| File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Support/GenericDomTree.h |
| Warning: | line 494, column 12 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===-- SILowerI1Copies.cpp - Lower I1 Copies -----------------------------===// | ||||||||
| 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 pass lowers all occurrences of i1 values (with a vreg_1 register class) | ||||||||
| 10 | // to lane masks (32 / 64-bit scalar registers). The pass assumes machine SSA | ||||||||
| 11 | // form and a wave-level control flow graph. | ||||||||
| 12 | // | ||||||||
| 13 | // Before this pass, values that are semantically i1 and are defined and used | ||||||||
| 14 | // within the same basic block are already represented as lane masks in scalar | ||||||||
| 15 | // registers. However, values that cross basic blocks are always transferred | ||||||||
| 16 | // between basic blocks in vreg_1 virtual registers and are lowered by this | ||||||||
| 17 | // pass. | ||||||||
| 18 | // | ||||||||
| 19 | // The only instructions that use or define vreg_1 virtual registers are COPY, | ||||||||
| 20 | // PHI, and IMPLICIT_DEF. | ||||||||
| 21 | // | ||||||||
| 22 | //===----------------------------------------------------------------------===// | ||||||||
| 23 | |||||||||
| 24 | #include "AMDGPU.h" | ||||||||
| 25 | #include "GCNSubtarget.h" | ||||||||
| 26 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" | ||||||||
| 27 | #include "llvm/CodeGen/MachineDominators.h" | ||||||||
| 28 | #include "llvm/CodeGen/MachineFunctionPass.h" | ||||||||
| 29 | #include "llvm/CodeGen/MachinePostDominators.h" | ||||||||
| 30 | #include "llvm/CodeGen/MachineSSAUpdater.h" | ||||||||
| 31 | #include "llvm/InitializePasses.h" | ||||||||
| 32 | |||||||||
| 33 | #define DEBUG_TYPE"si-i1-copies" "si-i1-copies" | ||||||||
| 34 | |||||||||
| 35 | using namespace llvm; | ||||||||
| 36 | |||||||||
| 37 | static unsigned createLaneMaskReg(MachineFunction &MF); | ||||||||
| 38 | static unsigned insertUndefLaneMask(MachineBasicBlock &MBB); | ||||||||
| 39 | |||||||||
| 40 | namespace { | ||||||||
| 41 | |||||||||
| 42 | class SILowerI1Copies : public MachineFunctionPass { | ||||||||
| 43 | public: | ||||||||
| 44 | static char ID; | ||||||||
| 45 | |||||||||
| 46 | private: | ||||||||
| 47 | bool IsWave32 = false; | ||||||||
| 48 | MachineFunction *MF = nullptr; | ||||||||
| 49 | MachineDominatorTree *DT = nullptr; | ||||||||
| 50 | MachinePostDominatorTree *PDT = nullptr; | ||||||||
| 51 | MachineRegisterInfo *MRI = nullptr; | ||||||||
| 52 | const GCNSubtarget *ST = nullptr; | ||||||||
| 53 | const SIInstrInfo *TII = nullptr; | ||||||||
| 54 | |||||||||
| 55 | unsigned ExecReg; | ||||||||
| 56 | unsigned MovOp; | ||||||||
| 57 | unsigned AndOp; | ||||||||
| 58 | unsigned OrOp; | ||||||||
| 59 | unsigned XorOp; | ||||||||
| 60 | unsigned AndN2Op; | ||||||||
| 61 | unsigned OrN2Op; | ||||||||
| 62 | |||||||||
| 63 | DenseSet<unsigned> ConstrainRegs; | ||||||||
| 64 | |||||||||
| 65 | public: | ||||||||
| 66 | SILowerI1Copies() : MachineFunctionPass(ID) { | ||||||||
| 67 | initializeSILowerI1CopiesPass(*PassRegistry::getPassRegistry()); | ||||||||
| 68 | } | ||||||||
| 69 | |||||||||
| 70 | bool runOnMachineFunction(MachineFunction &MF) override; | ||||||||
| 71 | |||||||||
| 72 | StringRef getPassName() const override { return "SI Lower i1 Copies"; } | ||||||||
| 73 | |||||||||
| 74 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||||||
| 75 | AU.setPreservesCFG(); | ||||||||
| 76 | AU.addRequired<MachineDominatorTree>(); | ||||||||
| 77 | AU.addRequired<MachinePostDominatorTree>(); | ||||||||
| 78 | MachineFunctionPass::getAnalysisUsage(AU); | ||||||||
| 79 | } | ||||||||
| 80 | |||||||||
| 81 | private: | ||||||||
| 82 | void lowerCopiesFromI1(); | ||||||||
| 83 | void lowerPhis(); | ||||||||
| 84 | void lowerCopiesToI1(); | ||||||||
| 85 | bool isConstantLaneMask(Register Reg, bool &Val) const; | ||||||||
| 86 | void buildMergeLaneMasks(MachineBasicBlock &MBB, | ||||||||
| 87 | MachineBasicBlock::iterator I, const DebugLoc &DL, | ||||||||
| 88 | unsigned DstReg, unsigned PrevReg, unsigned CurReg); | ||||||||
| 89 | MachineBasicBlock::iterator | ||||||||
| 90 | getSaluInsertionAtEnd(MachineBasicBlock &MBB) const; | ||||||||
| 91 | |||||||||
| 92 | bool isVreg1(Register Reg) const { | ||||||||
| 93 | return Reg.isVirtual() && MRI->getRegClass(Reg) == &AMDGPU::VReg_1RegClass; | ||||||||
| 94 | } | ||||||||
| 95 | |||||||||
| 96 | bool isLaneMaskReg(unsigned Reg) const { | ||||||||
| 97 | return TII->getRegisterInfo().isSGPRReg(*MRI, Reg) && | ||||||||
| 98 | TII->getRegisterInfo().getRegSizeInBits(Reg, *MRI) == | ||||||||
| 99 | ST->getWavefrontSize(); | ||||||||
| 100 | } | ||||||||
| 101 | }; | ||||||||
| 102 | |||||||||
| 103 | /// Helper class that determines the relationship between incoming values of a | ||||||||
| 104 | /// phi in the control flow graph to determine where an incoming value can | ||||||||
| 105 | /// simply be taken as a scalar lane mask as-is, and where it needs to be | ||||||||
| 106 | /// merged with another, previously defined lane mask. | ||||||||
| 107 | /// | ||||||||
| 108 | /// The approach is as follows: | ||||||||
| 109 | /// - Determine all basic blocks which, starting from the incoming blocks, | ||||||||
| 110 | /// a wave may reach before entering the def block (the block containing the | ||||||||
| 111 | /// phi). | ||||||||
| 112 | /// - If an incoming block has no predecessors in this set, we can take the | ||||||||
| 113 | /// incoming value as a scalar lane mask as-is. | ||||||||
| 114 | /// -- A special case of this is when the def block has a self-loop. | ||||||||
| 115 | /// - Otherwise, the incoming value needs to be merged with a previously | ||||||||
| 116 | /// defined lane mask. | ||||||||
| 117 | /// - If there is a path into the set of reachable blocks that does _not_ go | ||||||||
| 118 | /// through an incoming block where we can take the scalar lane mask as-is, | ||||||||
| 119 | /// we need to invent an available value for the SSAUpdater. Choices are | ||||||||
| 120 | /// 0 and undef, with differing consequences for how to merge values etc. | ||||||||
| 121 | /// | ||||||||
| 122 | /// TODO: We could use region analysis to quickly skip over SESE regions during | ||||||||
| 123 | /// the traversal. | ||||||||
| 124 | /// | ||||||||
| 125 | class PhiIncomingAnalysis { | ||||||||
| 126 | MachinePostDominatorTree &PDT; | ||||||||
| 127 | |||||||||
| 128 | // For each reachable basic block, whether it is a source in the induced | ||||||||
| 129 | // subgraph of the CFG. | ||||||||
| 130 | DenseMap<MachineBasicBlock *, bool> ReachableMap; | ||||||||
| 131 | SmallVector<MachineBasicBlock *, 4> ReachableOrdered; | ||||||||
| 132 | SmallVector<MachineBasicBlock *, 4> Stack; | ||||||||
| 133 | SmallVector<MachineBasicBlock *, 4> Predecessors; | ||||||||
| 134 | |||||||||
| 135 | public: | ||||||||
| 136 | PhiIncomingAnalysis(MachinePostDominatorTree &PDT) : PDT(PDT) {} | ||||||||
| 137 | |||||||||
| 138 | /// Returns whether \p MBB is a source in the induced subgraph of reachable | ||||||||
| 139 | /// blocks. | ||||||||
| 140 | bool isSource(MachineBasicBlock &MBB) const { | ||||||||
| 141 | return ReachableMap.find(&MBB)->second; | ||||||||
| 142 | } | ||||||||
| 143 | |||||||||
| 144 | ArrayRef<MachineBasicBlock *> predecessors() const { return Predecessors; } | ||||||||
| 145 | |||||||||
| 146 | void analyze(MachineBasicBlock &DefBlock, | ||||||||
| 147 | ArrayRef<MachineBasicBlock *> IncomingBlocks) { | ||||||||
| 148 | assert(Stack.empty())((void)0); | ||||||||
| 149 | ReachableMap.clear(); | ||||||||
| 150 | ReachableOrdered.clear(); | ||||||||
| 151 | Predecessors.clear(); | ||||||||
| 152 | |||||||||
| 153 | // Insert the def block first, so that it acts as an end point for the | ||||||||
| 154 | // traversal. | ||||||||
| 155 | ReachableMap.try_emplace(&DefBlock, false); | ||||||||
| 156 | ReachableOrdered.push_back(&DefBlock); | ||||||||
| 157 | |||||||||
| 158 | for (MachineBasicBlock *MBB : IncomingBlocks) { | ||||||||
| 159 | if (MBB == &DefBlock) { | ||||||||
| 160 | ReachableMap[&DefBlock] = true; // self-loop on DefBlock | ||||||||
| 161 | continue; | ||||||||
| 162 | } | ||||||||
| 163 | |||||||||
| 164 | ReachableMap.try_emplace(MBB, false); | ||||||||
| 165 | ReachableOrdered.push_back(MBB); | ||||||||
| 166 | |||||||||
| 167 | // If this block has a divergent terminator and the def block is its | ||||||||
| 168 | // post-dominator, the wave may first visit the other successors. | ||||||||
| 169 | bool Divergent = false; | ||||||||
| 170 | for (MachineInstr &MI : MBB->terminators()) { | ||||||||
| 171 | if (MI.getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO || | ||||||||
| 172 | MI.getOpcode() == AMDGPU::SI_IF || | ||||||||
| 173 | MI.getOpcode() == AMDGPU::SI_ELSE || | ||||||||
| 174 | MI.getOpcode() == AMDGPU::SI_LOOP) { | ||||||||
| 175 | Divergent = true; | ||||||||
| 176 | break; | ||||||||
| 177 | } | ||||||||
| 178 | } | ||||||||
| 179 | |||||||||
| 180 | if (Divergent && PDT.dominates(&DefBlock, MBB)) | ||||||||
| 181 | append_range(Stack, MBB->successors()); | ||||||||
