty(), hasVal(false) {} |
| 70 | |
| 71 | constexpr OptionalStorage(OptionalStorage const &other) : OptionalStorage() { |
| 72 | if (other.hasValue()) { |
| 73 | emplace(other.value); |
| 74 | } |
| 75 | } |
| 76 | constexpr OptionalStorage(OptionalStorage &&other) : OptionalStorage() { |
| 77 | if (other.hasValue()) { |
| 78 | emplace(std::move(other.value)); |
| 79 | } |
| 80 | } |
| 81 | |
| 82 | template <class... Args> |
| 83 | constexpr explicit OptionalStorage(in_place_t, Args &&... args) |
| 84 | : value(std::forward<Args>(args)...), hasVal(true) {} |
| 85 | |
| 86 | void reset() noexcept { |
| 87 | if (hasVal) { |
| 88 | value.~T(); |
| 89 | hasVal = false; |
| 90 | } |
| 91 | } |
| 92 | |
| 93 | constexpr bool hasValue() const noexcept { return hasVal; } |
| 94 | |
| 95 | T &getValue() LLVM_LVALUE_FUNCTION& noexcept { |
| 96 | assert(hasVal)((void)0); |
| 97 | return value; |
| 98 | } |
| 99 | constexpr T const &getValue() const LLVM_LVALUE_FUNCTION& noexcept { |
| 100 | assert(hasVal)((void)0); |
| 101 | return value; |
| 102 | } |
| 103 | #if LLVM_HAS_RVALUE_REFERENCE_THIS1 |
| 104 | T &&getValue() && noexcept { |
| 105 | assert(hasVal)((void)0); |
| 106 | return std::move(value); |
| 107 | } |
| 108 | #endif |
| 109 | |
| 110 | template <class... Args> void emplace(Args &&... args) { |
| 111 | reset(); |
| 112 | ::new ((void *)std::addressof(value)) T(std::forward<Args>(args)...); |
| 113 | hasVal = true; |
| 114 | } |
| 115 | |
| 116 | OptionalStorage &operator=(T const &y) { |
| 117 | if (hasValue()) { |
| 118 | value = y; |
| 119 | } else { |
| 120 | ::new ((void *)std::addressof(value)) T(y); |
| 121 | hasVal = true; |
| 122 | } |
| 123 | return *this; |
| 124 | } |
| 125 | OptionalStorage &operator=(T &&y) { |
| 126 | if (hasValue()) { |
| 127 | value = std::move(y); |
| 128 | } else { |
| 129 | ::new ((void *)std::addressof(value)) T(std::move(y)); |
| 130 | hasVal = true; |
| 131 | } |
| 132 | return *this; |
| 133 | } |
| 134 | |
| 135 | OptionalStorage &operator=(OptionalStorage const &other) { |
| 136 | if (other.hasValue()) { |
| 137 | if (hasValue()) { |
| 138 | value = other.value; |
| 139 | } else { |
| 140 | ::new ((void *)std::addressof(value)) T(other.value); |
| 141 | hasVal = true; |
| 142 | } |
| 143 | } else { |
| 144 | reset(); |
| 145 | } |
| 146 | return *this; |
| 147 | } |
| 148 | |
| 149 | OptionalStorage &operator=(OptionalStorage &&other) { |
| 150 | if (other.hasValue()) { |
| 151 | if (hasValue()) { |
| 152 | value = std::move(other.value); |
| 153 | } else { |
| 154 | ::new ((void *)std::addressof(value)) T(std::move(other.value)); |
| 155 | hasVal = true; |
| 156 | } |
| 157 | } else { |
| 158 | reset(); |
| 159 | } |
| 160 | return *this; |
| 161 | } |
| 162 | }; |
| 163 | |
| 164 | template <typename T> class OptionalStorage<T, true> { |
| 165 | union { |
| 166 | char empty; |
| 167 | T value; |
| 168 | }; |
| 169 | bool hasVal = false; |
| 170 | |
| 171 | public: |
| 172 | ~OptionalStorage() = default; |
| 173 | |
| 174 | constexpr OptionalStorage() noexcept : empty{} {} |
| 175 | |
| 176 | constexpr OptionalStorage(OptionalStorage const &other) = default; |
| 177 | constexpr OptionalStorage(OptionalStorage &&other) = default; |
| 178 | |
| 179 | OptionalStorage &operator=(OptionalStorage const &other) = default; |
| 180 | OptionalStorage &operator=(OptionalStorage &&other) = default; |
| 181 | |
| 182 | template <class... Args> |
| 183 | constexpr explicit OptionalStorage(in_place_t, Args &&... args) |
| 184 | : value(std::forward<Args>(args)...), hasVal(true) {} |
| 185 | |
| 186 | void reset() noexcept { |
| 187 | if (hasVal) { |
| 188 | value.~T(); |
| 189 | hasVal = false; |
| 190 | } |
| 191 | } |
| 192 | |
| 193 | constexpr bool hasValue() const noexcept { return hasVal; } |
| 194 | |
| 195 | T &getValue() LLVM_LVALUE_FUNCTION& noexcept { |
| 196 | assert(hasVal)((void)0); |
| 197 | return value; |
| 198 | } |
| 199 | constexpr T const &getValue() const LLVM_LVALUE_FUNCTION& noexcept { |
| 200 | assert(hasVal)((void)0); |
| 201 | return value; |
| 202 | } |
| 203 | #if LLVM_HAS_RVALUE_REFERENCE_THIS1 |
| 204 | T &&getValue() && noexcept { |
| 205 | assert(hasVal)((void)0); |
| 206 | return std::move(value); |
| 207 | } |
| 208 | #endif |
| 209 | |
| 210 | template <class... Args> void emplace(Args &&... args) { |
| 211 | reset(); |
| 212 | ::new ((void *)std::addressof(value)) T(std::forward<Args>(args)...); |
| 213 | hasVal = true; |
| 214 | } |
| 215 | |
| 216 | OptionalStorage &operator=(T const &y) { |
| 217 | if (hasValue()) { |
| 218 | value = y; |
| 219 | } else { |
| 220 | ::new ((void *)std::addressof(value)) T(y); |
| 221 | hasVal = true; |
| 222 | } |
| 223 | return *this; |
| 224 | } |
| 225 | OptionalStorage &operator=(T &&y) { |
| 226 | if (hasValue()) { |
| 227 | value = std::move(y); |
| 228 | } else { |
| 229 | ::new ((void *)std::addressof(value)) T(std::move(y)); |
| 230 | hasVal = true; |
| 231 | } |
| 232 | return *this; |
| 233 | } |
| 234 | }; |
| 235 | |
| 236 | } // namespace optional_detail |
| 237 | |
| 238 | template <typename T> class Optional { |
| 239 | optional_detail::OptionalStorage<T> Storage; |
| 240 | |
| 241 | public: |
| 242 | using value_type = T; |
| 243 | |
| 244 | constexpr Optional() {} |
| 245 | constexpr Optional(NoneType) {} |
| 246 | |
| 247 | constexpr Optional(const T &y) : Storage(in_place, y) {} |
| 248 | constexpr Optional(const Optional &O) = default; |
| 249 | |
| 250 | constexpr Optional(T &&y) : Storage(in_place, std::move(y)) {} |
| 251 | constexpr Optional(Optional &&O) = default; |
| 252 | |
| 253 | template <typename... ArgTypes> |
| 254 | constexpr Optional(in_place_t, ArgTypes &&...Args) |
| 255 | : Storage(in_place, std::forward<ArgTypes>(Args)...) {} |
| 256 | |
| 257 | Optional &operator=(T &&y) { |
| 258 | Storage = std::move(y); |
| 259 | return *this; |
| 260 | } |
| 261 | Optional &operator=(Optional &&O) = default; |
| 262 | |
| 263 | /// Create a new object by constructing it in place with the given arguments. |
| 264 | template <typename... ArgTypes> void emplace(ArgTypes &&... Args) { |
| 265 | Storage.emplace(std::forward<ArgTypes>(Args)...); |
| 266 | } |
| 267 | |
| 268 | static constexpr Optional create(const T *y) { |
| 269 | return y ? Optional(*y) : Optional(); |
| 270 | } |
| 271 | |
| 272 | Optional &operator=(const T &y) { |
| 273 | Storage = y; |
| 274 | return *this; |
| 275 | } |
| 276 | Optional &operator=(const Optional &O) = default; |
| 277 | |
| 278 | void reset() { Storage.reset(); } |
| 279 | |
| 280 | constexpr const T *getPointer() const { return &Storage.getValue(); } |
| 281 | T *getPointer() { return &Storage.getValue(); } |
| 282 | constexpr const T &getValue() const LLVM_LVALUE_FUNCTION& { |
| 283 | return Storage.getValue(); |
| 284 | } |
| 285 | T &getValue() LLVM_LVALUE_FUNCTION& { return Storage.getValue(); } |
| 286 | |
| 287 | constexpr explicit operator bool() const { return hasValue(); } |
| 288 | constexpr bool hasValue() const { return Storage.hasValue(); } |
| 289 | constexpr const T *operator->() const { return getPointer(); } |
| 290 | T *operator->() { return getPointer(); } |
| 291 | constexpr const T &operator*() const LLVM_LVALUE_FUNCTION& { |
| 292 | return getValue(); |
| 293 | } |
| 294 | T &operator*() LLVM_LVALUE_FUNCTION& { return getValue(); } |
| 295 | |
| 296 | template <typename U> |
| 297 | constexpr T getValueOr(U &&value) const LLVM_LVALUE_FUNCTION& { |
| 298 | return hasValue() ? getValue() : std::forward<U>(value); |
| 299 | } |
| 300 | |
| 301 | /// Apply a function to the value if present; otherwise return None. |
| 302 | template <class Function> |
| 303 | auto map(const Function &F) const LLVM_LVALUE_FUNCTION& |
| 304 | -> Optional<decltype(F(getValue()))> { |
| 305 | if (*this) return F(getValue()); |
| 306 | return None; |
| 307 | } |
| 308 | |
| 309 | #if LLVM_HAS_RVALUE_REFERENCE_THIS1 |
| 310 | T &&getValue() && { return std::move(Storage.getValue()); } |
| 311 | T &&operator*() && { return std::move(Storage.getValue()); } |
| 312 | |
| 313 | template <typename U> |
| 314 | T getValueOr(U &&value) && { |
| 315 | return hasValue() ? std::move(getValue()) : std::forward<U>(value); |
| 316 | } |
| 317 | |
| 318 | /// Apply a function to the value if present; otherwise return None. |
| 319 | template <class Function> |
| 320 | auto map(const Function &F) && |
| 321 | -> Optional<decltype(F(std::move(*this).getValue()))> { |
| 322 | if (*this) return F(std::move(*this).getValue()); |
| 323 | return None; |
| 324 | } |
| 325 | #endif |
| 326 | }; |
| 327 | |
| 328 | template <class T> llvm::hash_code hash_value(const Optional<T> &O) { |
| 329 | return O ? hash_combine(true, *O) : hash_value(false); |
| 330 | } |
| 331 | |
| 332 | template <typename T, typename U> |
| 333 | constexpr bool operator==(const Optional<T> &X, const Optional<U> &Y) { |
| 334 | if (X && Y) |
| 335 | return *X == *Y; |
| 336 | return X.hasValue() == Y.hasValue(); |
| 337 | } |
| 338 | |
| 339 | template <typename T, typename U> |
| 340 | constexpr bool operator!=(const Optional<T> &X, const Optional<U> &Y) { |
| 341 | return !(X == Y); |
| 342 | } |
| 343 | |
| 344 | template <typename T, typename U> |
| 345 | constexpr bool operator<(const Optional<T> &X, const Optional<U> &Y) { |
| 346 | if (X && Y) |
| 347 | return *X < *Y; |
| 348 | return X.hasValue() < Y.hasValue(); |
| 349 | } |
| 350 | |
| 351 | template <typename T, typename U> |
| 352 | constexpr bool operator<=(const Optional<T> &X, const Optional<U> &Y) { |
| 353 | return !(Y < X); |
| 354 | } |
| 355 | |
| 356 | template <typename T, typename U> |
| 357 | constexpr bool operator>(const Optional<T> &X, const Optional<U> &Y) { |
| 358 | return Y < X; |
| 359 | } |
| 360 | |
| 361 | template <typename T, typename U> |
| 362 | constexpr bool operator>=(const Optional<T> &X, const Optional<U> &Y) { |
| 363 | return !(X < Y); |
| 364 | } |
| 365 | |
| 366 | template <typename T> |
| 367 | constexpr bool operator==(const Optional<T> &X, NoneType) { |
| 368 | return !X; |
| 369 | } |
| 370 | |
| 371 | template <typename T> |
| 372 | constexpr bool operator==(NoneType, const Optional<T> &X) { |
| 373 | return X == None; |
| 374 | } |
| 375 | |
| 376 | template <typename T> |
| 377 | constexpr bool operator!=(const Optional<T> &X, NoneType) { |
| 378 | return !(X == None); |
| 379 | } |
| 380 | |
| 381 | template <typename T> |
| 382 | constexpr bool operator!=(NoneType, const Optional<T> &X) { |
| 383 | return X != None; |
| 384 | } |
| 385 | |
| 386 | template <typename T> constexpr bool operator<(const Optional<T> &, NoneType) { |
| 387 | return false; |
| 388 | } |
| 389 | |
| 390 | template <typename T> constexpr bool operator<(NoneType, const Optional<T> &X) { |
| 391 | return X.hasValue(); |
| 392 | } |
| 393 | |
| 394 | template <typename T> |
| 395 | constexpr bool operator<=(const Optional<T> &X, NoneType) { |
| 396 | return !(None < X); |
| 397 | } |
| 398 | |
| 399 | template <typename T> |
| 400 | constexpr bool operator<=(NoneType, const Optional<T> &X) { |
| 401 | return !(X < None); |
| 402 | } |
| 403 | |
| 404 | template <typename T> constexpr bool operator>(const Optional<T> &X, NoneType) { |
| 405 | return None < X; |
| 406 | } |
| 407 | |
| 408 | template <typename T> constexpr bool operator>(NoneType, const Optional<T> &X) { |
| 409 | return X < None; |
| 410 | } |
| 411 | |
| 412 | template <typename T> |
| 413 | constexpr bool operator>=(const Optional<T> &X, NoneType) { |
| 414 | return None <= X; |
| 415 | } |
| 416 | |
| 417 | template <typename T> |
| 418 | constexpr bool operator>=(NoneType, const Optional<T> &X) { |
| 419 | return X <= None; |
| 420 | } |
| 421 | |
| 422 | template <typename T> |
| 423 | constexpr bool operator==(const Optional<T> &X, const T &Y) { |
| 424 | return X && *X == Y; |
| 425 | } |
| 426 | |
| 427 | template <typename T> |
| 428 | constexpr bool operator==(const T &X, const Optional<T> &Y) { |
| 429 | return Y && X == *Y; |
| 430 | } |
| 431 | |
| 432 | template <typename T> |
| 433 | constexpr bool operator!=(const Optional<T> &X, const T &Y) { |
| 434 | return !(X == Y); |
| 435 | } |
| 436 | |
| 437 | template <typename T> |
| 438 | constexpr bool operator!=(const T &X, const Optional<T> &Y) { |
| 439 | return !(X == Y); |
| 440 | } |
| 441 | |
| 442 | template <typename T> |
| 443 | constexpr bool operator<(const Optional<T> &X, const T &Y) { |
| 444 | return !X || *X < Y; |
| 445 | } |
| 446 | |
| 447 | template <typename T> |
| 448 | constexpr bool operator<(const T &X, const Optional<T> &Y) { |
| 449 | return Y && X < *Y; |
| 450 | } |
| 451 | |
| 452 | template <typename T> |
| 453 | constexpr bool operator<=(const Optional<T> &X, const T &Y) { |
| 454 | return !(Y < X); |
| 455 | } |
| 456 | |
| 457 | template <typename T> |
| 458 | constexpr bool operator<=(const T &X, const Optional<T> &Y) { |
| 459 | return !(Y < X); |
| 460 | } |
| 461 | |
| 462 | template <typename T> |
| 463 | constexpr bool operator>(const Optional<T> &X, const T &Y) { |
| 464 | return Y < X; |
| 465 | } |
| 466 | |
| 467 | template <typename T> |
| 468 | constexpr bool operator>(const T &X, const Optional<T> &Y) { |
| 469 | return Y < X; |
| 470 | } |
| 471 | |
| 472 | template <typename T> |
| 473 | constexpr bool operator>=(const Optional<T> &X, const T &Y) { |
| 474 | return !(X < Y); |
| 475 | } |
| 476 | |
| 477 | template <typename T> |
| 478 | constexpr bool operator>=(const T &X, const Optional<T> &Y) { |
| 479 | return !(X < Y); |
| 480 | } |
| 481 | |
| 482 | raw_ostream &operator<<(raw_ostream &OS, NoneType); |
| 483 | |
| 484 | template <typename T, typename = decltype(std::declval<raw_ostream &>() |
| 485 | << std::declval<const T &>())> |
| 486 | raw_ostream &operator<<(raw_ostream &OS, const Optional<T> &O) { |
| 487 | if (O) |
| 488 | OS << *O; |
| 489 | else |
| 490 | OS << None; |
| 491 | return OS; |
| 492 | } |
| 493 | |
| 494 | } // end namespace llvm |
| 495 | |
| 496 | #endif // LLVM_ADT_OPTIONAL_H |
| 1 | //===- llvm/ADT/SmallVector.h - 'Normally small' vectors --------*- C++ -*-===// |
| 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 defines the SmallVector class. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #ifndef LLVM_ADT_SMALLVECTOR_H |
| 14 | #define LLVM_ADT_SMALLVECTOR_H |
| 15 | |
| 16 | #include "llvm/ADT/iterator_range.h" |
| 17 | #include "llvm/Support/Compiler.h" |
| 18 | #include "llvm/Support/ErrorHandling.h" |
| 19 | #include "llvm/Support/MemAlloc.h" |
| 20 | #include "llvm/Support/type_traits.h" |
| 21 | #include <algorithm> |
| 22 | #include <cassert> |
| 23 | #include <cstddef> |
| 24 | #include <cstdlib> |
| 25 | #include <cstring> |
| 26 | #include <functional> |
| 27 | #include <initializer_list> |
| 28 | #include <iterator> |
| 29 | #include <limits> |
| 30 | #include <memory> |
| 31 | #include <new> |
| 32 | #include <type_traits> |
| 33 | #include <utility> |
| 34 | |
| 35 | namespace llvm { |
| 36 | |
| 37 | /// This is all the stuff common to all SmallVectors. |
| 38 | /// |
| 39 | /// The template parameter specifies the type which should be used to hold the |
| 40 | /// Size and Capacity of the SmallVector, so it can be adjusted. |
| 41 | /// Using 32 bit size is desirable to shrink the size of the SmallVector. |
| 42 | /// Using 64 bit size is desirable for cases like SmallVector<char>, where a |
| 43 | /// 32 bit size would limit the vector to ~4GB. SmallVectors are used for |
| 44 | /// buffering bitcode output - which can exceed 4GB. |
| 45 | template <class Size_T> class SmallVectorBase { |
| 46 | protected: |
| 47 | void *BeginX; |
| 48 | Size_T Size = 0, Capacity; |
| 49 | |
| 50 | /// The maximum value of the Size_T used. |
| 51 | static constexpr size_t SizeTypeMax() { |
| 52 | return std::numeric_limits<Size_T>::max(); |
| 53 | } |
| 54 | |
| 55 | SmallVectorBase() = delete; |
| 56 | SmallVectorBase(void *FirstEl, size_t TotalCapacity) |
| 57 | : BeginX(FirstEl), Capacity(TotalCapacity) {} |
| 58 | |
| 59 | /// This is a helper for \a grow() that's out of line to reduce code |
| 60 | /// duplication. This function will report a fatal error if it can't grow at |
| 61 | /// least to \p MinSize. |
| 62 | void *mallocForGrow(size_t MinSize, size_t TSize, size_t &NewCapacity); |
| 63 | |
| 64 | /// This is an implementation of the grow() method which only works |
| 65 | /// on POD-like data types and is out of line to reduce code duplication. |
| 66 | /// This function will report a fatal error if it cannot increase capacity. |
| 67 | void grow_pod(void *FirstEl, size_t MinSize, size_t TSize); |
| 68 | |
| 69 | public: |
| 70 | size_t size() const { return Size; } |
| 71 | size_t capacity() const { return Capacity; } |
| 72 | |
| 73 | LLVM_NODISCARD[[clang::warn_unused_result]] bool empty() const { return !Size; } |
| 74 | |
| 75 | /// Set the array size to \p N, which the current array must have enough |
| 76 | /// capacity for. |
| 77 | /// |
| 78 | /// This does not construct or destroy any elements in the vector. |
| 79 | /// |