| 182 | } | ||||||||
| 183 | |||||||||
| 184 | while (!Stack.empty()) { | ||||||||
| 185 | MachineBasicBlock *MBB = Stack.pop_back_val(); | ||||||||
| 186 | if (!ReachableMap.try_emplace(MBB, false).second) | ||||||||
| 187 | continue; | ||||||||
| 188 | ReachableOrdered.push_back(MBB); | ||||||||
| 189 | |||||||||
| 190 | append_range(Stack, MBB->successors()); | ||||||||
| 191 | } | ||||||||
| 192 | |||||||||
| 193 | for (MachineBasicBlock *MBB : ReachableOrdered) { | ||||||||
| 194 | bool HaveReachablePred = false; | ||||||||
| 195 | for (MachineBasicBlock *Pred : MBB->predecessors()) { | ||||||||
| 196 | if (ReachableMap.count(Pred)) { | ||||||||
| 197 | HaveReachablePred = true; | ||||||||
| 198 | } else { | ||||||||
| 199 | Stack.push_back(Pred); | ||||||||
| 200 | } | ||||||||
| 201 | } | ||||||||
| 202 | if (!HaveReachablePred) | ||||||||
| 203 | ReachableMap[MBB] = true; | ||||||||
| 204 | if (HaveReachablePred) { | ||||||||
| 205 | for (MachineBasicBlock *UnreachablePred : Stack) { | ||||||||
| 206 | if (!llvm::is_contained(Predecessors, UnreachablePred)) | ||||||||
| 207 | Predecessors.push_back(UnreachablePred); | ||||||||
| 208 | } | ||||||||
| 209 | } | ||||||||
| 210 | Stack.clear(); | ||||||||
| 211 | } | ||||||||
| 212 | } | ||||||||
| 213 | }; | ||||||||
| 214 | |||||||||
| 215 | /// Helper class that detects loops which require us to lower an i1 COPY into | ||||||||
| 216 | /// bitwise manipulation. | ||||||||
| 217 | /// | ||||||||
| 218 | /// Unfortunately, we cannot use LoopInfo because LoopInfo does not distinguish | ||||||||
| 219 | /// between loops with the same header. Consider this example: | ||||||||
| 220 | /// | ||||||||
| 221 | /// A-+-+ | ||||||||
| 222 | /// | | | | ||||||||
| 223 | /// B-+ | | ||||||||
| 224 | /// | | | ||||||||
| 225 | /// C---+ | ||||||||
| 226 | /// | ||||||||
| 227 | /// A is the header of a loop containing A, B, and C as far as LoopInfo is | ||||||||
| 228 | /// concerned. However, an i1 COPY in B that is used in C must be lowered to | ||||||||
| 229 | /// bitwise operations to combine results from different loop iterations when | ||||||||
| 230 | /// B has a divergent branch (since by default we will compile this code such | ||||||||
| 231 | /// that threads in a wave are merged at the entry of C). | ||||||||
| 232 | /// | ||||||||
| 233 | /// The following rule is implemented to determine whether bitwise operations | ||||||||
| 234 | /// are required: use the bitwise lowering for a def in block B if a backward | ||||||||
| 235 | /// edge to B is reachable without going through the nearest common | ||||||||
| 236 | /// post-dominator of B and all uses of the def. | ||||||||
| 237 | /// | ||||||||
| 238 | /// TODO: This rule is conservative because it does not check whether the | ||||||||
| 239 | /// relevant branches are actually divergent. | ||||||||
| 240 | /// | ||||||||
| 241 | /// The class is designed to cache the CFG traversal so that it can be re-used | ||||||||
| 242 | /// for multiple defs within the same basic block. | ||||||||
| 243 | /// | ||||||||
| 244 | /// TODO: We could use region analysis to quickly skip over SESE regions during | ||||||||
| 245 | /// the traversal. | ||||||||
| 246 | /// | ||||||||
| 247 | class LoopFinder { | ||||||||
| 248 | MachineDominatorTree &DT; | ||||||||
| 249 | MachinePostDominatorTree &PDT; | ||||||||
| 250 | |||||||||
| 251 | // All visited / reachable block, tagged by level (level 0 is the def block, | ||||||||
| 252 | // level 1 are all blocks reachable including but not going through the def | ||||||||
| 253 | // block's IPDOM, etc.). | ||||||||
| 254 | DenseMap<MachineBasicBlock *, unsigned> Visited; | ||||||||
| 255 | |||||||||
| 256 | // Nearest common dominator of all visited blocks by level (level 0 is the | ||||||||
| 257 | // def block). Used for seeding the SSAUpdater. | ||||||||
| 258 | SmallVector<MachineBasicBlock *, 4> CommonDominators; | ||||||||
| 259 | |||||||||
| 260 | // Post-dominator of all visited blocks. | ||||||||
| 261 | MachineBasicBlock *VisitedPostDom = nullptr; | ||||||||
| 262 | |||||||||
| 263 | // Level at which a loop was found: 0 is not possible; 1 = a backward edge is | ||||||||
| 264 | // reachable without going through the IPDOM of the def block (if the IPDOM | ||||||||
| 265 | // itself has an edge to the def block, the loop level is 2), etc. | ||||||||
| 266 | unsigned FoundLoopLevel = ~0u; | ||||||||
| 267 | |||||||||
| 268 | MachineBasicBlock *DefBlock = nullptr; | ||||||||
| 269 | SmallVector<MachineBasicBlock *, 4> Stack; | ||||||||
| 270 | SmallVector<MachineBasicBlock *, 4> NextLevel; | ||||||||
| 271 | |||||||||
| 272 | public: | ||||||||
| 273 | LoopFinder(MachineDominatorTree &DT, MachinePostDominatorTree &PDT) | ||||||||
| 274 | : DT(DT), PDT(PDT) {} | ||||||||
| 275 | |||||||||
| 276 | void initialize(MachineBasicBlock &MBB) { | ||||||||
| 277 | Visited.clear(); | ||||||||
| 278 | CommonDominators.clear(); | ||||||||
| 279 | Stack.clear(); | ||||||||
| 280 | NextLevel.clear(); | ||||||||
| 281 | VisitedPostDom = nullptr; | ||||||||
| 282 | FoundLoopLevel = ~0u; | ||||||||
| 283 | |||||||||
| 284 | DefBlock = &MBB; | ||||||||
| 285 | } | ||||||||
| 286 | |||||||||
| 287 | /// Check whether a backward edge can be reached without going through the | ||||||||
| 288 | /// given \p PostDom of the def block. | ||||||||
| 289 | /// | ||||||||
| 290 | /// Return the level of \p PostDom if a loop was found, or 0 otherwise. | ||||||||
| 291 | unsigned findLoop(MachineBasicBlock *PostDom) { | ||||||||
| 292 | MachineDomTreeNode *PDNode = PDT.getNode(DefBlock); | ||||||||
| 293 | |||||||||
| 294 | if (!VisitedPostDom) | ||||||||
| 295 | advanceLevel(); | ||||||||
| 296 | |||||||||
| 297 | unsigned Level = 0; | ||||||||
| 298 | while (PDNode->getBlock() != PostDom) { | ||||||||
| 299 | if (PDNode->getBlock() == VisitedPostDom) | ||||||||
| 300 | advanceLevel(); | ||||||||
| 301 | PDNode = PDNode->getIDom(); | ||||||||
| 302 | Level++; | ||||||||
| 303 | if (FoundLoopLevel == Level) | ||||||||
| 304 | return Level; | ||||||||
| 305 | } | ||||||||
| 306 | |||||||||
| 307 | return 0; | ||||||||
| 308 | } | ||||||||
| 309 | |||||||||
| 310 | /// Add undef values dominating the loop and the optionally given additional | ||||||||
| 311 | /// blocks, so that the SSA updater doesn't have to search all the way to the | ||||||||
| 312 | /// function entry. | ||||||||
| 313 | void addLoopEntries(unsigned LoopLevel, MachineSSAUpdater &SSAUpdater, | ||||||||
| 314 | ArrayRef<MachineBasicBlock *> Blocks = {}) { | ||||||||
| 315 | assert(LoopLevel < CommonDominators.size())((void)0); | ||||||||
| 316 | |||||||||
| 317 | MachineBasicBlock *Dom = CommonDominators[LoopLevel]; | ||||||||
| 318 | for (MachineBasicBlock *MBB : Blocks) | ||||||||
| 319 | Dom = DT.findNearestCommonDominator(Dom, MBB); | ||||||||
| 320 | |||||||||
| 321 | if (!inLoopLevel(*Dom, LoopLevel, Blocks)) { | ||||||||
| 322 | SSAUpdater.AddAvailableValue(Dom, insertUndefLaneMask(*Dom)); | ||||||||
| 323 | } else { | ||||||||
| 324 | // The dominator is part of the loop or the given blocks, so add the | ||||||||
| 325 | // undef value to unreachable predecessors instead. | ||||||||
| 326 | for (MachineBasicBlock *Pred : Dom->predecessors()) { | ||||||||
| 327 | if (!inLoopLevel(*Pred, LoopLevel, Blocks)) | ||||||||
| 328 | SSAUpdater.AddAvailableValue(Pred, insertUndefLaneMask(*Pred)); | ||||||||
| 329 | } | ||||||||
| 330 | } | ||||||||
| 331 | } | ||||||||
| 332 | |||||||||
| 333 | private: | ||||||||
| 334 | bool inLoopLevel(MachineBasicBlock &MBB, unsigned LoopLevel, | ||||||||
| 335 | ArrayRef<MachineBasicBlock *> Blocks) const { | ||||||||
| 336 | auto DomIt = Visited.find(&MBB); | ||||||||
| 337 | if (DomIt != Visited.end() && DomIt->second <= LoopLevel) | ||||||||
| 338 | return true; | ||||||||
| 339 | |||||||||
| 340 | if (llvm::is_contained(Blocks, &MBB)) | ||||||||
| 341 | return true; | ||||||||
| 342 | |||||||||
| 343 | return false; | ||||||||
| 344 | } | ||||||||
| 345 | |||||||||
| 346 | void advanceLevel() { | ||||||||
| 347 | MachineBasicBlock *VisitedDom; | ||||||||
| 348 | |||||||||
| 349 | if (!VisitedPostDom) { | ||||||||
| 350 | VisitedPostDom = DefBlock; | ||||||||
| 351 | VisitedDom = DefBlock; | ||||||||
| 352 | Stack.push_back(DefBlock); | ||||||||
| 353 | } else { | ||||||||
| 354 | VisitedPostDom = PDT.getNode(VisitedPostDom)->getIDom()->getBlock(); | ||||||||
| 355 | VisitedDom = CommonDominators.back(); | ||||||||
| 356 | |||||||||
| 357 | for (unsigned i = 0; i < NextLevel.size();) { | ||||||||
| 358 | if (PDT.dominates(VisitedPostDom, NextLevel[i])) { | ||||||||
| 359 | Stack.push_back(NextLevel[i]); | ||||||||
| 360 | |||||||||
| 361 | NextLevel[i] = NextLevel.back(); | ||||||||
| 362 | NextLevel.pop_back(); | ||||||||
| 363 | } else { | ||||||||
| 364 | i++; | ||||||||
| 365 | } | ||||||||
| 366 | } | ||||||||
| 367 | } | ||||||||
| 368 | |||||||||
| 369 | unsigned Level = CommonDominators.size(); | ||||||||
| 370 | while (!Stack.empty()) { | ||||||||
| 371 | MachineBasicBlock *MBB = Stack.pop_back_val(); | ||||||||
| 372 | if (!PDT.dominates(VisitedPostDom, MBB)) | ||||||||