| 80 | /// Clients can use this in conjunction with capacity() to write past the end |
| 81 | /// of the buffer when they know that more elements are available, and only |
| 82 | /// update the size later. This avoids the cost of value initializing elements |
| 83 | /// which will only be overwritten. |
| 84 | void set_size(size_t N) { |
| 85 | assert(N <= capacity())((void)0); |
| 86 | Size = N; |
| 87 | } |
| 88 | }; |
| 89 | |
| 90 | template <class T> |
| 91 | using SmallVectorSizeType = |
| 92 | typename std::conditional<sizeof(T) < 4 && sizeof(void *) >= 8, uint64_t, |
| 93 | uint32_t>::type; |
| 94 | |
| 95 | /// Figure out the offset of the first element. |
| 96 | template <class T, typename = void> struct SmallVectorAlignmentAndSize { |
| 97 | alignas(SmallVectorBase<SmallVectorSizeType<T>>) char Base[sizeof( |
| 98 | SmallVectorBase<SmallVectorSizeType<T>>)]; |
| 99 | alignas(T) char FirstEl[sizeof(T)]; |
| 100 | }; |
| 101 | |
| 102 | /// This is the part of SmallVectorTemplateBase which does not depend on whether |
| 103 | /// the type T is a POD. The extra dummy template argument is used by ArrayRef |
| 104 | /// to avoid unnecessarily requiring T to be complete. |
| 105 | template <typename T, typename = void> |
| 106 | class SmallVectorTemplateCommon |
| 107 | : public SmallVectorBase<SmallVectorSizeType<T>> { |
| 108 | using Base = SmallVectorBase<SmallVectorSizeType<T>>; |
| 109 | |
| 110 | /// Find the address of the first element. For this pointer math to be valid |
| 111 | /// with small-size of 0 for T with lots of alignment, it's important that |
| 112 | /// SmallVectorStorage is properly-aligned even for small-size of 0. |
| 113 | void *getFirstEl() const { |
| 114 | return const_cast<void *>(reinterpret_cast<const void *>( |
| 115 | reinterpret_cast<const char *>(this) + |
| 116 | offsetof(SmallVectorAlignmentAndSize<T>, FirstEl)__builtin_offsetof(SmallVectorAlignmentAndSize<T>, FirstEl ))); |
| 117 | } |
| 118 | // Space after 'FirstEl' is clobbered, do not add any instance vars after it. |
| 119 | |
| 120 | protected: |
| 121 | SmallVectorTemplateCommon(size_t Size) : Base(getFirstEl(), Size) {} |
| 122 | |
| 123 | void grow_pod(size_t MinSize, size_t TSize) { |
| 124 | Base::grow_pod(getFirstEl(), MinSize, TSize); |
| 125 | } |
| 126 | |
| 127 | /// Return true if this is a smallvector which has not had dynamic |
| 128 | /// memory allocated for it. |
| 129 | bool isSmall() const { return this->BeginX == getFirstEl(); } |
| 130 | |
| 131 | /// Put this vector in a state of being small. |
| 132 | void resetToSmall() { |
| 133 | this->BeginX = getFirstEl(); |
| 134 | this->Size = this->Capacity = 0; // FIXME: Setting Capacity to 0 is suspect. |
| 135 | } |
| 136 | |
| 137 | /// Return true if V is an internal reference to the given range. |
| 138 | bool isReferenceToRange(const void *V, const void *First, const void *Last) const { |
| 139 | // Use std::less to avoid UB. |
| 140 | std::less<> LessThan; |
| 141 | return !LessThan(V, First) && LessThan(V, Last); |
| 142 | } |
| 143 | |
| 144 | /// Return true if V is an internal reference to this vector. |
| 145 | bool isReferenceToStorage(const void *V) const { |
| 146 | return isReferenceToRange(V, this->begin(), this->end()); |
| 147 | } |
| 148 | |
| 149 | /// Return true if First and Last form a valid (possibly empty) range in this |
| 150 | /// vector's storage. |
| 151 | bool isRangeInStorage(const void *First, const void *Last) const { |
| 152 | // Use std::less to avoid UB. |
| 153 | std::less<> LessThan; |
| 154 | return !LessThan(First, this->begin()) && !LessThan(Last, First) && |
| 155 | !LessThan(this->end(), Last); |
| 156 | } |
| 157 | |
| 158 | /// Return true unless Elt will be invalidated by resizing the vector to |
| 159 | /// NewSize. |
| 160 | bool isSafeToReferenceAfterResize(const void *Elt, size_t NewSize) { |
| 161 | // Past the end. |
| 162 | if (LLVM_LIKELY(!isReferenceToStorage(Elt))__builtin_expect((bool)(!isReferenceToStorage(Elt)), true)) |
| 163 | return true; |
| 164 | |
| 165 | // Return false if Elt will be destroyed by shrinking. |
| 166 | if (NewSize <= this->size()) |
| 167 | return Elt < this->begin() + NewSize; |
| 168 | |
| 169 | // Return false if we need to grow. |
| 170 | return NewSize <= this->capacity(); |
| 171 | } |
| 172 | |
| 173 | /// Check whether Elt will be invalidated by resizing the vector to NewSize. |
| 174 | void assertSafeToReferenceAfterResize(const void *Elt, size_t NewSize) { |
| 175 | assert(isSafeToReferenceAfterResize(Elt, NewSize) &&((void)0) |
| 176 | "Attempting to reference an element of the vector in an operation "((void)0) |
| 177 | "that invalidates it")((void)0); |
| 178 | } |
| 179 | |
| 180 | /// Check whether Elt will be invalidated by increasing the size of the |
| 181 | /// vector by N. |
| 182 | void assertSafeToAdd(const void *Elt, size_t N = 1) { |
| 183 | this->assertSafeToReferenceAfterResize(Elt, this->size() + N); |
| 184 | } |
| 185 | |
| 186 | /// Check whether any part of the range will be invalidated by clearing. |
| 187 | void assertSafeToReferenceAfterClear(const T *From, const T *To) { |
| 188 | if (From == To) |
| 189 | return; |
| 190 | this->assertSafeToReferenceAfterResize(From, 0); |
| 191 | this->assertSafeToReferenceAfterResize(To - 1, 0); |
| 192 | } |
| 193 | template < |
| 194 | class ItTy, |
| 195 | std::enable_if_t<!std::is_same<std::remove_const_t<ItTy>, T *>::value, |
| 196 | bool> = false> |
| 197 | void assertSafeToReferenceAfterClear(ItTy, ItTy) {} |
| 198 | |
| 199 | /// Check whether any part of the range will be invalidated by growing. |
| 200 | void assertSafeToAddRange(const T *From, const T *To) { |
| 201 | if (From == To) |
| 202 | return; |
| 203 | this->assertSafeToAdd(From, To - From); |
| 204 | this->assertSafeToAdd(To - 1, To - From); |
| 205 | } |
| 206 | template < |
| 207 | class ItTy, |
| 208 | std::enable_if_t<!std::is_same<std::remove_const_t<ItTy>, T *>::value, |
| 209 | bool> = false> |
| 210 | void assertSafeToAddRange(ItTy, ItTy) {} |
| 211 | |
| 212 | /// Reserve enough space to add one element, and return the updated element |
| 213 | /// pointer in case it was a reference to the storage. |
| 214 | template <class U> |
| 215 | static const T *reserveForParamAndGetAddressImpl(U *This, const T &Elt, |
| 216 | size_t N) { |
| 217 | size_t NewSize = This->size() + N; |
| 218 | if (LLVM_LIKELY(NewSize <= This->capacity())__builtin_expect((bool)(NewSize <= This->capacity()), true )) |
| 219 | return &Elt; |
| 220 | |
| 221 | bool ReferencesStorage = false; |
| 222 | int64_t Index = -1; |
| 223 | if (!U::TakesParamByValue) { |
| 224 | if (LLVM_UNLIKELY(This->isReferenceToStorage(&Elt))__builtin_expect((bool)(This->isReferenceToStorage(&Elt )), false)) { |
| 225 | ReferencesStorage = true; |
| 226 | Index = &Elt - This->begin(); |
| 227 | } |
| 228 | } |
| 229 | This->grow(NewSize); |
| 230 | return ReferencesStorage ? This->begin() + Index : &Elt; |
| 231 | } |
| 232 | |
| 233 | public: |
| 234 | using size_type = size_t; |
| 235 | using difference_type = ptrdiff_t; |
| 236 | using value_type = T; |
| 237 | using iterator = T *; |
| 238 | using const_iterator = const T *; |
| 239 | |
| 240 | using const_reverse_iterator = std::reverse_iterator<const_iterator>; |
| 241 | using reverse_iterator = std::reverse_iterator<iterator>; |
| 242 | |
| 243 | using reference = T &; |
| 244 | using const_reference = const T &; |
| 245 | using pointer = T *; |
| 246 | using const_pointer = const T *; |
| 247 | |
| 248 | using Base::capacity; |
| 249 | using Base::empty; |
| 250 | using Base::size; |
| 251 | |
| 252 | // forward iterator creation methods. |
| 253 | iterator begin() { return (iterator)this->BeginX; } |
| 254 | const_iterator begin() const { return (const_iterator)this->BeginX; } |
| 255 | iterator end() { return begin() + size(); } |
| 256 | const_iterator end() const { return begin() + size(); } |
| 257 | |
| 258 | // reverse iterator creation methods. |
| 259 | reverse_iterator rbegin() { return reverse_iterator(end()); } |
| 260 | const_reverse_iterator rbegin() const{ return const_reverse_iterator(end()); } |
| 261 | reverse_iterator rend() { return reverse_iterator(begin()); } |
| 262 | const_reverse_iterator rend() const { return const_reverse_iterator(begin());} |
| 263 | |
| 264 | size_type size_in_bytes() const { return size() * sizeof(T); } |
| 265 | size_type max_size() const { |
| 266 | return std::min(this->SizeTypeMax(), size_type(-1) / sizeof(T)); |
| 267 | } |
| 268 | |