| 373 | NextLevel.push_back(MBB); | ||||||||
| 374 | |||||||||
| 375 | Visited[MBB] = Level; | ||||||||
| 376 | VisitedDom = DT.findNearestCommonDominator(VisitedDom, MBB); | ||||||||
| 377 | |||||||||
| 378 | for (MachineBasicBlock *Succ : MBB->successors()) { | ||||||||
| 379 | if (Succ == DefBlock) { | ||||||||
| 380 | if (MBB == VisitedPostDom) | ||||||||
| 381 | FoundLoopLevel = std::min(FoundLoopLevel, Level + 1); | ||||||||
| 382 | else | ||||||||
| 383 | FoundLoopLevel = std::min(FoundLoopLevel, Level); | ||||||||
| 384 | continue; | ||||||||
| 385 | } | ||||||||
| 386 | |||||||||
| 387 | if (Visited.try_emplace(Succ, ~0u).second) { | ||||||||
| 388 | if (MBB == VisitedPostDom) | ||||||||
| 389 | NextLevel.push_back(Succ); | ||||||||
| 390 | else | ||||||||
| 391 | Stack.push_back(Succ); | ||||||||
| 392 | } | ||||||||
| 393 | } | ||||||||
| 394 | } | ||||||||
| 395 | |||||||||
| 396 | CommonDominators.push_back(VisitedDom); | ||||||||
| 397 | } | ||||||||
| 398 | }; | ||||||||
| 399 | |||||||||
| 400 | } // End anonymous namespace. | ||||||||
| 401 | |||||||||
| 402 | INITIALIZE_PASS_BEGIN(SILowerI1Copies, DEBUG_TYPE, "SI Lower i1 Copies", false,static void *initializeSILowerI1CopiesPassOnce(PassRegistry & Registry) { | ||||||||
| 403 | false)static void *initializeSILowerI1CopiesPassOnce(PassRegistry & Registry) { | ||||||||
| 404 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)initializeMachineDominatorTreePass(Registry); | ||||||||
| 405 | INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)initializeMachinePostDominatorTreePass(Registry); | ||||||||
| 406 | INITIALIZE_PASS_END(SILowerI1Copies, DEBUG_TYPE, "SI Lower i1 Copies", false,PassInfo *PI = new PassInfo( "SI Lower i1 Copies", "si-i1-copies" , &SILowerI1Copies::ID, PassInfo::NormalCtor_t(callDefaultCtor <SILowerI1Copies>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeSILowerI1CopiesPassFlag ; void llvm::initializeSILowerI1CopiesPass(PassRegistry & Registry) { llvm::call_once(InitializeSILowerI1CopiesPassFlag , initializeSILowerI1CopiesPassOnce, std::ref(Registry)); } | ||||||||
| 407 | false)PassInfo *PI = new PassInfo( "SI Lower i1 Copies", "si-i1-copies" , &SILowerI1Copies::ID, PassInfo::NormalCtor_t(callDefaultCtor <SILowerI1Copies>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeSILowerI1CopiesPassFlag ; void llvm::initializeSILowerI1CopiesPass(PassRegistry & Registry) { llvm::call_once(InitializeSILowerI1CopiesPassFlag , initializeSILowerI1CopiesPassOnce, std::ref(Registry)); } | ||||||||
| 408 | |||||||||
| 409 | char SILowerI1Copies::ID = 0; | ||||||||
| 410 | |||||||||
| 411 | char &llvm::SILowerI1CopiesID = SILowerI1Copies::ID; | ||||||||
| 412 | |||||||||
| 413 | FunctionPass *llvm::createSILowerI1CopiesPass() { | ||||||||
| 414 | return new SILowerI1Copies(); | ||||||||
| 415 | } | ||||||||
| 416 | |||||||||
| 417 | static unsigned createLaneMaskReg(MachineFunction &MF) { | ||||||||
| 418 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); | ||||||||
| 419 | MachineRegisterInfo &MRI = MF.getRegInfo(); | ||||||||
| 420 | return MRI.createVirtualRegister(ST.isWave32() ? &AMDGPU::SReg_32RegClass | ||||||||
| 421 | : &AMDGPU::SReg_64RegClass); | ||||||||
| 422 | } | ||||||||
| 423 | |||||||||
| 424 | static unsigned insertUndefLaneMask(MachineBasicBlock &MBB) { | ||||||||
| 425 | MachineFunction &MF = *MBB.getParent(); | ||||||||
| 426 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); | ||||||||
| 427 | const SIInstrInfo *TII = ST.getInstrInfo(); | ||||||||
| 428 | unsigned UndefReg = createLaneMaskReg(MF); | ||||||||
| 429 | BuildMI(MBB, MBB.getFirstTerminator(), {}, TII->get(AMDGPU::IMPLICIT_DEF), | ||||||||
| 430 | UndefReg); | ||||||||
| 431 | return UndefReg; | ||||||||
| 432 | } | ||||||||
| 433 | |||||||||
| 434 | /// Lower all instructions that def or use vreg_1 registers. | ||||||||
| 435 | /// | ||||||||
| 436 | /// In a first pass, we lower COPYs from vreg_1 to vector registers, as can | ||||||||
| 437 | /// occur around inline assembly. We do this first, before vreg_1 registers | ||||||||
| 438 | /// are changed to scalar mask registers. | ||||||||
| 439 | /// | ||||||||
| 440 | /// Then we lower all defs of vreg_1 registers. Phi nodes are lowered before | ||||||||
| 441 | /// all others, because phi lowering looks through copies and can therefore | ||||||||
| 442 | /// often make copy lowering unnecessary. | ||||||||
| 443 | bool SILowerI1Copies::runOnMachineFunction(MachineFunction &TheMF) { | ||||||||
| 444 | // Only need to run this in SelectionDAG path. | ||||||||
| 445 | if (TheMF.getProperties().hasProperty( | ||||||||
| |||||||||
| 446 | MachineFunctionProperties::Property::Selected)) | ||||||||
| 447 | return false; | ||||||||
| 448 | |||||||||
| 449 | MF = &TheMF; | ||||||||
| 450 | MRI = &MF->getRegInfo(); | ||||||||
| 451 | DT = &getAnalysis<MachineDominatorTree>(); | ||||||||
| 452 | PDT = &getAnalysis<MachinePostDominatorTree>(); | ||||||||
| 453 | |||||||||
| 454 | ST = &MF->getSubtarget<GCNSubtarget>(); | ||||||||
| 455 | TII = ST->getInstrInfo(); | ||||||||
| 456 | IsWave32 = ST->isWave32(); | ||||||||
| 457 | |||||||||
| 458 | if (IsWave32
| ||||||||
| 459 | ExecReg = AMDGPU::EXEC_LO; | ||||||||
| 460 | MovOp = AMDGPU::S_MOV_B32; | ||||||||
| 461 | AndOp = AMDGPU::S_AND_B32; | ||||||||
| 462 | OrOp = AMDGPU::S_OR_B32; | ||||||||
| 463 | XorOp = AMDGPU::S_XOR_B32; | ||||||||
| 464 | AndN2Op = AMDGPU::S_ANDN2_B32; | ||||||||
| 465 | OrN2Op = AMDGPU::S_ORN2_B32; | ||||||||
| 466 | } else { | ||||||||
| 467 | ExecReg = AMDGPU::EXEC; | ||||||||
| 468 | MovOp = AMDGPU::S_MOV_B64; | ||||||||
| 469 | AndOp = AMDGPU::S_AND_B64; | ||||||||
| 470 | OrOp = AMDGPU::S_OR_B64; | ||||||||
| 471 | XorOp = AMDGPU::S_XOR_B64; | ||||||||
| 472 | AndN2Op = AMDGPU::S_ANDN2_B64; | ||||||||
| 473 | OrN2Op = AMDGPU::S_ORN2_B64; | ||||||||
| 474 | } | ||||||||
| 475 | |||||||||
| 476 | lowerCopiesFromI1(); | ||||||||
| 477 | lowerPhis(); | ||||||||
| 478 | lowerCopiesToI1(); | ||||||||
| 479 | |||||||||
| 480 | for (unsigned Reg : ConstrainRegs) | ||||||||
| 481 | MRI->constrainRegClass(Reg, &AMDGPU::SReg_1_XEXECRegClass); | ||||||||
| 482 | ConstrainRegs.clear(); | ||||||||
| 483 | |||||||||
| 484 | return true; | ||||||||
| 485 | } | ||||||||
| 486 | |||||||||
| 487 | #ifndef NDEBUG1 | ||||||||
| 488 | static bool isVRegCompatibleReg(const SIRegisterInfo &TRI, | ||||||||
| 489 | const MachineRegisterInfo &MRI, | ||||||||
| 490 | Register Reg) { | ||||||||
| 491 | unsigned Size = TRI.getRegSizeInBits(Reg, MRI); | ||||||||
| 492 | return Size == 1 || Size == 32; | ||||||||
| 493 | } | ||||||||
| 494 | #endif | ||||||||
| 495 | |||||||||
| 496 | void SILowerI1Copies::lowerCopiesFromI1() { | ||||||||
| 497 | SmallVector<MachineInstr *, 4> DeadCopies; | ||||||||
| 498 | |||||||||
| 499 | for (MachineBasicBlock &MBB : *MF) { | ||||||||
| 500 | for (MachineInstr &MI : MBB) { | ||||||||
| 501 | if (MI.getOpcode() != AMDGPU::COPY) | ||||||||
| 502 | continue; | ||||||||
| 503 | |||||||||
| 504 | Register DstReg = MI.getOperand(0).getReg(); | ||||||||
| 505 | Register SrcReg = MI.getOperand(1).getReg(); | ||||||||
| 506 | if (!isVreg1(SrcReg)) | ||||||||
| 507 | continue; | ||||||||
| 508 | |||||||||
| 509 | if (isLaneMaskReg(DstReg) || isVreg1(DstReg)) | ||||||||
| 510 | continue; | ||||||||
| 511 | |||||||||
| 512 | // Copy into a 32-bit vector register. | ||||||||
| 513 | LLVM_DEBUG(dbgs() << "Lower copy from i1: " << MI)do { } while (false); | ||||||||
| 514 | DebugLoc DL = MI.getDebugLoc(); | ||||||||
| 515 | |||||||||
| 516 | assert(isVRegCompatibleReg(TII->getRegisterInfo(), *MRI, DstReg))((void)0); | ||||||||
| 517 | assert(!MI.getOperand(0).getSubReg())((void)0); | ||||||||
| 518 | |||||||||
| 519 | ConstrainRegs.insert(SrcReg); | ||||||||
| 520 | BuildMI(MBB, MI, DL, TII->get(AMDGPU::V_CNDMASK_B32_e64), DstReg) | ||||||||
| 521 | .addImm(0) | ||||||||
| 522 | .addImm(0) | ||||||||
| 523 | .addImm(0) | ||||||||
| 524 | .addImm(-1) | ||||||||
| 525 | .addReg(SrcReg); | ||||||||
| 526 | DeadCopies.push_back(&MI); | ||||||||
| 527 | } | ||||||||
| 528 | |||||||||
| 529 | for (MachineInstr *MI : DeadCopies) | ||||||||
| 530 | MI->eraseFromParent(); | ||||||||
| 531 | DeadCopies.clear(); | ||||||||
| 532 | } | ||||||||
| 533 | } | ||||||||
| 534 | |||||||||
| 535 | void SILowerI1Copies::lowerPhis() { | ||||||||
| 536 | MachineSSAUpdater SSAUpdater(*MF); | ||||||||
| 537 | LoopFinder LF(*DT, *PDT); | ||||||||
| 538 | PhiIncomingAnalysis PIA(*PDT); | ||||||||
| 539 | SmallVector<MachineInstr *, 4> Vreg1Phis; | ||||||||
| 540 | SmallVector<MachineBasicBlock *, 4> IncomingBlocks; | ||||||||
| 541 | SmallVector<unsigned, 4> IncomingRegs; | ||||||||
| 542 | SmallVector<unsigned, 4> IncomingUpdated; | ||||||||
| 543 | #ifndef NDEBUG1 | ||||||||
| 544 | DenseSet<unsigned> PhiRegisters; | ||||||||
| 545 | #endif | ||||||||
| 546 | |||||||||
| 547 | for (MachineBasicBlock &MBB : *MF) { | ||||||||
| 548 | for (MachineInstr &MI : MBB.phis()) { | ||||||||
| 549 | if (isVreg1(MI.getOperand(0).getReg())) | ||||||||
| 550 | Vreg1Phis.push_back(&MI); | ||||||||
| 551 | } | ||||||||
| 552 | } | ||||||||
| 553 | |||||||||