| 269 | size_t capacity_in_bytes() const { return capacity() * sizeof(T); } |
| 270 | |
| 271 | /// Return a pointer to the vector's buffer, even if empty(). |
| 272 | pointer data() { return pointer(begin()); } |
| 273 | /// Return a pointer to the vector's buffer, even if empty(). |
| 274 | const_pointer data() const { return const_pointer(begin()); } |
| 275 | |
| 276 | reference operator[](size_type idx) { |
| 277 | assert(idx < size())((void)0); |
| 278 | return begin()[idx]; |
| 279 | } |
| 280 | const_reference operator[](size_type idx) const { |
| 281 | assert(idx < size())((void)0); |
| 282 | return begin()[idx]; |
| 283 | } |
| 284 | |
| 285 | reference front() { |
| 286 | assert(!empty())((void)0); |
| 287 | return begin()[0]; |
| 288 | } |
| 289 | const_reference front() const { |
| 290 | assert(!empty())((void)0); |
| 291 | return begin()[0]; |
| 292 | } |
| 293 | |
| 294 | reference back() { |
| 295 | assert(!empty())((void)0); |
| 296 | return end()[-1]; |
| 297 | } |
| 298 | const_reference back() const { |
| 299 | assert(!empty())((void)0); |
| 300 | return end()[-1]; |
| 301 | } |
| 302 | }; |
| 303 | |
| 304 | /// SmallVectorTemplateBase<TriviallyCopyable = false> - This is where we put |
| 305 | /// method implementations that are designed to work with non-trivial T's. |
| 306 | /// |
| 307 | /// We approximate is_trivially_copyable with trivial move/copy construction and |
| 308 | /// trivial destruction. While the standard doesn't specify that you're allowed |
| 309 | /// copy these types with memcpy, there is no way for the type to observe this. |
| 310 | /// This catches the important case of std::pair<POD, POD>, which is not |
| 311 | /// trivially assignable. |
| 312 | template <typename T, bool = (is_trivially_copy_constructible<T>::value) && |
| 313 | (is_trivially_move_constructible<T>::value) && |
| 314 | std::is_trivially_destructible<T>::value> |
| 315 | class SmallVectorTemplateBase : public SmallVectorTemplateCommon<T> { |
| 316 | friend class SmallVectorTemplateCommon<T>; |
| 317 | |
| 318 | protected: |
| 319 | static constexpr bool TakesParamByValue = false; |
| 320 | using ValueParamT = const T &; |
| 321 | |
| 322 | SmallVectorTemplateBase(size_t Size) : SmallVectorTemplateCommon<T>(Size) {} |
| 323 | |
| 324 | static void destroy_range(T *S, T *E) { |
| 325 | while (S != E) { |
| 326 | --E; |
| 327 | E->~T(); |
| 328 | } |
| 329 | } |
| 330 | |
| 331 | /// Move the range [I, E) into the uninitialized memory starting with "Dest", |
| 332 | /// constructing elements as needed. |
| 333 | template<typename It1, typename It2> |
| 334 | static void uninitialized_move(It1 I, It1 E, It2 Dest) { |
| 335 | std::uninitialized_copy(std::make_move_iterator(I), |
| 336 | std::make_move_iterator(E), Dest); |
| 337 | } |
| 338 | |
| 339 | /// Copy the range [I, E) onto the uninitialized memory starting with "Dest", |
| 340 | /// constructing elements as needed. |
| 341 | template<typename It1, typename It2> |
| 342 | static void uninitialized_copy(It1 I, It1 E, It2 Dest) { |
| 343 | std::uninitialized_copy(I, E, Dest); |
| 344 | } |
| 345 | |
| 346 | /// Grow the allocated memory (without initializing new elements), doubling |
| 347 | /// the size of the allocated memory. Guarantees space for at least one more |
| 348 | /// element, or MinSize more elements if specified. |
| 349 | void grow(size_t MinSize = 0); |
| 350 | |
| 351 | /// Create a new allocation big enough for \p MinSize and pass back its size |
| 352 | /// in \p NewCapacity. This is the first section of \a grow(). |
| 353 | T *mallocForGrow(size_t MinSize, size_t &NewCapacity) { |
| 354 | return static_cast<T *>( |
| 355 | SmallVectorBase<SmallVectorSizeType<T>>::mallocForGrow( |
| 356 | MinSize, sizeof(T), NewCapacity)); |
| 357 | } |
| 358 | |
| 359 | /// Move existing elements over to the new allocation \p NewElts, the middle |
| 360 | /// section of \a grow(). |
| 361 | void moveElementsForGrow(T *NewElts); |
| 362 | |
| 363 | /// Transfer ownership of the allocation, finishing up \a grow(). |
| 364 | void takeAllocationForGrow(T *NewElts, size_t NewCapacity); |
| 365 | |
| 366 | /// Reserve enough space to add one element, and return the updated element |
| 367 | /// pointer in case it was a reference to the storage. |
| 368 | const T *reserveForParamAndGetAddress(const T &Elt, size_t N = 1) { |
| 369 | return this->reserveForParamAndGetAddressImpl(this, Elt, N); |
| 370 | } |
| 371 | |
| 372 | /// Reserve enough space to add one element, and return the updated element |
| 373 | /// pointer in case it was a reference to the storage. |
| 374 | T *reserveForParamAndGetAddress(T &Elt, size_t N = 1) { |
| 375 | return const_cast<T *>( |
| 376 | this->reserveForParamAndGetAddressImpl(this, Elt, N)); |
| 377 | } |
| 378 | |
| 379 | static T &&forward_value_param(T &&V) { return std::move(V); } |
| 380 | static const T &forward_value_param(const T &V) { return V; } |
| 381 | |
| 382 | void growAndAssign(size_t NumElts, const T &Elt) { |
| 383 | // Grow manually in case Elt is an internal reference. |
| 384 | size_t NewCapacity; |
| 385 | T *NewElts = mallocForGrow(NumElts, NewCapacity); |
| 386 | std::uninitialized_fill_n(NewElts, NumElts, Elt); |
| 387 | this->destroy_range(this->begin(), this->end()); |
| 388 | takeAllocationForGrow(NewElts, NewCapacity); |
| 389 | this->set_size(NumElts); |
| 390 | } |
| 391 | |
| 392 | template <typename... ArgTypes> T &growAndEmplaceBack(ArgTypes &&... Args) { |
| 393 | // Grow manually in case one of Args is an internal reference. |
| 394 | size_t NewCapacity; |
| 395 | T *NewElts = mallocForGrow(0, NewCapacity); |
| 396 | ::new ((void *)(NewElts + this->size())) T(std::forward<ArgTypes>(Args)...); |
| 397 | moveElementsForGrow(NewElts); |
| 398 | takeAllocationForGrow(NewElts, NewCapacity); |
| 399 | this->set_size(this->size() + 1); |
| 400 | return this->back(); |
| 401 | } |
| 402 | |
| 403 | public: |
| 404 | void push_back(const T &Elt) { |
| 405 | const T *EltPtr = reserveForParamAndGetAddress(Elt); |
| 406 | ::new ((void *)this->end()) T(*EltPtr); |
| 407 | this->set_size(this->size() + 1); |
| 408 | } |
| 409 | |
| 410 | void push_back(T &&Elt) { |
| 411 | T *EltPtr = reserveForParamAndGetAddress(Elt); |
| 412 | ::new ((void *)this->end()) T(::std::move(*EltPtr)); |
| 413 | this->set_size(this->size() + 1); |
| 414 | } |
| 415 | |
| 416 | void pop_back() { |
| 417 | this->set_size(this->size() - 1); |
| 418 | this->end()->~T(); |
| 419 | } |
| 420 | }; |
| 421 | |
| 422 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
| 423 | template <typename T, bool TriviallyCopyable> |
| 424 | void SmallVectorTemplateBase<T, TriviallyCopyable>::grow(size_t MinSize) { |
| 425 | size_t NewCapacity; |
| 426 | T *NewElts = mallocForGrow(MinSize, NewCapacity); |
| 427 | moveElementsForGrow(NewElts); |
| 428 | takeAllocationForGrow(NewElts, NewCapacity); |
| 429 | } |
| 430 | |
| 431 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
| 432 | template <typename T, bool TriviallyCopyable> |
| 433 | void SmallVectorTemplateBase<T, TriviallyCopyable>::moveElementsForGrow( |
| 434 | T *NewElts) { |
| 435 | // Move the elements over. |
| 436 | this->uninitialized_move(this->begin(), this->end(), NewElts); |
| 437 | |
| 438 | // Destroy the original elements. |
| 439 | destroy_range(this->begin(), this->end()); |
| 440 | } |
| 441 | |
| 442 | // Define this out-of-line to dissuade the C++ compiler from inlining it. |
| 443 | template <typename T, bool TriviallyCopyable> |
| 444 | void SmallVectorTemplateBase<T, TriviallyCopyable>::takeAllocationForGrow( |
| 445 | T *NewElts, size_t NewCapacity) { |
| 446 | // If this wasn't grown from the inline copy, deallocate the old space. |
| 447 | if (!this->isSmall()) |
| 448 | free(this->begin()); |
| 449 | |
| 450 | this->BeginX = NewElts; |
| 451 | this->Capacity = NewCapacity; |
| 452 | } |
| 453 | |
| 454 | /// SmallVectorTemplateBase<TriviallyCopyable = true> - This is where we put |
| 455 | /// method implementations that are designed to work with trivially copyable |
| 456 | /// T's. This allows using memcpy in place of copy/move construction and |
| 457 | /// skipping destruction. |
| 458 | template <typename T> |
| 459 | class SmallVectorTemplateBase<T, true> : public SmallVectorTemplateCommon<T> { |
| 460 | friend class SmallVectorTemplateCommon<T>; |
| 461 | |
| 462 | protected: |
| 463 | /// True if it's cheap enough to take parameters by value. Doing so avoids |
| 464 | /// overhead related to mitigations for reference invalidation. |