| 554 | MachineBasicBlock *PrevMBB = nullptr; | ||||||||
| 555 | for (MachineInstr *MI : Vreg1Phis) { | ||||||||
| 556 | MachineBasicBlock &MBB = *MI->getParent(); | ||||||||
| 557 | if (&MBB != PrevMBB) { | ||||||||
| 558 | LF.initialize(MBB); | ||||||||
| 559 | PrevMBB = &MBB; | ||||||||
| 560 | } | ||||||||
| 561 | |||||||||
| 562 | LLVM_DEBUG(dbgs() << "Lower PHI: " << *MI)do { } while (false); | ||||||||
| 563 | |||||||||
| 564 | Register DstReg = MI->getOperand(0).getReg(); | ||||||||
| 565 | MRI->setRegClass(DstReg, IsWave32
| ||||||||
| 566 | : &AMDGPU::SReg_64RegClass); | ||||||||
| 567 | |||||||||
| 568 | // Collect incoming values. | ||||||||
| 569 | for (unsigned i = 1; i < MI->getNumOperands(); i += 2) { | ||||||||
| 570 | assert(i + 1 < MI->getNumOperands())((void)0); | ||||||||
| 571 | Register IncomingReg = MI->getOperand(i).getReg(); | ||||||||
| 572 | MachineBasicBlock *IncomingMBB = MI->getOperand(i + 1).getMBB(); | ||||||||
| 573 | MachineInstr *IncomingDef = MRI->getUniqueVRegDef(IncomingReg); | ||||||||
| 574 | |||||||||
| 575 | if (IncomingDef->getOpcode() == AMDGPU::COPY) { | ||||||||
| 576 | IncomingReg = IncomingDef->getOperand(1).getReg(); | ||||||||
| 577 | assert(isLaneMaskReg(IncomingReg) || isVreg1(IncomingReg))((void)0); | ||||||||
| 578 | assert(!IncomingDef->getOperand(1).getSubReg())((void)0); | ||||||||
| 579 | } else if (IncomingDef->getOpcode() == AMDGPU::IMPLICIT_DEF) { | ||||||||
| 580 | continue; | ||||||||
| 581 | } else { | ||||||||
| 582 | assert(IncomingDef->isPHI() || PhiRegisters.count(IncomingReg))((void)0); | ||||||||
| 583 | } | ||||||||
| 584 | |||||||||
| 585 | IncomingBlocks.push_back(IncomingMBB); | ||||||||
| 586 | IncomingRegs.push_back(IncomingReg); | ||||||||
| 587 | } | ||||||||
| 588 | |||||||||
| 589 | #ifndef NDEBUG1 | ||||||||
| 590 | PhiRegisters.insert(DstReg); | ||||||||
| 591 | #endif | ||||||||
| 592 | |||||||||
| 593 | // Phis in a loop that are observed outside the loop receive a simple but | ||||||||
| 594 | // conservatively correct treatment. | ||||||||
| 595 | std::vector<MachineBasicBlock *> DomBlocks = {&MBB}; | ||||||||
| 596 | for (MachineInstr &Use : MRI->use_instructions(DstReg)) | ||||||||
| 597 | DomBlocks.push_back(Use.getParent()); | ||||||||
| 598 | |||||||||
| 599 | MachineBasicBlock *PostDomBound = | ||||||||
| 600 | PDT->findNearestCommonDominator(DomBlocks); | ||||||||
| 601 | |||||||||
| 602 | // FIXME: This fails to find irreducible cycles. If we have a def (other | ||||||||
| 603 | // than a constant) in a pair of blocks that end up looping back to each | ||||||||
| 604 | // other, it will be mishandle. Due to structurization this shouldn't occur | ||||||||
| 605 | // in practice. | ||||||||
| 606 | unsigned FoundLoopLevel = LF.findLoop(PostDomBound); | ||||||||
| 607 | |||||||||
| 608 | SSAUpdater.Initialize(DstReg); | ||||||||
| 609 | |||||||||
| 610 | if (FoundLoopLevel
| ||||||||
| 611 | LF.addLoopEntries(FoundLoopLevel, SSAUpdater, IncomingBlocks); | ||||||||
| 612 | |||||||||
| 613 | for (unsigned i = 0; i < IncomingRegs.size(); ++i) { | ||||||||
| 614 | IncomingUpdated.push_back(createLaneMaskReg(*MF)); | ||||||||
| 615 | SSAUpdater.AddAvailableValue(IncomingBlocks[i], | ||||||||
| 616 | IncomingUpdated.back()); | ||||||||
| 617 | } | ||||||||
| 618 | |||||||||
| 619 | for (unsigned i = 0; i < IncomingRegs.size(); ++i) { | ||||||||
| 620 | MachineBasicBlock &IMBB = *IncomingBlocks[i]; | ||||||||
| 621 | buildMergeLaneMasks( | ||||||||
| 622 | IMBB, getSaluInsertionAtEnd(IMBB), {}, IncomingUpdated[i], | ||||||||
| 623 | SSAUpdater.GetValueInMiddleOfBlock(&IMBB), IncomingRegs[i]); | ||||||||
| 624 | } | ||||||||
| 625 | } else { | ||||||||
| 626 | // The phi is not observed from outside a loop. Use a more accurate | ||||||||
| 627 | // lowering. | ||||||||
| 628 | PIA.analyze(MBB, IncomingBlocks); | ||||||||
| 629 | |||||||||
| 630 | for (MachineBasicBlock *MBB : PIA.predecessors()) | ||||||||
| 631 | SSAUpdater.AddAvailableValue(MBB, insertUndefLaneMask(*MBB)); | ||||||||
| 632 | |||||||||
| 633 | for (unsigned i = 0; i < IncomingRegs.size(); ++i) { | ||||||||
| 634 | MachineBasicBlock &IMBB = *IncomingBlocks[i]; | ||||||||
| 635 | if (PIA.isSource(IMBB)) { | ||||||||
| 636 | IncomingUpdated.push_back(0); | ||||||||
| 637 | SSAUpdater.AddAvailableValue(&IMBB, IncomingRegs[i]); | ||||||||
| 638 | } else { | ||||||||
| 639 | IncomingUpdated.push_back(createLaneMaskReg(*MF)); | ||||||||
| 640 | SSAUpdater.AddAvailableValue(&IMBB, IncomingUpdated.back()); | ||||||||
| 641 | } | ||||||||
| 642 | } | ||||||||
| 643 | |||||||||
| 644 | for (unsigned i = 0; i < IncomingRegs.size(); ++i) { | ||||||||
| 645 | if (!IncomingUpdated[i]) | ||||||||
| 646 | continue; | ||||||||
| 647 | |||||||||
| 648 | MachineBasicBlock &IMBB = *IncomingBlocks[i]; | ||||||||
| 649 | buildMergeLaneMasks( | ||||||||
| 650 | IMBB, getSaluInsertionAtEnd(IMBB), {}, IncomingUpdated[i], | ||||||||
| 651 | SSAUpdater.GetValueInMiddleOfBlock(&IMBB), IncomingRegs[i]); | ||||||||
| 652 | } | ||||||||
| 653 | } | ||||||||
| 654 | |||||||||
| 655 | Register NewReg = SSAUpdater.GetValueInMiddleOfBlock(&MBB); | ||||||||
| 656 | if (NewReg != DstReg) { | ||||||||
| 657 | MRI->replaceRegWith(NewReg, DstReg); | ||||||||
| 658 | MI->eraseFromParent(); | ||||||||
| 659 | } | ||||||||
| 660 | |||||||||
| 661 | IncomingBlocks.clear(); | ||||||||
| 662 | IncomingRegs.clear(); | ||||||||
| 663 | IncomingUpdated.clear(); | ||||||||
| 664 | } | ||||||||
| 665 | } | ||||||||
| 666 | |||||||||
| 667 | void SILowerI1Copies::lowerCopiesToI1() { | ||||||||
| 668 | MachineSSAUpdater SSAUpdater(*MF); | ||||||||
| 669 | LoopFinder LF(*DT, *PDT); | ||||||||
| 670 | SmallVector<MachineInstr *, 4> DeadCopies; | ||||||||
| 671 | |||||||||
| 672 | for (MachineBasicBlock &MBB : *MF) { | ||||||||
| 673 | LF.initialize(MBB); | ||||||||
| 674 | |||||||||
| 675 | for (MachineInstr &MI : MBB) { | ||||||||
| 676 | if (MI.getOpcode() != AMDGPU::IMPLICIT_DEF && | ||||||||
| 677 | MI.getOpcode() != AMDGPU::COPY) | ||||||||
| 678 | continue; | ||||||||
| 679 | |||||||||
| 680 | Register DstReg = MI.getOperand(0).getReg(); | ||||||||
| 681 | if (!isVreg1(DstReg)) | ||||||||
| 682 | continue; | ||||||||
| 683 | |||||||||
| 684 | if (MRI->use_empty(DstReg)) { | ||||||||
| 685 | DeadCopies.push_back(&MI); | ||||||||
| 686 | continue; | ||||||||
| 687 | } | ||||||||
| 688 | |||||||||
| 689 | LLVM_DEBUG(dbgs() << "Lower Other: " << MI)do { } while (false); | ||||||||
| 690 | |||||||||
| 691 | MRI->setRegClass(DstReg, IsWave32 ? &AMDGPU::SReg_32RegClass | ||||||||
| 692 | : &AMDGPU::SReg_64RegClass); | ||||||||
| 693 | if (MI.getOpcode() == AMDGPU::IMPLICIT_DEF) | ||||||||
| 694 | continue; | ||||||||
| 695 | |||||||||
| 696 | DebugLoc DL = MI.getDebugLoc(); | ||||||||
| 697 | Register SrcReg = MI.getOperand(1).getReg(); | ||||||||
| 698 | assert(!MI.getOperand(1).getSubReg())((void)0); | ||||||||
| 699 | |||||||||
| 700 | if (!SrcReg.isVirtual() || (!isLaneMaskReg(SrcReg) && !isVreg1(SrcReg))) { | ||||||||
| 701 | assert(TII->getRegisterInfo().getRegSizeInBits(SrcReg, *MRI) == 32)((void)0); | ||||||||
| 702 | unsigned TmpReg = createLaneMaskReg(*MF); | ||||||||
| 703 | BuildMI(MBB, MI, DL, TII->get(AMDGPU::V_CMP_NE_U32_e64), TmpReg) | ||||||||
| 704 | .addReg(SrcReg) | ||||||||
| 705 | .addImm(0); | ||||||||
| 706 | MI.getOperand(1).setReg(TmpReg); | ||||||||
| 707 | SrcReg = TmpReg; | ||||||||
| 708 | } | ||||||||
| 709 | |||||||||
| 710 | // Defs in a loop that are observed outside the loop must be transformed | ||||||||
| 711 | // into appropriate bit manipulation. | ||||||||
| 712 | std::vector<MachineBasicBlock *> DomBlocks = {&MBB}; | ||||||||
| 713 | for (MachineInstr &Use : MRI->use_instructions(DstReg)) | ||||||||
| 714 | DomBlocks.push_back(Use.getParent()); | ||||||||
| 715 | |||||||||
| 716 | MachineBasicBlock *PostDomBound = | ||||||||
| 717 | PDT->findNearestCommonDominator(DomBlocks); | ||||||||
| 718 | unsigned FoundLoopLevel = LF.findLoop(PostDomBound); | ||||||||
| 719 | if (FoundLoopLevel) { | ||||||||
| 720 | SSAUpdater.Initialize(DstReg); | ||||||||
| 721 | SSAUpdater.AddAvailableValue(&MBB, DstReg); | ||||||||
| 722 | LF.addLoopEntries(FoundLoopLevel, SSAUpdater); | ||||||||
| 723 | |||||||||
| 724 | buildMergeLaneMasks(MBB, MI, DL, DstReg, | ||||||||
| 725 | SSAUpdater.GetValueInMiddleOfBlock(&MBB), SrcReg); | ||||||||
| 726 | DeadCopies.push_back(&MI); | ||||||||
| 727 | } | ||||||||
| 728 | } | ||||||||
| 729 | |||||||||
| 730 | for (MachineInstr *MI : DeadCopies) | ||||||||
| 731 | MI->eraseFromParent(); | ||||||||
| 732 | DeadCopies.clear(); | ||||||||
| 733 | } | ||||||||
| 734 | } | ||||||||
| 735 | |||||||||
| 736 | bool SILowerI1Copies::isConstantLaneMask(Register Reg, bool &Val) const { | ||||||||
| 737 | const MachineInstr *MI; | ||||||||
| 738 | for (;;) { | ||||||||
| 739 | MI = MRI->getUniqueVRegDef(Reg); | ||||||||
| 740 | if (MI->getOpcode() == AMDGPU::IMPLICIT_DEF) | ||||||||
| 741 | return true; | ||||||||
| 742 | |||||||||
| 743 | if (MI->getOpcode() != AMDGPU::COPY) | ||||||||
| 744 | break; | ||||||||
| 745 | |||||||||
| 746 | Reg = MI->getOperand(1).getReg(); | ||||||||