| 465 | static constexpr bool TakesParamByValue = sizeof(T) <= 2 * sizeof(void *); |
| 466 | |
| 467 | /// Either const T& or T, depending on whether it's cheap enough to take |
| 468 | /// parameters by value. |
| 469 | using ValueParamT = |
| 470 | typename std::conditional<TakesParamByValue, T, const T &>::type; |
| 471 | |
| 472 | SmallVectorTemplateBase(size_t Size) : SmallVectorTemplateCommon<T>(Size) {} |
| 473 | |
| 474 | // No need to do a destroy loop for POD's. |
| 475 | static void destroy_range(T *, T *) {} |
| 476 | |
| 477 | /// Move the range [I, E) onto the uninitialized memory |
| 478 | /// starting with "Dest", constructing elements into it as needed. |
| 479 | template<typename It1, typename It2> |
| 480 | static void uninitialized_move(It1 I, It1 E, It2 Dest) { |
| 481 | // Just do a copy. |
| 482 | uninitialized_copy(I, E, Dest); |
| 483 | } |
| 484 | |
| 485 | /// Copy the range [I, E) onto the uninitialized memory |
| 486 | /// starting with "Dest", constructing elements into it as needed. |
| 487 | template<typename It1, typename It2> |
| 488 | static void uninitialized_copy(It1 I, It1 E, It2 Dest) { |
| 489 | // Arbitrary iterator types; just use the basic implementation. |
| 490 | std::uninitialized_copy(I, E, Dest); |
| 491 | } |
| 492 | |
| 493 | /// Copy the range [I, E) onto the uninitialized memory |
| 494 | /// starting with "Dest", constructing elements into it as needed. |
| 495 | template <typename T1, typename T2> |
| 496 | static void uninitialized_copy( |
| 497 | T1 *I, T1 *E, T2 *Dest, |
| 498 | std::enable_if_t<std::is_same<typename std::remove_const<T1>::type, |
| 499 | T2>::value> * = nullptr) { |
| 500 | // Use memcpy for PODs iterated by pointers (which includes SmallVector |
| 501 | // iterators): std::uninitialized_copy optimizes to memmove, but we can |
| 502 | // use memcpy here. Note that I and E are iterators and thus might be |
| 503 | // invalid for memcpy if they are equal. |
| 504 | if (I != E) |
| 505 | memcpy(reinterpret_cast<void *>(Dest), I, (E - I) * sizeof(T)); |
| 506 | } |
| 507 | |
| 508 | /// Double the size of the allocated memory, guaranteeing space for at |
| 509 | /// least one more element or MinSize if specified. |
| 510 | void grow(size_t MinSize = 0) { this->grow_pod(MinSize, sizeof(T)); } |
| 511 | |
| 512 | /// Reserve enough space to add one element, and return the updated element |
| 513 | /// pointer in case it was a reference to the storage. |
| 514 | const T *reserveForParamAndGetAddress(const T &Elt, size_t N = 1) { |
| 515 | return this->reserveForParamAndGetAddressImpl(this, Elt, N); |
| 516 | } |
| 517 | |
| 518 | /// Reserve enough space to add one element, and return the updated element |
| 519 | /// pointer in case it was a reference to the storage. |
| 520 | T *reserveForParamAndGetAddress(T &Elt, size_t N = 1) { |
| 521 | return const_cast<T *>( |
| 522 | this->reserveForParamAndGetAddressImpl(this, Elt, N)); |
| 523 | } |
| 524 | |
| 525 | /// Copy \p V or return a reference, depending on \a ValueParamT. |
| 526 | static ValueParamT forward_value_param(ValueParamT V) { return V; } |
| 527 | |
| 528 | void growAndAssign(size_t NumElts, T Elt) { |
| 529 | // Elt has been copied in case it's an internal reference, side-stepping |
| 530 | // reference invalidation problems without losing the realloc optimization. |
| 531 | this->set_size(0); |
| 532 | this->grow(NumElts); |
| 533 | std::uninitialized_fill_n(this->begin(), NumElts, Elt); |
| 534 | this->set_size(NumElts); |
| 535 | } |
| 536 | |
| 537 | template <typename... ArgTypes> T &growAndEmplaceBack(ArgTypes &&... Args) { |
| 538 | // Use push_back with a copy in case Args has an internal reference, |
| 539 | // side-stepping reference invalidation problems without losing the realloc |
| 540 | // optimization. |
| 541 | push_back(T(std::forward<ArgTypes>(Args)...)); |
| 542 | return this->back(); |
| 543 | } |
| 544 | |
| 545 | public: |
| 546 | void push_back(ValueParamT Elt) { |
| 547 | const T *EltPtr = reserveForParamAndGetAddress(Elt); |
| 548 | memcpy(reinterpret_cast<void *>(this->end()), EltPtr, sizeof(T)); |
| 549 | this->set_size(this->size() + 1); |
| 550 | } |
| 551 | |
| 552 | void pop_back() { this->set_size(this->size() - 1); } |
| 553 | }; |
| 554 | |
| 555 | /// This class consists of common code factored out of the SmallVector class to |
| 556 | /// reduce code duplication based on the SmallVector 'N' template parameter. |
| 557 | template <typename T> |
| 558 | class SmallVectorImpl : public SmallVectorTemplateBase<T> { |
| 559 | using SuperClass = SmallVectorTemplateBase<T>; |
| 560 | |
| 561 | public: |
| 562 | using iterator = typename SuperClass::iterator; |
| 563 | using const_iterator = typename SuperClass::const_iterator; |
| 564 | using reference = typename SuperClass::reference; |
| 565 | using size_type = typename SuperClass::size_type; |
| 566 | |
| 567 | protected: |
| 568 | using SmallVectorTemplateBase<T>::TakesParamByValue; |
| 569 | using ValueParamT = typename SuperClass::ValueParamT; |
| 570 | |
| 571 | // Default ctor - Initialize to empty. |
| 572 | explicit SmallVectorImpl(unsigned N) |
| 573 | : SmallVectorTemplateBase<T>(N) {} |
| 574 | |
| 575 | public: |
| 576 | SmallVectorImpl(const SmallVectorImpl &) = delete; |
| 577 | |
| 578 | ~SmallVectorImpl() { |
| 579 | // Subclass has already destructed this vector's elements. |
| 580 | // If this wasn't grown from the inline copy, deallocate the old space. |
| 581 | if (!this->isSmall()) |
| 582 | free(this->begin()); |
| 583 | } |
| 584 | |
| 585 | void clear() { |
| 586 | this->destroy_range(this->begin(), this->end()); |
| 587 | this->Size = 0; |
| 588 | } |
| 589 | |
| 590 | private: |
| 591 | template <bool ForOverwrite> void resizeImpl(size_type N) { |
| 592 | if (N < this->size()) { |
| 593 | this->pop_back_n(this->size() - N); |
| 594 | } else if (N > this->size()) { |
| 595 | this->reserve(N); |
| 596 | for (auto I = this->end(), E = this->begin() + N; I != E; ++I) |
| 597 | if (ForOverwrite) |
| 598 | new (&*I) T; |
| 599 | else |
| 600 | new (&*I) T(); |
| 601 | this->set_size(N); |
| 602 | } |
| 603 | } |
| 604 | |
| 605 | public: |
| 606 | void resize(size_type N) { resizeImpl<false>(N); } |
| 607 | |
| 608 | /// Like resize, but \ref T is POD, the new values won't be initialized. |
| 609 | void resize_for_overwrite(size_type N) { resizeImpl<true>(N); } |
| 610 | |
| 611 | void resize(size_type N, ValueParamT NV) { |
| 612 | if (N == this->size()) |
| 613 | return; |
| 614 | |
| 615 | if (N < this->size()) { |
| 616 | this->pop_back_n(this->size() - N); |
| 617 | return; |
| 618 | } |
| 619 | |
| 620 | // N > this->size(). Defer to append. |
| 621 | this->append(N - this->size(), NV); |
| 622 | } |
| 623 | |
| 624 | void reserve(size_type N) { |
| 625 | if (this->capacity() < N) |
| 626 | this->grow(N); |
| 627 | } |
| 628 | |
| 629 | void pop_back_n(size_type NumItems) { |
| 630 | assert(this->size() >= NumItems)((void)0); |
| 631 | this->destroy_range(this->end() - NumItems, this->end()); |
| 632 | this->set_size(this->size() - NumItems); |
| 633 | } |
| 634 | |
| 635 | LLVM_NODISCARD[[clang::warn_unused_result]] T pop_back_val() { |
| 636 | T Result = ::std::move(this->back()); |
| 637 | this->pop_back(); |
| 638 | return Result; |
| 639 | } |
| 640 | |
| 641 | void swap(SmallVectorImpl &RHS); |
| 642 | |
| 643 | /// Add the specified range to the end of the SmallVector. |
| 644 | template <typename in_iter, |
| 645 | typename = std::enable_if_t<std::is_convertible< |
| 646 | typename std::iterator_traits<in_iter>::iterator_category, |
| 647 | std::input_iterator_tag>::value>> |
| 648 | void append(in_iter in_start, in_iter in_end) { |
| 649 | this->assertSafeToAddRange(in_start, in_end); |
| 650 | size_type NumInputs = std::distance(in_start, in_end); |
| 651 | this->reserve(this->size() + NumInputs); |
| 652 | this->uninitialized_copy(in_start, in_end, this->end()); |
| 653 | this->set_size(this->size() + NumInputs); |
| 654 | } |
| 655 | |
| 656 | /// Append \p NumInputs copies of \p Elt to the end. |
| 657 | void append(size_type NumInputs, ValueParamT Elt) { |
| 658 | const T *EltPtr = this->reserveForParamAndGetAddress(Elt, NumInputs); |
| 659 | std::uninitialized_fill_n(this->end(), NumInputs, *EltPtr); |
| 660 | this->set_size(this->size() + NumInputs); |
| 661 | } |
| 662 | |
| 663 | void append(std::initializer_list<T> IL) { |
| 664 | append(IL.begin(), IL.end()); |
| 665 | } |
| 666 | |
| 667 | void append(const SmallVectorImpl &RHS) { append(RHS.begin(), RHS.end()); } |
| 668 | |
| 669 | void assign(size_type NumElts, ValueParamT Elt) { |
| 670 | // Note that Elt could be an internal reference. |
| 671 | if (NumElts > this->capacity()) { |
| 672 | this->growAndAssign(NumElts, Elt); |
| 673 | return; |
| 674 | } |
| 675 | |
| 676 | // Assign over existing elements. |
| 677 | std::fill_n(this->begin(), std::min(NumElts, this->size()), Elt); |
| 678 | if (NumElts > this->size()) |
| 679 | std::uninitialized_fill_n(this->end(), NumElts - this->size(), Elt); |
| 680 | else if (NumElts < this->size()) |
| 681 | this->destroy_range(this->begin() + NumElts, this->end()); |
| 682 | this->set_size(NumElts); |
| 683 | } |
| 684 | |
| 685 | // FIXME: Consider assigning over existing elements, rather than clearing & |
| 686 | // re-initializing them - for all assign(...) variants. |
| 687 | |
| 688 | template <typename in_iter, |
| 689 | typename = std::enable_if_t<std::is_convertible< |
| 690 | typename std::iterator_traits<in_iter>::iterator_category, |
| 691 | std::input_iterator_tag>::value>> |
| 692 | void assign(in_iter in_start, in_iter in_end) { |
| 693 | this->assertSafeToReferenceAfterClear(in_start, in_end); |
| 694 | clear(); |
| 695 | append(in_start, in_end); |
| 696 | } |
| 697 | |
| 698 | void assign(std::initializer_list<T> IL) { |
| 699 | clear(); |
| 700 | append(IL); |
| 701 | } |
| 702 | |
| 703 | void assign(const SmallVectorImpl &RHS) { assign(RHS.begin(), RHS.end()); } |
| 704 | |
| 705 | iterator erase(const_iterator CI) { |
| 706 | // Just cast away constness because this is a non-const member function. |
| 707 | iterator I = const_cast<iterator>(CI); |
| 708 | |
| 709 | assert(this->isReferenceToStorage(CI) && "Iterator to erase is out of bounds.")((void)0); |
| 710 | |
| 711 | iterator N = I; |
| 712 | // Shift all elts down one. |
| 713 | std::move(I+1, this->end(), I); |
| 714 | // Drop the last elt. |
| 715 | this->pop_back(); |
| 716 | return(N); |
| 717 | } |
| 718 | |
| 719 | iterator erase(const_iterator CS, const_iterator CE) { |
| 720 | // Just cast away constness because this is a non-const member function. |
| 721 | iterator S = const_cast<iterator>(CS); |
| 722 | iterator E = const_cast<iterator>(CE); |
| 723 | |
| 724 | assert(this->isRangeInStorage(S, E) && "Range to erase is out of bounds.")((void)0); |
| 725 | |
| 726 | iterator N = S; |
| 727 | // Shift all elts down. |
| 728 | iterator I = std::move(E, this->end(), S); |
| 729 | // Drop the last elts. |
| 730 | this->destroy_range(I, this->end()); |
| 731 | this->set_size(I - this->begin()); |
| 732 | return(N); |
| 733 | } |
| 734 | |
| 735 | private: |
| 736 | template <class ArgType> iterator insert_one_impl(iterator I, ArgType &&Elt) { |
| 737 | // Callers ensure that ArgType is derived from T. |
| 738 | static_assert( |
| 739 | std::is_same<std::remove_const_t<std::remove_reference_t<ArgType>>, |
| 740 | T>::value, |
| 741 | "ArgType must be derived from T!"); |
| 742 | |
| 743 | if (I == this->end()) { // Important special case for empty vector. |
| 744 | this->push_back(::std::forward<ArgType>(Elt)); |
| 745 | return this->end()-1; |
| 746 | } |
| 747 | |
| 748 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")((void)0); |
| 749 | |
| 750 | // Grow if necessary. |
| 751 | size_t Index = I - this->begin(); |
| 752 | std::remove_reference_t<ArgType> *EltPtr = |
| 753 | this->reserveForParamAndGetAddress(Elt); |
| 754 | I = this->begin() + Index; |
| 755 | |
| 756 | ::new ((void*) this->end()) T(::std::move(this->back())); |
| 757 | // Push everything else over. |
| 758 | std::move_backward(I, this->end()-1, this->end()); |
| 759 | this->set_size(this->size() + 1); |
| 760 | |
| 761 | // If we just moved the element we're inserting, be sure to update |
| 762 | // the reference (never happens if TakesParamByValue). |
| 763 | static_assert(!TakesParamByValue || std::is_same<ArgType, T>::value, |
| 764 | "ArgType must be 'T' when taking by value!"); |
| 765 | if (!TakesParamByValue && this->isReferenceToRange(EltPtr, I, this->end())) |
| 766 | ++EltPtr; |
| 767 | |
| 768 | *I = ::std::forward<ArgType>(*EltPtr); |
| 769 | return I; |
| 770 | } |
| 771 | |
| 772 | public: |
| 773 | iterator insert(iterator I, T &&Elt) { |
| 774 | return insert_one_impl(I, this->forward_value_param(std::move(Elt))); |
| 775 | } |
| 776 | |
| 777 | iterator insert(iterator I, const T &Elt) { |
| 778 | return insert_one_impl(I, this->forward_value_param(Elt)); |
| 779 | } |
| 780 | |
| 781 | iterator insert(iterator I, size_type NumToInsert, ValueParamT Elt) { |
| 782 | // Convert iterator to elt# to avoid invalidating iterator when we reserve() |
| 783 | size_t InsertElt = I - this->begin(); |
| 784 | |
| 785 | if (I == this->end()) { // Important special case for empty vector. |
| 786 | append(NumToInsert, Elt); |
| 787 | return this->begin()+InsertElt; |
| 788 | } |
| 789 | |
| 790 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")((void)0); |
| 791 | |
| 792 | // Ensure there is enough space, and get the (maybe updated) address of |
| 793 | // Elt. |
| 794 | const T *EltPtr = this->reserveForParamAndGetAddress(Elt, NumToInsert); |
| 795 | |
| 796 | // Uninvalidate the iterator. |
| 797 | I = this->begin()+InsertElt; |
| 798 | |
| 799 | // If there are more elements between the insertion point and the end of the |
| 800 | // range than there are being inserted, we can use a simple approach to |
| 801 | // insertion. Since we already reserved space, we know that this won't |
| 802 | // reallocate the vector. |
| 803 | if (size_t(this->end()-I) >= NumToInsert) { |
| 804 | T *OldEnd = this->end(); |
| 805 | append(std::move_iterator<iterator>(this->end() - NumToInsert), |
| 806 | std::move_iterator<iterator>(this->end())); |
| 807 | |
| 808 | // Copy the existing elements that get replaced. |
| 809 | std::move_backward(I, OldEnd-NumToInsert, OldEnd); |
| 810 | |
| 811 | // If we just moved the element we're inserting, be sure to update |
| 812 | // the reference (never happens if TakesParamByValue). |
| 813 | if (!TakesParamByValue && I <= EltPtr && EltPtr < this->end()) |
| 814 | EltPtr += NumToInsert; |
| 815 | |
| 816 | std::fill_n(I, NumToInsert, *EltPtr); |
| 817 | return I; |
| 818 | } |
| 819 | |
| 820 | // Otherwise, we're inserting more elements than exist already, and we're |
| 821 | // not inserting at the end. |
| 822 | |
| 823 | // Move over the elements that we're about to overwrite. |
| 824 | T *OldEnd = this->end(); |
| 825 | this->set_size(this->size() + NumToInsert); |
| 826 | size_t NumOverwritten = OldEnd-I; |
| 827 | this->uninitialized_move(I, OldEnd, this->end()-NumOverwritten); |
| 828 | |
| 829 | // If we just moved the element we're inserting, be sure to update |
| 830 | // the reference (never happens if TakesParamByValue). |
| 831 | if (!TakesParamByValue && I <= EltPtr && EltPtr < this->end()) |
| 832 | EltPtr += NumToInsert; |
| 833 | |
| 834 | // Replace the overwritten part. |
| 835 | std::fill_n(I, NumOverwritten, *EltPtr); |
| 836 | |
| 837 | // Insert the non-overwritten middle part. |
| 838 | std::uninitialized_fill_n(OldEnd, NumToInsert - NumOverwritten, *EltPtr); |
| 839 | return I; |
| 840 | } |
| 841 | |
| 842 | template <typename ItTy, |
| 843 | typename = std::enable_if_t<std::is_convertible< |
| 844 | typename std::iterator_traits<ItTy>::iterator_category, |
| 845 | std::input_iterator_tag>::value>> |
| 846 | iterator insert(iterator I, ItTy From, ItTy To) { |
| 847 | // Convert iterator to elt# to avoid invalidating iterator when we reserve() |
| 848 | size_t InsertElt = I - this->begin(); |
| 849 | |
| 850 | if (I == this->end()) { // Important special case for empty vector. |
| 851 | append(From, To); |
| 852 | return this->begin()+InsertElt; |
| 853 | } |
| 854 | |
| 855 | assert(this->isReferenceToStorage(I) && "Insertion iterator is out of bounds.")((void)0); |
| 856 | |
| 857 | // Check that the reserve that follows doesn't invalidate the iterators. |
| 858 | this->assertSafeToAddRange(From, To); |
| 859 | |
| 860 | size_t NumToInsert = std::distance(From, To); |
| 861 | |
| 862 | // Ensure there is enough space. |
| 863 | reserve(this->size() + NumToInsert); |
| 864 | |
| 865 | // Uninvalidate the iterator. |
| 866 | I = this->begin()+InsertElt; |
| 867 | |
| 868 | // If there are more elements between the insertion point and the end of the |
| 869 | // range than there are being inserted, we can use a simple approach to |
| 870 | // insertion. Since we already reserved space, we know that this won't |