| 747 | if (!Reg.isVirtual()) | ||||||||
| 748 | return false; | ||||||||
| 749 | if (!isLaneMaskReg(Reg)) | ||||||||
| 750 | return false; | ||||||||
| 751 | } | ||||||||
| 752 | |||||||||
| 753 | if (MI->getOpcode() != MovOp) | ||||||||
| 754 | return false; | ||||||||
| 755 | |||||||||
| 756 | if (!MI->getOperand(1).isImm()) | ||||||||
| 757 | return false; | ||||||||
| 758 | |||||||||
| 759 | int64_t Imm = MI->getOperand(1).getImm(); | ||||||||
| 760 | if (Imm == 0) { | ||||||||
| 761 | Val = false; | ||||||||
| 762 | return true; | ||||||||
| 763 | } | ||||||||
| 764 | if (Imm == -1) { | ||||||||
| 765 | Val = true; | ||||||||
| 766 | return true; | ||||||||
| 767 | } | ||||||||
| 768 | |||||||||
| 769 | return false; | ||||||||
| 770 | } | ||||||||
| 771 | |||||||||
| 772 | static void instrDefsUsesSCC(const MachineInstr &MI, bool &Def, bool &Use) { | ||||||||
| 773 | Def = false; | ||||||||
| 774 | Use = false; | ||||||||
| 775 | |||||||||
| 776 | for (const MachineOperand &MO : MI.operands()) { | ||||||||
| 777 | if (MO.isReg() && MO.getReg() == AMDGPU::SCC) { | ||||||||
| 778 | if (MO.isUse()) | ||||||||
| 779 | Use = true; | ||||||||
| 780 | else | ||||||||
| 781 | Def = true; | ||||||||
| 782 | } | ||||||||
| 783 | } | ||||||||
| 784 | } | ||||||||
| 785 | |||||||||
| 786 | /// Return a point at the end of the given \p MBB to insert SALU instructions | ||||||||
| 787 | /// for lane mask calculation. Take terminators and SCC into account. | ||||||||
| 788 | MachineBasicBlock::iterator | ||||||||
| 789 | SILowerI1Copies::getSaluInsertionAtEnd(MachineBasicBlock &MBB) const { | ||||||||
| 790 | auto InsertionPt = MBB.getFirstTerminator(); | ||||||||
| 791 | bool TerminatorsUseSCC = false; | ||||||||
| 792 | for (auto I = InsertionPt, E = MBB.end(); I != E; ++I) { | ||||||||
| 793 | bool DefsSCC; | ||||||||
| 794 | instrDefsUsesSCC(*I, DefsSCC, TerminatorsUseSCC); | ||||||||
| 795 | if (TerminatorsUseSCC || DefsSCC) | ||||||||
| 796 | break; | ||||||||
| 797 | } | ||||||||
| 798 | |||||||||
| 799 | if (!TerminatorsUseSCC) | ||||||||
| 800 | return InsertionPt; | ||||||||
| 801 | |||||||||
| 802 | while (InsertionPt != MBB.begin()) { | ||||||||
| 803 | InsertionPt--; | ||||||||
| 804 | |||||||||
| 805 | bool DefSCC, UseSCC; | ||||||||
| 806 | instrDefsUsesSCC(*InsertionPt, DefSCC, UseSCC); | ||||||||
| 807 | if (DefSCC) | ||||||||
| 808 | return InsertionPt; | ||||||||
| 809 | } | ||||||||
| 810 | |||||||||
| 811 | // We should have at least seen an IMPLICIT_DEF or COPY | ||||||||
| 812 | llvm_unreachable("SCC used by terminator but no def in block")__builtin_unreachable(); | ||||||||
| 813 | } | ||||||||
| 814 | |||||||||
| 815 | void SILowerI1Copies::buildMergeLaneMasks(MachineBasicBlock &MBB, | ||||||||
| 816 | MachineBasicBlock::iterator I, | ||||||||
| 817 | const DebugLoc &DL, unsigned DstReg, | ||||||||
| 818 | unsigned PrevReg, unsigned CurReg) { | ||||||||
| 819 | bool PrevVal = false; | ||||||||
| 820 | bool PrevConstant = isConstantLaneMask(PrevReg, PrevVal); | ||||||||
| 821 | bool CurVal = false; | ||||||||
| 822 | bool CurConstant = isConstantLaneMask(CurReg, CurVal); | ||||||||
| 823 | |||||||||
| 824 | if (PrevConstant && CurConstant) { | ||||||||
| 825 | if (PrevVal == CurVal) { | ||||||||
| 826 | BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg).addReg(CurReg); | ||||||||
| 827 | } else if (CurVal) { | ||||||||
| 828 | BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg).addReg(ExecReg); | ||||||||
| 829 | } else { | ||||||||
| 830 | BuildMI(MBB, I, DL, TII->get(XorOp), DstReg) | ||||||||
| 831 | .addReg(ExecReg) | ||||||||
| 832 | .addImm(-1); | ||||||||
| 833 | } | ||||||||
| 834 | return; | ||||||||
| 835 | } | ||||||||
| 836 | |||||||||
| 837 | unsigned PrevMaskedReg = 0; | ||||||||
| 838 | unsigned CurMaskedReg = 0; | ||||||||
| 839 | if (!PrevConstant) { | ||||||||
| 840 | if (CurConstant && CurVal) { | ||||||||
| 841 | PrevMaskedReg = PrevReg; | ||||||||
| 842 | } else { | ||||||||
| 843 | PrevMaskedReg = createLaneMaskReg(*MF); | ||||||||
| 844 | BuildMI(MBB, I, DL, TII->get(AndN2Op), PrevMaskedReg) | ||||||||
| 845 | .addReg(PrevReg) | ||||||||
| 846 | .addReg(ExecReg); | ||||||||
| 847 | } | ||||||||
| 848 | } | ||||||||
| 849 | if (!CurConstant) { | ||||||||
| 850 | // TODO: check whether CurReg is already masked by EXEC | ||||||||
| 851 | if (PrevConstant && PrevVal) { | ||||||||
| 852 | CurMaskedReg = CurReg; | ||||||||
| 853 | } else { | ||||||||
| 854 | CurMaskedReg = createLaneMaskReg(*MF); | ||||||||
| 855 | BuildMI(MBB, I, DL, TII->get(AndOp), CurMaskedReg) | ||||||||
| 856 | .addReg(CurReg) | ||||||||
| 857 | .addReg(ExecReg); | ||||||||
| 858 | } | ||||||||
| 859 | } | ||||||||
| 860 | |||||||||
| 861 | if (PrevConstant && !PrevVal) { | ||||||||
| 862 | BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg) | ||||||||
| 863 | .addReg(CurMaskedReg); | ||||||||
| 864 | } else if (CurConstant && !CurVal) { | ||||||||
| 865 | BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg) | ||||||||
| 866 | .addReg(PrevMaskedReg); | ||||||||
| 867 | } else if (PrevConstant && PrevVal) { | ||||||||
| 868 | BuildMI(MBB, I, DL, TII->get(OrN2Op), DstReg) | ||||||||
| 869 | .addReg(CurMaskedReg) | ||||||||
| 870 | .addReg(ExecReg); | ||||||||
| 871 | } else { | ||||||||
| 872 | BuildMI(MBB, I, DL, TII->get(OrOp), DstReg) | ||||||||
| 873 | .addReg(PrevMaskedReg) | ||||||||
| 874 | .addReg(CurMaskedReg ? CurMaskedReg : ExecReg); | ||||||||
| 875 | } | ||||||||
| 876 | } |
| 1 | //==- llvm/CodeGen/MachineDominators.h - Machine Dom Calculation -*- 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 classes mirroring those in llvm/Analysis/Dominators.h, |
| 10 | // but for target-specific code rather than target-independent IR. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef LLVM_CODEGEN_MACHINEDOMINATORS_H |
| 15 | #define LLVM_CODEGEN_MACHINEDOMINATORS_H |
| 16 | |
| 17 | #include "llvm/ADT/SmallSet.h" |
| 18 | #include "llvm/ADT/SmallVector.h" |
| 19 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 20 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 21 | #include "llvm/CodeGen/MachineInstr.h" |
| 22 | #include "llvm/Support/GenericDomTree.h" |
| 23 | #include "llvm/Support/GenericDomTreeConstruction.h" |
| 24 | #include <cassert> |
| 25 | #include <memory> |
| 26 | |
| 27 | namespace llvm { |
| 28 | |
| 29 | template <> |
| 30 | inline void DominatorTreeBase<MachineBasicBlock, false>::addRoot( |
| 31 | MachineBasicBlock *MBB) { |
| 32 | this->Roots.push_back(MBB); |
| 33 | } |
| 34 | |
| 35 | extern template class DomTreeNodeBase<MachineBasicBlock>; |
| 36 | extern template class DominatorTreeBase<MachineBasicBlock, false>; // DomTree |
| 37 | extern template class DominatorTreeBase<MachineBasicBlock, true>; // PostDomTree |
| 38 | |
| 39 | using MachineDomTreeNode = DomTreeNodeBase<MachineBasicBlock>; |
| 40 | |
| 41 | //===------------------------------------- |
| 42 | /// DominatorTree Class - Concrete subclass of DominatorTreeBase that is used to |
| 43 | /// compute a normal dominator tree. |
| 44 | /// |
| 45 | class MachineDominatorTree : public MachineFunctionPass { |
| 46 | using DomTreeT = DomTreeBase<MachineBasicBlock>; |
| 47 | |
| 48 | /// Helper structure used to hold all the basic blocks |
| 49 | /// involved in the split of a critical edge. |
| 50 | struct CriticalEdge { |
| 51 | MachineBasicBlock *FromBB; |
| 52 | MachineBasicBlock *ToBB; |
| 53 | MachineBasicBlock *NewBB; |
| 54 | }; |
| 55 | |
| 56 | /// Pile up all the critical edges to be split. |
| 57 | /// The splitting of a critical edge is local and thus, it is possible |
| 58 | /// to apply several of those changes at the same time. |
| 59 | mutable SmallVector<CriticalEdge, 32> CriticalEdgesToSplit; |
| 60 | |
| 61 | /// Remember all the basic blocks that are inserted during |
| 62 | /// edge splitting. |
| 63 | /// Invariant: NewBBs == all the basic blocks contained in the NewBB |
| 64 | /// field of all the elements of CriticalEdgesToSplit. |
| 65 | /// I.e., forall elt in CriticalEdgesToSplit, it exists BB in NewBBs |
| 66 | /// such as BB == elt.NewBB. |
| 67 | mutable SmallSet<MachineBasicBlock *, 32> NewBBs; |
| 68 | |
| 69 | /// The DominatorTreeBase that is used to compute a normal dominator tree. |
| 70 | std::unique_ptr<DomTreeT> DT; |
| 71 | |
| 72 | /// Apply all the recorded critical edges to the DT. |
| 73 | /// This updates the underlying DT information in a way that uses |
| 74 | /// the fast query path of DT as much as possible. |
| 75 | /// |
| 76 | /// \post CriticalEdgesToSplit.empty(). |
| 77 | void applySplitCriticalEdges() const; |
| 78 | |
| 79 | public: |
| 80 | static char ID; // Pass ID, replacement for typeid |
| 81 | |
| 82 | MachineDominatorTree(); |
| 83 | explicit MachineDominatorTree(MachineFunction &MF) : MachineFunctionPass(ID) { |
| 84 | calculate(MF); |
| 85 | } |
| 86 | |
| 87 | DomTreeT &getBase() { |
| 88 | if (!DT) DT.reset(new DomTreeT()); |
| 89 | applySplitCriticalEdges(); |
| 90 | return *DT; |
| 91 | } |
| 92 | |
| 93 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
| 94 | |
| 95 | MachineBasicBlock *getRoot() const { |
| 96 | applySplitCriticalEdges(); |
| 97 | return DT->getRoot(); |
| 98 | } |
| 99 | |
| 100 | MachineDomTreeNode *getRootNode() const { |
| 101 | applySplitCriticalEdges(); |
| 102 | return DT->getRootNode(); |
| 103 | } |
| 104 | |
| 105 | bool runOnMachineFunction(MachineFunction &F) override; |
| 106 | |
| 107 | void calculate(MachineFunction &F); |
| 108 | |
| 109 | bool dominates(const MachineDomTreeNode *A, |
| 110 | const MachineDomTreeNode *B) const { |
| 111 | applySplitCriticalEdges(); |
| 112 | return DT->dominates(A, B); |
| 113 | } |
| 114 | |
| 115 | bool dominates(const MachineBasicBlock *A, const MachineBasicBlock *B) const { |
| 116 | applySplitCriticalEdges(); |
| 117 | return DT->dominates(A, B); |
| 118 | } |
| 119 | |
| 120 | // dominates - Return true if A dominates B. This performs the |
| 121 | // special checks necessary if A and B are in the same basic block. |
| 122 | bool dominates(const MachineInstr *A, const MachineInstr *B) const { |
| 123 | applySplitCriticalEdges(); |
| 124 | const MachineBasicBlock *BBA = A->getParent(), *BBB = B->getParent(); |
| 125 | if (BBA != BBB) return DT->dominates(BBA, BBB); |
| 126 | |
| 127 | // Loop through the basic block until we find A or B. |
| 128 | MachineBasicBlock::const_iterator I = BBA->begin(); |
| 129 | for (; &*I != A && &*I != B; ++I) |
| 130 | /*empty*/ ; |
| 131 | |
| 132 | return &*I == A; |
| 133 | } |
| 134 | |
| 135 | bool properlyDominates(const MachineDomTreeNode *A, |
| 136 | const MachineDomTreeNode *B) const { |
| 137 | applySplitCriticalEdges(); |
| 138 | return DT->properlyDominates(A, B); |
| 139 | } |
| 140 | |
| 141 | bool properlyDominates(const MachineBasicBlock *A, |
| 142 | const MachineBasicBlock *B) const { |
| 143 | applySplitCriticalEdges(); |
| 144 | return DT->properlyDominates(A, B); |
| 145 | } |
| 146 | |
| 147 | /// findNearestCommonDominator - Find nearest common dominator basic block |
| 148 | /// for basic block A and B. If there is no such block then return NULL. |
| 149 | MachineBasicBlock *findNearestCommonDominator(MachineBasicBlock *A, |
| 150 | MachineBasicBlock *B) { |
| 151 | applySplitCriticalEdges(); |
| 152 | return DT->findNearestCommonDominator(A, B); |
| 153 | } |
| 154 | |
| 155 | MachineDomTreeNode *operator[](MachineBasicBlock *BB) const { |
| 156 | applySplitCriticalEdges(); |
| 157 | return DT->getNode(BB); |
| 158 | } |
| 159 | |
| 160 | /// getNode - return the (Post)DominatorTree node for the specified basic |
| 161 | /// block. This is the same as using operator[] on this class. |
| 162 | /// |
| 163 | MachineDomTreeNode *getNode(MachineBasicBlock *BB) const { |
| 164 | applySplitCriticalEdges(); |
| 165 | return DT->getNode(BB); |
| 166 | } |
| 167 | |
| 168 | /// addNewBlock - Add a new node to the dominator tree information. This |
| 169 | /// creates a new node as a child of DomBB dominator node,linking it into |
| 170 | /// the children list of the immediate dominator. |
| 171 | MachineDomTreeNode *addNewBlock(MachineBasicBlock *BB, |
| 172 | MachineBasicBlock *DomBB) { |
| 173 | applySplitCriticalEdges(); |
| 174 | return DT->addNewBlock(BB, DomBB); |
| 175 | } |
| 176 | |
| 177 | /// changeImmediateDominator - This method is used to update the dominator |
| 178 | /// tree information when a node's immediate dominator changes. |
| 179 | /// |
| 180 | void changeImmediateDominator(MachineBasicBlock *N, |
| 181 | MachineBasicBlock *NewIDom) { |
| 182 | applySplitCriticalEdges(); |
| 183 | DT->changeImmediateDominator(N, NewIDom); |
| 184 | } |
| 185 | |
| 186 | void changeImmediateDominator(MachineDomTreeNode *N, |
| 187 | MachineDomTreeNode *NewIDom) { |
| 188 | applySplitCriticalEdges(); |
| 189 | DT->changeImmediateDominator(N, NewIDom); |
| 190 | } |
| 191 | |
| 192 | /// eraseNode - Removes a node from the dominator tree. Block must not |
| 193 | /// dominate any other blocks. Removes node from its immediate dominator's |
| 194 | /// children list. Deletes dominator node associated with basic block BB. |
| 195 | void eraseNode(MachineBasicBlock *BB) { |
| 196 | applySplitCriticalEdges(); |
| 197 | DT->eraseNode(BB); |
| 198 | } |
| 199 | |
| 200 | /// splitBlock - BB is split and now it has one successor. Update dominator |
| 201 | /// tree to reflect this change. |
| 202 | void splitBlock(MachineBasicBlock* NewBB) { |
| 203 | applySplitCriticalEdges(); |
| 204 | DT->splitBlock(NewBB); |
| 205 | } |
| 206 | |
| 207 | /// isReachableFromEntry - Return true if A is dominated by the entry |
| 208 | /// block of the function containing it. |
| 209 | bool isReachableFromEntry(const MachineBasicBlock *A) { |
| 210 | applySplitCriticalEdges(); |
| 211 | return DT->isReachableFromEntry(A); |
| 212 | } |
| 213 | |
| 214 | void releaseMemory() override; |
| 215 | |
| 216 | void verifyAnalysis() const override; |
| 217 | |
| 218 | void print(raw_ostream &OS, const Module*) const override; |
| 219 | |
| 220 | /// Record that the critical edge (FromBB, ToBB) has been |
| 221 | /// split with NewBB. |
| 222 | /// This is best to use this method instead of directly update the |
| 223 | /// underlying information, because this helps mitigating the |
| 224 | /// number of time the DT information is invalidated. |
| 225 | /// |
| 226 | /// \note Do not use this method with regular edges. |
| 227 | /// |
| 228 | /// \note To benefit from the compile time improvement incurred by this |
| 229 | /// method, the users of this method have to limit the queries to the DT |
| 230 | /// interface between two edges splitting. In other words, they have to |
| 231 | /// pack the splitting of critical edges as much as possible. |
| 232 | void recordSplitCriticalEdge(MachineBasicBlock *FromBB, |
| 233 | MachineBasicBlock *ToBB, |
| 234 | MachineBasicBlock *NewBB) { |
| 235 | bool Inserted = NewBBs.insert(NewBB).second; |
| 236 | (void)Inserted; |
| 237 | assert(Inserted &&((void)0) |
| 238 | "A basic block inserted via edge splitting cannot appear twice")((void)0); |
| 239 | CriticalEdgesToSplit.push_back({FromBB, ToBB, NewBB}); |
| 240 | } |
| 241 | }; |
| 242 | |
| 243 | //===------------------------------------- |
| 244 | /// DominatorTree GraphTraits specialization so the DominatorTree can be |
| 245 | /// iterable by generic graph iterators. |
| 246 | /// |
| 247 | |
| 248 | template <class Node, class ChildIterator> |
| 249 | struct MachineDomTreeGraphTraitsBase { |
| 250 | using NodeRef = Node *; |
| 251 | using ChildIteratorType = ChildIterator; |
| 252 | |
| 253 | static NodeRef getEntryNode(NodeRef N) { return N; } |
| 254 | static ChildIteratorType child_begin(NodeRef N) { return N->begin(); } |
| 255 | static ChildIteratorType child_end(NodeRef N) { return N->end(); } |
| 256 | }; |
| 257 | |
| 258 | template <class T> struct GraphTraits; |
| 259 | |
| 260 | template <> |
| 261 | struct GraphTraits<MachineDomTreeNode *> |
| 262 | : public MachineDomTreeGraphTraitsBase<MachineDomTreeNode, |
| 263 | MachineDomTreeNode::const_iterator> { |
| 264 | }; |
| 265 | |
| 266 | template <> |
| 267 | struct GraphTraits<const MachineDomTreeNode *> |
| 268 | : public MachineDomTreeGraphTraitsBase<const MachineDomTreeNode, |
| 269 | MachineDomTreeNode::const_iterator> { |
| 270 | }; |
| 271 | |
| 272 | template <> struct GraphTraits<MachineDominatorTree*> |
| 273 | : public GraphTraits<MachineDomTreeNode *> { |
| 274 | static NodeRef getEntryNode(MachineDominatorTree *DT) { |
| 275 | return DT->getRootNode(); |
| 276 | } |
| 277 | }; |
| 278 | |
| 279 | } // end namespace llvm |
| 280 | |
| 281 | #endif // LLVM_CODEGEN_MACHINEDOMINATORS_H |
| 1 | //===- GenericDomTree.h - Generic dominator trees for graphs ----*- 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 | /// \file | |||
| 9 | /// | |||
| 10 | /// This file defines a set of templates that efficiently compute a dominator | |||
| 11 | /// tree over a generic graph. This is used typically in LLVM for fast | |||
| 12 | /// dominance queries on the CFG, but is fully generic w.r.t. the underlying | |||
| 13 | /// graph types. | |||
| 14 | /// | |||
| 15 | /// Unlike ADT/* graph algorithms, generic dominator tree has more requirements | |||
| 16 | /// on the graph's NodeRef. The NodeRef should be a pointer and, | |||
| 17 | /// NodeRef->getParent() must return the parent node that is also a pointer. | |||
| 18 | /// | |||
| 19 | /// FIXME: Maybe GenericDomTree needs a TreeTraits, instead of GraphTraits. | |||
| 20 | /// | |||
| 21 | //===----------------------------------------------------------------------===// | |||
| 22 | ||||
| 23 | #ifndef LLVM_SUPPORT_GENERICDOMTREE_H | |||
| 24 | #define LLVM_SUPPORT_GENERICDOMTREE_H | |||
| 25 | ||||
| 26 | #include "llvm/ADT/DenseMap.h" | |||
| 27 | #include "llvm/ADT/GraphTraits.h" | |||
| 28 | #include "llvm/ADT/STLExtras.h" | |||
| 29 | #include "llvm/ADT/SmallPtrSet.h" | |||
| 30 | #include "llvm/ADT/SmallVector.h" | |||
| 31 | #include "llvm/Support/CFGDiff.h" | |||
| 32 | #include "llvm/Support/CFGUpdate.h" | |||
| 33 | #include "llvm/Support/raw_ostream.h" | |||
| 34 | #include <algorithm> | |||
| 35 | #include <cassert> | |||
| 36 | #include <cstddef> | |||
| 37 | #include <iterator> | |||
| 38 | #include <memory> | |||