| 871 | // reallocate the vector. |
| 872 | if (size_t(this->end()-I) >= NumToInsert) { |
| 873 | T *OldEnd = this->end(); |
| 874 | append(std::move_iterator<iterator>(this->end() - NumToInsert), |
| 875 | std::move_iterator<iterator>(this->end())); |
| 876 | |
| 877 | // Copy the existing elements that get replaced. |
| 878 | std::move_backward(I, OldEnd-NumToInsert, OldEnd); |
| 879 | |
| 880 | std::copy(From, To, I); |
| 881 | return I; |
| 882 | } |
| 883 | |
| 884 | // Otherwise, we're inserting more elements than exist already, and we're |
| 885 | // not inserting at the end. |
| 886 | |
| 887 | // Move over the elements that we're about to overwrite. |
| 888 | T *OldEnd = this->end(); |
| 889 | this->set_size(this->size() + NumToInsert); |
| 890 | size_t NumOverwritten = OldEnd-I; |
| 891 | this->uninitialized_move(I, OldEnd, this->end()-NumOverwritten); |
| 892 | |
| 893 | // Replace the overwritten part. |
| 894 | for (T *J = I; NumOverwritten > 0; --NumOverwritten) { |
| 895 | *J = *From; |
| 896 | ++J; ++From; |
| 897 | } |
| 898 | |
| 899 | // Insert the non-overwritten middle part. |
| 900 | this->uninitialized_copy(From, To, OldEnd); |
| 901 | return I; |
| 902 | } |
| 903 | |
| 904 | void insert(iterator I, std::initializer_list<T> IL) { |
| 905 | insert(I, IL.begin(), IL.end()); |
| 906 | } |
| 907 | |
| 908 | template <typename... ArgTypes> reference emplace_back(ArgTypes &&... Args) { |
| 909 | if (LLVM_UNLIKELY(this->size() >= this->capacity())__builtin_expect((bool)(this->size() >= this->capacity ()), false)) |
| 910 | return this->growAndEmplaceBack(std::forward<ArgTypes>(Args)...); |
| 911 | |
| 912 | ::new ((void *)this->end()) T(std::forward<ArgTypes>(Args)...); |
| 913 | this->set_size(this->size() + 1); |
| 914 | return this->back(); |
| 915 | } |
| 916 | |
| 917 | SmallVectorImpl &operator=(const SmallVectorImpl &RHS); |
| 918 | |
| 919 | SmallVectorImpl &operator=(SmallVectorImpl &&RHS); |
| 920 | |
| 921 | bool operator==(const SmallVectorImpl &RHS) const { |
| 922 | if (this->size() != RHS.size()) return false; |
| 923 | return std::equal(this->begin(), this->end(), RHS.begin()); |
| 924 | } |
| 925 | bool operator!=(const SmallVectorImpl &RHS) const { |
| 926 | return !(*this == RHS); |
| 927 | } |
| 928 | |
| 929 | bool operator<(const SmallVectorImpl &RHS) const { |
| 930 | return std::lexicographical_compare(this->begin(), this->end(), |
| 931 | RHS.begin(), RHS.end()); |
| 932 | } |
| 933 | }; |
| 934 | |
| 935 | template <typename T> |
| 936 | void SmallVectorImpl<T>::swap(SmallVectorImpl<T> &RHS) { |
| 937 | if (this == &RHS) return; |
| 938 | |
| 939 | // We can only avoid copying elements if neither vector is small. |
| 940 | if (!this->isSmall() && !RHS.isSmall()) { |
| 941 | std::swap(this->BeginX, RHS.BeginX); |
| 942 | std::swap(this->Size, RHS.Size); |
| 943 | std::swap(this->Capacity, RHS.Capacity); |
| 944 | return; |
| 945 | } |
| 946 | this->reserve(RHS.size()); |
| 947 | RHS.reserve(this->size()); |
| 948 | |
| 949 | // Swap the shared elements. |
| 950 | size_t NumShared = this->size(); |
| 951 | if (NumShared > RHS.size()) NumShared = RHS.size(); |
| 952 | for (size_type i = 0; i != NumShared; ++i) |
| 953 | std::swap((*this)[i], RHS[i]); |
| 954 | |
| 955 | // Copy over the extra elts. |
| 956 | if (this->size() > RHS.size()) { |
| 957 | size_t EltDiff = this->size() - RHS.size(); |
| 958 | this->uninitialized_copy(this->begin()+NumShared, this->end(), RHS.end()); |
| 959 | RHS.set_size(RHS.size() + EltDiff); |
| 960 | this->destroy_range(this->begin()+NumShared, this->end()); |
| 961 | this->set_size(NumShared); |
| 962 | } else if (RHS.size() > this->size()) { |
| 963 | size_t EltDiff = RHS.size() - this->size(); |
| 964 | this->uninitialized_copy(RHS.begin()+NumShared, RHS.end(), this->end()); |
| 965 | this->set_size(this->size() + EltDiff); |
| 966 | this->destroy_range(RHS.begin()+NumShared, RHS.end()); |
| 967 | RHS.set_size(NumShared); |
| 968 | } |
| 969 | } |
| 970 | |
| 971 | template <typename T> |
| 972 | SmallVectorImpl<T> &SmallVectorImpl<T>:: |
| 973 | operator=(const SmallVectorImpl<T> &RHS) { |
| 974 | // Avoid self-assignment. |
| 975 | if (this == &RHS) return *this; |
| 976 | |
| 977 | // If we already have sufficient space, assign the common elements, then |
| 978 | // destroy any excess. |
| 979 | size_t RHSSize = RHS.size(); |
| 980 | size_t CurSize = this->size(); |
| 981 | if (CurSize >= RHSSize) { |
| 982 | // Assign common elements. |
| 983 | iterator NewEnd; |
| 984 | if (RHSSize) |
| 985 | NewEnd = std::copy(RHS.begin(), RHS.begin()+RHSSize, this->begin()); |
| 986 | else |
| 987 | NewEnd = this->begin(); |
| 988 | |
| 989 | // Destroy excess elements. |
| 990 | this->destroy_range(NewEnd, this->end()); |
| 991 | |
| 992 | // Trim. |
| 993 | this->set_size(RHSSize); |
| 994 | return *this; |
| 995 | } |
| 996 | |
| 997 | // If we have to grow to have enough elements, destroy the current elements. |
| 998 | // This allows us to avoid copying them during the grow. |
| 999 | // FIXME: don't do this if they're efficiently moveable. |
| 1000 | if (this->capacity() < RHSSize) { |
| 1001 | // Destroy current elements. |
| 1002 | this->clear(); |
| 1003 | CurSize = 0; |
| 1004 | this->grow(RHSSize); |
| 1005 | } else if (CurSize) { |
| 1006 | // Otherwise, use assignment for the already-constructed elements. |
| 1007 | std::copy(RHS.begin(), RHS.begin()+CurSize, this->begin()); |
| 1008 | } |
| 1009 | |
| 1010 | // Copy construct the new elements in place. |
| 1011 | this->uninitialized_copy(RHS.begin()+CurSize, RHS.end(), |
| 1012 | this->begin()+CurSize); |
| 1013 | |
| 1014 | // Set end. |
| 1015 | this->set_size(RHSSize); |
| 1016 | return *this; |
| 1017 | } |
| 1018 | |
| 1019 | template <typename T> |
| 1020 | SmallVectorImpl<T> &SmallVectorImpl<T>::operator=(SmallVectorImpl<T> &&RHS) { |
| 1021 | // Avoid self-assignment. |
| 1022 | if (this == &RHS) return *this; |
| 1023 | |
| 1024 | // If the RHS isn't small, clear this vector and then steal its buffer. |
| 1025 | if (!RHS.isSmall()) { |
| 1026 | this->destroy_range(this->begin(), this->end()); |
| 1027 | if (!this->isSmall()) free(this->begin()); |
| 1028 | this->BeginX = RHS.BeginX; |
| 1029 | this->Size = RHS.Size; |
| 1030 | this->Capacity = RHS.Capacity; |
| 1031 | RHS.resetToSmall(); |
| 1032 | return *this; |
| 1033 | } |
| 1034 | |
| 1035 | // If we already have sufficient space, assign the common elements, then |
| 1036 | // destroy any excess. |
| 1037 | size_t RHSSize = RHS.size(); |
| 1038 | size_t CurSize = this->size(); |
| 1039 | if (CurSize >= RHSSize) { |
| 1040 | // Assign common elements. |
| 1041 | iterator NewEnd = this->begin(); |
| 1042 | if (RHSSize) |
| 1043 | NewEnd = std::move(RHS.begin(), RHS.end(), NewEnd); |
| 1044 | |
| 1045 | // Destroy excess elements and trim the bounds. |
| 1046 | this->destroy_range(NewEnd, this->end()); |
| 1047 | this->set_size(RHSSize); |
| 1048 | |
| 1049 | // Clear the RHS. |
| 1050 | RHS.clear(); |
| 1051 | |
| 1052 | return *this; |
| 1053 | } |
| 1054 | |
| 1055 | // If we have to grow to have enough elements, destroy the current elements. |
| 1056 | // This allows us to avoid copying them during the grow. |
| 1057 | // FIXME: this may not actually make any sense if we can efficiently move |
| 1058 | // elements. |
| 1059 | if (this->capacity() < RHSSize) { |
| 1060 | // Destroy current elements. |
| 1061 | this->clear(); |
| 1062 | CurSize = 0; |
| 1063 | this->grow(RHSSize); |
| 1064 | } else if (CurSize) { |
| 1065 | // Otherwise, use assignment for the already-constructed elements. |
| 1066 | std::move(RHS.begin(), RHS.begin()+CurSize, this->begin()); |
| 1067 | } |
| 1068 | |
| 1069 | // Move-construct the new elements in place. |
| 1070 | this->uninitialized_move(RHS.begin()+CurSize, RHS.end(), |
| 1071 | this->begin()+CurSize); |
| 1072 | |
| 1073 | // Set end. |
| 1074 | this->set_size(RHSSize); |
| 1075 | |
| 1076 | RHS.clear(); |
| 1077 | return *this; |
| 1078 | } |
| 1079 | |
| 1080 | /// Storage for the SmallVector elements. This is specialized for the N=0 case |
| 1081 | /// to avoid allocating unnecessary storage. |
| 1082 | template <typename T, unsigned N> |
| 1083 | struct SmallVectorStorage { |
| 1084 | alignas(T) char InlineElts[N * sizeof(T)]; |
| 1085 | }; |
| 1086 | |
| 1087 | /// We need the storage to be properly aligned even for small-size of 0 so that |
| 1088 | /// the pointer math in \a SmallVectorTemplateCommon::getFirstEl() is |
| 1089 | /// well-defined. |