| 39 | #include <type_traits> | |||
| 40 | #include <utility> | |||
| 41 | ||||
| 42 | namespace llvm { | |||
| 43 | ||||
| 44 | template <typename NodeT, bool IsPostDom> | |||
| 45 | class DominatorTreeBase; | |||
| 46 | ||||
| 47 | namespace DomTreeBuilder { | |||
| 48 | template <typename DomTreeT> | |||
| 49 | struct SemiNCAInfo; | |||
| 50 | } // namespace DomTreeBuilder | |||
| 51 | ||||
| 52 | /// Base class for the actual dominator tree node. | |||
| 53 | template <class NodeT> class DomTreeNodeBase { | |||
| 54 | friend class PostDominatorTree; | |||
| 55 | friend class DominatorTreeBase<NodeT, false>; | |||
| 56 | friend class DominatorTreeBase<NodeT, true>; | |||
| 57 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, false>>; | |||
| 58 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase<NodeT, true>>; | |||
| 59 | ||||
| 60 | NodeT *TheBB; | |||
| 61 | DomTreeNodeBase *IDom; | |||
| 62 | unsigned Level; | |||
| 63 | SmallVector<DomTreeNodeBase *, 4> Children; | |||
| 64 | mutable unsigned DFSNumIn = ~0; | |||
| 65 | mutable unsigned DFSNumOut = ~0; | |||
| 66 | ||||
| 67 | public: | |||
| 68 | DomTreeNodeBase(NodeT *BB, DomTreeNodeBase *iDom) | |||
| 69 | : TheBB(BB), IDom(iDom), Level(IDom ? IDom->Level + 1 : 0) {} | |||
| 70 | ||||
| 71 | using iterator = typename SmallVector<DomTreeNodeBase *, 4>::iterator; | |||
| 72 | using const_iterator = | |||
| 73 | typename SmallVector<DomTreeNodeBase *, 4>::const_iterator; | |||
| 74 | ||||
| 75 | iterator begin() { return Children.begin(); } | |||
| 76 | iterator end() { return Children.end(); } | |||
| 77 | const_iterator begin() const { return Children.begin(); } | |||
| 78 | const_iterator end() const { return Children.end(); } | |||
| 79 | ||||
| 80 | DomTreeNodeBase *const &back() const { return Children.back(); } | |||
| 81 | DomTreeNodeBase *&back() { return Children.back(); } | |||
| 82 | ||||
| 83 | iterator_range<iterator> children() { return make_range(begin(), end()); } | |||
| 84 | iterator_range<const_iterator> children() const { | |||
| 85 | return make_range(begin(), end()); | |||
| 86 | } | |||
| 87 | ||||
| 88 | NodeT *getBlock() const { return TheBB; } | |||
| 89 | DomTreeNodeBase *getIDom() const { return IDom; } | |||
| 90 | unsigned getLevel() const { return Level; } | |||
| 91 | ||||
| 92 | std::unique_ptr<DomTreeNodeBase> addChild( | |||
| 93 | std::unique_ptr<DomTreeNodeBase> C) { | |||
| 94 | Children.push_back(C.get()); | |||
| 95 | return C; | |||
| 96 | } | |||
| 97 | ||||
| 98 | bool isLeaf() const { return Children.empty(); } | |||
| 99 | size_t getNumChildren() const { return Children.size(); } | |||
| 100 | ||||
| 101 | void clearAllChildren() { Children.clear(); } | |||
| 102 | ||||
| 103 | bool compare(const DomTreeNodeBase *Other) const { | |||
| 104 | if (getNumChildren() != Other->getNumChildren()) | |||
| 105 | return true; | |||
| 106 | ||||
| 107 | if (Level != Other->Level) return true; | |||
| 108 | ||||
| 109 | SmallPtrSet<const NodeT *, 4> OtherChildren; | |||
| 110 | for (const DomTreeNodeBase *I : *Other) { | |||
| 111 | const NodeT *Nd = I->getBlock(); | |||
| 112 | OtherChildren.insert(Nd); | |||
| 113 | } | |||
| 114 | ||||
| 115 | for (const DomTreeNodeBase *I : *this) { | |||
| 116 | const NodeT *N = I->getBlock(); | |||
| 117 | if (OtherChildren.count(N) == 0) | |||
| 118 | return true; | |||
| 119 | } | |||
| 120 | return false; | |||
| 121 | } | |||
| 122 | ||||
| 123 | void setIDom(DomTreeNodeBase *NewIDom) { | |||
| 124 | assert(IDom && "No immediate dominator?")((void)0); | |||
| 125 | if (IDom == NewIDom) return; | |||
| 126 | ||||
| 127 | auto I = find(IDom->Children, this); | |||
| 128 | assert(I != IDom->Children.end() &&((void)0) | |||
| 129 | "Not in immediate dominator children set!")((void)0); | |||
| 130 | // I am no longer your child... | |||
| 131 | IDom->Children.erase(I); | |||
| 132 | ||||
| 133 | // Switch to new dominator | |||
| 134 | IDom = NewIDom; | |||
| 135 | IDom->Children.push_back(this); | |||
| 136 | ||||
| 137 | UpdateLevel(); | |||
| 138 | } | |||
| 139 | ||||
| 140 | /// getDFSNumIn/getDFSNumOut - These return the DFS visitation order for nodes | |||
| 141 | /// in the dominator tree. They are only guaranteed valid if | |||
| 142 | /// updateDFSNumbers() has been called. | |||
| 143 | unsigned getDFSNumIn() const { return DFSNumIn; } | |||
| 144 | unsigned getDFSNumOut() const { return DFSNumOut; } | |||
| 145 | ||||
| 146 | private: | |||
| 147 | // Return true if this node is dominated by other. Use this only if DFS info | |||
| 148 | // is valid. | |||
| 149 | bool DominatedBy(const DomTreeNodeBase *other) const { | |||
| 150 | return this->DFSNumIn >= other->DFSNumIn && | |||
| 151 | this->DFSNumOut <= other->DFSNumOut; | |||
| 152 | } | |||
| 153 | ||||
| 154 | void UpdateLevel() { | |||
| 155 | assert(IDom)((void)0); | |||
| 156 | if (Level == IDom->Level + 1) return; | |||
| 157 | ||||
| 158 | SmallVector<DomTreeNodeBase *, 64> WorkStack = {this}; | |||
| 159 | ||||
| 160 | while (!WorkStack.empty()) { | |||
| 161 | DomTreeNodeBase *Current = WorkStack.pop_back_val(); | |||
| 162 | Current->Level = Current->IDom->Level + 1; | |||
| 163 | ||||
| 164 | for (DomTreeNodeBase *C : *Current) { | |||
| 165 | assert(C->IDom)((void)0); | |||
| 166 | if (C->Level != C->IDom->Level + 1) WorkStack.push_back(C); | |||
| 167 | } | |||
| 168 | } | |||
| 169 | } | |||
| 170 | }; | |||
| 171 | ||||
| 172 | template <class NodeT> | |||
| 173 | raw_ostream &operator<<(raw_ostream &O, const DomTreeNodeBase<NodeT> *Node) { | |||
| 174 | if (Node->getBlock()) | |||
| 175 | Node->getBlock()->printAsOperand(O, false); | |||
| 176 | else | |||
| 177 | O << " <<exit node>>"; | |||
| 178 | ||||
| 179 | O << " {" << Node->getDFSNumIn() << "," << Node->getDFSNumOut() << "} [" | |||
| 180 | << Node->getLevel() << "]\n"; | |||
| 181 | ||||
| 182 | return O; | |||
| 183 | } | |||
| 184 | ||||
| 185 | template <class NodeT> | |||
| 186 | void PrintDomTree(const DomTreeNodeBase<NodeT> *N, raw_ostream &O, | |||
| 187 | unsigned Lev) { | |||
| 188 | O.indent(2 * Lev) << "[" << Lev << "] " << N; | |||
| 189 | for (typename DomTreeNodeBase<NodeT>::const_iterator I = N->begin(), | |||
| 190 | E = N->end(); | |||
| 191 | I != E; ++I) | |||
| 192 | PrintDomTree<NodeT>(*I, O, Lev + 1); | |||
| 193 | } | |||
| 194 | ||||
| 195 | namespace DomTreeBuilder { | |||
| 196 | // The routines below are provided in a separate header but referenced here. | |||
| 197 | template <typename DomTreeT> | |||
| 198 | void Calculate(DomTreeT &DT); | |||
| 199 | ||||
| 200 | template <typename DomTreeT> | |||
| 201 | void CalculateWithUpdates(DomTreeT &DT, | |||
| 202 | ArrayRef<typename DomTreeT::UpdateType> Updates); | |||
| 203 | ||||
| 204 | template <typename DomTreeT> | |||
| 205 | void InsertEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, | |||
| 206 | typename DomTreeT::NodePtr To); | |||
| 207 | ||||
| 208 | template <typename DomTreeT> | |||
| 209 | void DeleteEdge(DomTreeT &DT, typename DomTreeT::NodePtr From, | |||
| 210 | typename DomTreeT::NodePtr To); | |||
| 211 | ||||
| 212 | template <typename DomTreeT> | |||
| 213 | void ApplyUpdates(DomTreeT &DT, | |||
| 214 | GraphDiff<typename DomTreeT::NodePtr, | |||
| 215 | DomTreeT::IsPostDominator> &PreViewCFG, | |||
| 216 | GraphDiff<typename DomTreeT::NodePtr, | |||
| 217 | DomTreeT::IsPostDominator> *PostViewCFG); | |||
| 218 | ||||
| 219 | template <typename DomTreeT> | |||
| 220 | bool Verify(const DomTreeT &DT, typename DomTreeT::VerificationLevel VL); | |||
| 221 | } // namespace DomTreeBuilder | |||
| 222 | ||||
| 223 | /// Core dominator tree base class. | |||
| 224 | /// | |||
| 225 | /// This class is a generic template over graph nodes. It is instantiated for | |||
| 226 | /// various graphs in the LLVM IR or in the code generator. | |||
| 227 | template <typename NodeT, bool IsPostDom> | |||
| 228 | class DominatorTreeBase { | |||
| 229 | public: | |||
| 230 | static_assert(std::is_pointer<typename GraphTraits<NodeT *>::NodeRef>::value, | |||
| 231 | "Currently DominatorTreeBase supports only pointer nodes"); | |||
| 232 | using NodeType = NodeT; | |||
| 233 | using NodePtr = NodeT *; | |||
| 234 | using ParentPtr = decltype(std::declval<NodeT *>()->getParent()); | |||
| 235 | static_assert(std::is_pointer<ParentPtr>::value, | |||
| 236 | "Currently NodeT's parent must be a pointer type"); | |||
| 237 | using ParentType = std::remove_pointer_t<ParentPtr>; | |||
| 238 | static constexpr bool IsPostDominator = IsPostDom; | |||
| 239 | ||||
| 240 | using UpdateType = cfg::Update<NodePtr>; | |||
| 241 | using UpdateKind = cfg::UpdateKind; | |||
| 242 | static constexpr UpdateKind Insert = UpdateKind::Insert; | |||
| 243 | static constexpr UpdateKind Delete = UpdateKind::Delete; | |||
| 244 | ||||
| 245 | enum class VerificationLevel { Fast, Basic, Full }; | |||
| 246 | ||||
| 247 | protected: | |||
| 248 | // Dominators always have a single root, postdominators can have more. | |||
| 249 | SmallVector<NodeT *, IsPostDom ? 4 : 1> Roots; | |||
| 250 | ||||
| 251 | using DomTreeNodeMapType = | |||
| 252 | DenseMap<NodeT *, std::unique_ptr<DomTreeNodeBase<NodeT>>>; | |||