| 1090 | template <typename T> struct alignas(T) SmallVectorStorage<T, 0> {}; |
| 1091 | |
| 1092 | /// Forward declaration of SmallVector so that |
| 1093 | /// calculateSmallVectorDefaultInlinedElements can reference |
| 1094 | /// `sizeof(SmallVector<T, 0>)`. |
| 1095 | template <typename T, unsigned N> class LLVM_GSL_OWNER[[gsl::Owner]] SmallVector; |
| 1096 | |
| 1097 | /// Helper class for calculating the default number of inline elements for |
| 1098 | /// `SmallVector<T>`. |
| 1099 | /// |
| 1100 | /// This should be migrated to a constexpr function when our minimum |
| 1101 | /// compiler support is enough for multi-statement constexpr functions. |
| 1102 | template <typename T> struct CalculateSmallVectorDefaultInlinedElements { |
| 1103 | // Parameter controlling the default number of inlined elements |
| 1104 | // for `SmallVector<T>`. |
| 1105 | // |
| 1106 | // The default number of inlined elements ensures that |
| 1107 | // 1. There is at least one inlined element. |
| 1108 | // 2. `sizeof(SmallVector<T>) <= kPreferredSmallVectorSizeof` unless |
| 1109 | // it contradicts 1. |
| 1110 | static constexpr size_t kPreferredSmallVectorSizeof = 64; |
| 1111 | |
| 1112 | // static_assert that sizeof(T) is not "too big". |
| 1113 | // |
| 1114 | // Because our policy guarantees at least one inlined element, it is possible |
| 1115 | // for an arbitrarily large inlined element to allocate an arbitrarily large |
| 1116 | // amount of inline storage. We generally consider it an antipattern for a |
| 1117 | // SmallVector to allocate an excessive amount of inline storage, so we want |
| 1118 | // to call attention to these cases and make sure that users are making an |
| 1119 | // intentional decision if they request a lot of inline storage. |
| 1120 | // |
| 1121 | // We want this assertion to trigger in pathological cases, but otherwise |
| 1122 | // not be too easy to hit. To accomplish that, the cutoff is actually somewhat |
| 1123 | // larger than kPreferredSmallVectorSizeof (otherwise, |
| 1124 | // `SmallVector<SmallVector<T>>` would be one easy way to trip it, and that |
| 1125 | // pattern seems useful in practice). |
| 1126 | // |
| 1127 | // One wrinkle is that this assertion is in theory non-portable, since |
| 1128 | // sizeof(T) is in general platform-dependent. However, we don't expect this |
| 1129 | // to be much of an issue, because most LLVM development happens on 64-bit |
| 1130 | // hosts, and therefore sizeof(T) is expected to *decrease* when compiled for |
| 1131 | // 32-bit hosts, dodging the issue. The reverse situation, where development |
| 1132 | // happens on a 32-bit host and then fails due to sizeof(T) *increasing* on a |
| 1133 | // 64-bit host, is expected to be very rare. |
| 1134 | static_assert( |
| 1135 | sizeof(T) <= 256, |
| 1136 | "You are trying to use a default number of inlined elements for " |
| 1137 | "`SmallVector<T>` but `sizeof(T)` is really big! Please use an " |
| 1138 | "explicit number of inlined elements with `SmallVector<T, N>` to make " |
| 1139 | "sure you really want that much inline storage."); |
| 1140 | |
| 1141 | // Discount the size of the header itself when calculating the maximum inline |
| 1142 | // bytes. |
| 1143 | static constexpr size_t PreferredInlineBytes = |
| 1144 | kPreferredSmallVectorSizeof - sizeof(SmallVector<T, 0>); |
| 1145 | static constexpr size_t NumElementsThatFit = PreferredInlineBytes / sizeof(T); |
| 1146 | static constexpr size_t value = |
| 1147 | NumElementsThatFit == 0 ? 1 : NumElementsThatFit; |
| 1148 | }; |
| 1149 | |
| 1150 | /// This is a 'vector' (really, a variable-sized array), optimized |
| 1151 | /// for the case when the array is small. It contains some number of elements |
| 1152 | /// in-place, which allows it to avoid heap allocation when the actual number of |
| 1153 | /// elements is below that threshold. This allows normal "small" cases to be |
| 1154 | /// fast without losing generality for large inputs. |
| 1155 | /// |
| 1156 | /// \note |
| 1157 | /// In the absence of a well-motivated choice for the number of inlined |
| 1158 | /// elements \p N, it is recommended to use \c SmallVector<T> (that is, |
| 1159 | /// omitting the \p N). This will choose a default number of inlined elements |
| 1160 | /// reasonable for allocation on the stack (for example, trying to keep \c |
| 1161 | /// sizeof(SmallVector<T>) around 64 bytes). |
| 1162 | /// |
| 1163 | /// \warning This does not attempt to be exception safe. |
| 1164 | /// |
| 1165 | /// \see https://llvm.org/docs/ProgrammersManual.html#llvm-adt-smallvector-h |
| 1166 | template <typename T, |
| 1167 | unsigned N = CalculateSmallVectorDefaultInlinedElements<T>::value> |
| 1168 | class LLVM_GSL_OWNER[[gsl::Owner]] SmallVector : public SmallVectorImpl<T>, |
| 1169 | SmallVectorStorage<T, N> { |
| 1170 | public: |
| 1171 | SmallVector() : SmallVectorImpl<T>(N) {} |
| 1172 | |
| 1173 | ~SmallVector() { |
| 1174 | // Destroy the constructed elements in the vector. |
| 1175 | this->destroy_range(this->begin(), this->end()); |
| 1176 | } |
| 1177 | |
| 1178 | explicit SmallVector(size_t Size, const T &Value = T()) |
| 1179 | : SmallVectorImpl<T>(N) { |
| 1180 | this->assign(Size, Value); |
| 1181 | } |
| 1182 | |
| 1183 | template <typename ItTy, |
| 1184 | typename = std::enable_if_t<std::is_convertible< |
| 1185 | typename std::iterator_traits<ItTy>::iterator_category, |
| 1186 | std::input_iterator_tag>::value>> |
| 1187 | SmallVector(ItTy S, ItTy E) : SmallVectorImpl<T>(N) { |
| 1188 | this->append(S, E); |
| 1189 | } |
| 1190 | |
| 1191 | template <typename RangeTy> |
| 1192 | explicit SmallVector(const iterator_range<RangeTy> &R) |
| 1193 | : SmallVectorImpl<T>(N) { |
| 1194 | this->append(R.begin(), R.end()); |
| 1195 | } |
| 1196 | |
| 1197 | SmallVector(std::initializer_list<T> IL) : SmallVectorImpl<T>(N) { |
| 1198 | this->assign(IL); |
| 1199 | } |
| 1200 | |
| 1201 | SmallVector(const SmallVector &RHS) : SmallVectorImpl<T>(N) { |
| 1202 | if (!RHS.empty()) |
| 1203 | SmallVectorImpl<T>::operator=(RHS); |
| 1204 | } |
| 1205 | |
| 1206 | SmallVector &operator=(const SmallVector &RHS) { |
| 1207 | SmallVectorImpl<T>::operator=(RHS); |
| 1208 | return *this; |
| 1209 | } |
| 1210 | |
| 1211 | SmallVector(SmallVector &&RHS) : SmallVectorImpl<T>(N) { |
| 1212 | if (!RHS.empty()) |
| 1213 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
| 1214 | } |
| 1215 | |
| 1216 | SmallVector(SmallVectorImpl<T> &&RHS) : SmallVectorImpl<T>(N) { |
| 1217 | if (!RHS.empty()) |
| 1218 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
| 1219 | } |
| 1220 | |
| 1221 | SmallVector &operator=(SmallVector &&RHS) { |
| 1222 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
| 1223 | return *this; |
| 1224 | } |
| 1225 | |
| 1226 | SmallVector &operator=(SmallVectorImpl<T> &&RHS) { |
| 1227 | SmallVectorImpl<T>::operator=(::std::move(RHS)); |
| 1228 | return *this; |
| 1229 | } |
| 1230 | |
| 1231 | SmallVector &operator=(std::initializer_list<T> IL) { |
| 1232 | this->assign(IL); |
| 1233 | return *this; |
| 1234 | } |
| 1235 | }; |
| 1236 | |
| 1237 | template <typename T, unsigned N> |
| 1238 | inline size_t capacity_in_bytes(const SmallVector<T, N> &X) { |
| 1239 | return X.capacity_in_bytes(); |
| 1240 | } |
| 1241 | |
| 1242 | /// Given a range of type R, iterate the entire range and return a |
| 1243 | /// SmallVector with elements of the vector. This is useful, for example, |
| 1244 | /// when you want to iterate a range and then sort the results. |
| 1245 | template <unsigned Size, typename R> |
| 1246 | SmallVector<typename std::remove_const<typename std::remove_reference< |
| 1247 | decltype(*std::begin(std::declval<R &>()))>::type>::type, |
| 1248 | Size> |
| 1249 | to_vector(R &&Range) { |
| 1250 | return {std::begin(Range), std::end(Range)}; |
| 1251 | } |
| 1252 | |
| 1253 | } // end namespace llvm |
| 1254 | |
| 1255 | namespace std { |
| 1256 | |
| 1257 | /// Implement std::swap in terms of SmallVector swap. |
| 1258 | template<typename T> |
| 1259 | inline void |
| 1260 | swap(llvm::SmallVectorImpl<T> &LHS, llvm::SmallVectorImpl<T> &RHS) { |
| 1261 | LHS.swap(RHS); |
| 1262 | } |
| 1263 | |
| 1264 | /// Implement std::swap in terms of SmallVector swap. |
| 1265 | template<typename T, unsigned N> |
| 1266 | inline void |
| 1267 | swap(llvm::SmallVector<T, N> &LHS, llvm::SmallVector<T, N> &RHS) { |
| 1268 | LHS.swap(RHS); |
| 1269 | } |
| 1270 | |
| 1271 | } // end namespace std |
| 1272 | |
| 1273 | #endif // LLVM_ADT_SMALLVECTOR_H |