| 253 | DomTreeNodeMapType DomTreeNodes; | |||
| 254 | DomTreeNodeBase<NodeT> *RootNode = nullptr; | |||
| 255 | ParentPtr Parent = nullptr; | |||
| 256 | ||||
| 257 | mutable bool DFSInfoValid = false; | |||
| 258 | mutable unsigned int SlowQueries = 0; | |||
| 259 | ||||
| 260 | friend struct DomTreeBuilder::SemiNCAInfo<DominatorTreeBase>; | |||
| 261 | ||||
| 262 | public: | |||
| 263 | DominatorTreeBase() {} | |||
| 264 | ||||
| 265 | DominatorTreeBase(DominatorTreeBase &&Arg) | |||
| 266 | : Roots(std::move(Arg.Roots)), | |||
| 267 | DomTreeNodes(std::move(Arg.DomTreeNodes)), | |||
| 268 | RootNode(Arg.RootNode), | |||
| 269 | Parent(Arg.Parent), | |||
| 270 | DFSInfoValid(Arg.DFSInfoValid), | |||
| 271 | SlowQueries(Arg.SlowQueries) { | |||
| 272 | Arg.wipe(); | |||
| 273 | } | |||
| 274 | ||||
| 275 | DominatorTreeBase &operator=(DominatorTreeBase &&RHS) { | |||
| 276 | Roots = std::move(RHS.Roots); | |||
| 277 | DomTreeNodes = std::move(RHS.DomTreeNodes); | |||
| 278 | RootNode = RHS.RootNode; | |||
| 279 | Parent = RHS.Parent; | |||
| 280 | DFSInfoValid = RHS.DFSInfoValid; | |||
| 281 | SlowQueries = RHS.SlowQueries; | |||
| 282 | RHS.wipe(); | |||
| 283 | return *this; | |||
| 284 | } | |||
| 285 | ||||
| 286 | DominatorTreeBase(const DominatorTreeBase &) = delete; | |||
| 287 | DominatorTreeBase &operator=(const DominatorTreeBase &) = delete; | |||
| 288 | ||||
| 289 | /// Iteration over roots. | |||
| 290 | /// | |||
| 291 | /// This may include multiple blocks if we are computing post dominators. | |||
| 292 | /// For forward dominators, this will always be a single block (the entry | |||
| 293 | /// block). | |||
| 294 | using root_iterator = typename SmallVectorImpl<NodeT *>::iterator; | |||
| 295 | using const_root_iterator = typename SmallVectorImpl<NodeT *>::const_iterator; | |||
| 296 | ||||
| 297 | root_iterator root_begin() { return Roots.begin(); } | |||
| 298 | const_root_iterator root_begin() const { return Roots.begin(); } | |||
| 299 | root_iterator root_end() { return Roots.end(); } | |||
| 300 | const_root_iterator root_end() const { return Roots.end(); } | |||
| 301 | ||||
| 302 | size_t root_size() const { return Roots.size(); } | |||
| 303 | ||||
| 304 | iterator_range<root_iterator> roots() { | |||
| 305 | return make_range(root_begin(), root_end()); | |||
| 306 | } | |||
| 307 | iterator_range<const_root_iterator> roots() const { | |||
| 308 | return make_range(root_begin(), root_end()); | |||
| 309 | } | |||
| 310 | ||||
| 311 | /// isPostDominator - Returns true if analysis based of postdoms | |||
| 312 | /// | |||
| 313 | bool isPostDominator() const { return IsPostDominator; } | |||
| 314 | ||||
| 315 | /// compare - Return false if the other dominator tree base matches this | |||
| 316 | /// dominator tree base. Otherwise return true. | |||
| 317 | bool compare(const DominatorTreeBase &Other) const { | |||
| 318 | if (Parent != Other.Parent) return true; | |||
| 319 | ||||
| 320 | if (Roots.size() != Other.Roots.size()) | |||
| 321 | return true; | |||
| 322 | ||||
| 323 | if (!std::is_permutation(Roots.begin(), Roots.end(), Other.Roots.begin())) | |||
| 324 | return true; | |||
| 325 | ||||
| 326 | const DomTreeNodeMapType &OtherDomTreeNodes = Other.DomTreeNodes; | |||
| 327 | if (DomTreeNodes.size() != OtherDomTreeNodes.size()) | |||
| 328 | return true; | |||
| 329 | ||||
| 330 | for (const auto &DomTreeNode : DomTreeNodes) { | |||
| 331 | NodeT *BB = DomTreeNode.first; | |||
| 332 | typename DomTreeNodeMapType::const_iterator OI = | |||
| 333 | OtherDomTreeNodes.find(BB); | |||
| 334 | if (OI == OtherDomTreeNodes.end()) | |||
| 335 | return true; | |||
| 336 | ||||
| 337 | DomTreeNodeBase<NodeT> &MyNd = *DomTreeNode.second; | |||
| 338 | DomTreeNodeBase<NodeT> &OtherNd = *OI->second; | |||
| 339 | ||||
| 340 | if (MyNd.compare(&OtherNd)) | |||
| 341 | return true; | |||
| 342 | } | |||
| 343 | ||||
| 344 | return false; | |||
| 345 | } | |||
| 346 | ||||
| 347 | /// getNode - return the (Post)DominatorTree node for the specified basic | |||
| 348 | /// block. This is the same as using operator[] on this class. The result | |||
| 349 | /// may (but is not required to) be null for a forward (backwards) | |||
| 350 | /// statically unreachable block. | |||
| 351 | DomTreeNodeBase<NodeT> *getNode(const NodeT *BB) const { | |||
| 352 | auto I = DomTreeNodes.find(BB); | |||
| 353 | if (I != DomTreeNodes.end()) | |||
| 354 | return I->second.get(); | |||
| 355 | return nullptr; | |||
| 356 | } | |||
| 357 | ||||
| 358 | /// See getNode. | |||
| 359 | DomTreeNodeBase<NodeT> *operator[](const NodeT *BB) const { | |||
| 360 | return getNode(BB); | |||
| 361 | } | |||
| 362 | ||||
| 363 | /// getRootNode - This returns the entry node for the CFG of the function. If | |||
| 364 | /// this tree represents the post-dominance relations for a function, however, | |||
| 365 | /// this root may be a node with the block == NULL. This is the case when | |||
| 366 | /// there are multiple exit nodes from a particular function. Consumers of | |||
| 367 | /// post-dominance information must be capable of dealing with this | |||
| 368 | /// possibility. | |||
| 369 | /// | |||
| 370 | DomTreeNodeBase<NodeT> *getRootNode() { return RootNode; } | |||
| 371 | const DomTreeNodeBase<NodeT> *getRootNode() const { return RootNode; } | |||
| 372 | ||||
| 373 | /// Get all nodes dominated by R, including R itself. | |||
| 374 | void getDescendants(NodeT *R, SmallVectorImpl<NodeT *> &Result) const { | |||
| 375 | Result.clear(); | |||
| 376 | const DomTreeNodeBase<NodeT> *RN = getNode(R); | |||
| 377 | if (!RN) | |||
| 378 | return; // If R is unreachable, it will not be present in the DOM tree. | |||
| 379 | SmallVector<const DomTreeNodeBase<NodeT> *, 8> WL; | |||
| 380 | WL.push_back(RN); | |||
| 381 | ||||
| 382 | while (!WL.empty()) { | |||
| 383 | const DomTreeNodeBase<NodeT> *N = WL.pop_back_val(); | |||
| 384 | Result.push_back(N->getBlock()); | |||
| 385 | WL.append(N->begin(), N->end()); | |||
| 386 | } | |||
| 387 | } | |||
| 388 | ||||
| 389 | /// properlyDominates - Returns true iff A dominates B and A != B. | |||
| 390 | /// Note that this is not a constant time operation! | |||
| 391 | /// | |||
| 392 | bool properlyDominates(const DomTreeNodeBase<NodeT> *A, | |||
| 393 | const DomTreeNodeBase<NodeT> *B) const { | |||
| 394 | if (!A || !B) | |||
| 395 | return false; | |||
| 396 | if (A == B) | |||
| 397 | return false; | |||
| 398 | return dominates(A, B); | |||
| 399 | } | |||
| 400 | ||||
| 401 | bool properlyDominates(const NodeT *A, const NodeT *B) const; | |||
| 402 | ||||
| 403 | /// isReachableFromEntry - Return true if A is dominated by the entry | |||
| 404 | /// block of the function containing it. | |||
| 405 | bool isReachableFromEntry(const NodeT *A) const { | |||
| 406 | assert(!this->isPostDominator() &&((void)0) | |||
| 407 | "This is not implemented for post dominators")((void)0); | |||
| 408 | return isReachableFromEntry(getNode(const_cast<NodeT *>(A))); | |||
| 409 | } | |||
| 410 | ||||
| 411 | bool isReachableFromEntry(const DomTreeNodeBase<NodeT> *A) const { return A; } | |||
| 412 | ||||
| 413 | /// dominates - Returns true iff A dominates B. Note that this is not a | |||
| 414 | /// constant time operation! | |||
| 415 | /// | |||
| 416 | bool dominates(const DomTreeNodeBase<NodeT> *A, | |||
| 417 | const DomTreeNodeBase<NodeT> *B) const { | |||
| 418 | // A node trivially dominates itself. | |||
| 419 | if (B == A) | |||
| 420 | return true; | |||
| 421 | ||||
| 422 | // An unreachable node is dominated by anything. | |||
| 423 | if (!isReachableFromEntry(B)) | |||
| 424 | return true; | |||
| 425 | ||||
| 426 | // And dominates nothing. | |||
| 427 | if (!isReachableFromEntry(A)) | |||
| 428 | return false; | |||
| 429 | ||||
| 430 | if (B->getIDom() == A) return true; | |||
| 431 | ||||
| 432 | if (A->getIDom() == B) return false; | |||
| 433 | ||||
| 434 | // A can only dominate B if it is higher in the tree. | |||
| 435 | if (A->getLevel() >= B->getLevel()) return false; | |||
| 436 | ||||
| 437 | // Compare the result of the tree walk and the dfs numbers, if expensive | |||
| 438 | // checks are enabled. | |||
| 439 | #ifdef EXPENSIVE_CHECKS | |||
| 440 | assert((!DFSInfoValid ||((void)0) | |||
| 441 | (dominatedBySlowTreeWalk(A, B) == B->DominatedBy(A))) &&((void)0) | |||
| 442 | "Tree walk disagrees with dfs numbers!")((void)0); | |||
| 443 | #endif | |||
| 444 | ||||
| 445 | if (DFSInfoValid) | |||
| 446 | return B->DominatedBy(A); | |||
| 447 | ||||
| 448 | // If we end up with too many slow queries, just update the | |||
| 449 | // DFS numbers on the theory that we are going to keep querying. | |||
| 450 | SlowQueries++; | |||
| 451 | if (SlowQueries > 32) { | |||
| 452 | updateDFSNumbers(); | |||
| 453 | return B->DominatedBy(A); | |||
| 454 | } | |||
| 455 | ||||
| 456 | return dominatedBySlowTreeWalk(A, B); | |||
| 457 | } | |||
| 458 | ||||
| 459 | bool dominates(const NodeT *A, const NodeT *B) const; | |||
| 460 | ||||
| 461 | NodeT *getRoot() const { | |||
| 462 | assert(this->Roots.size() == 1 && "Should always have entry node!")((void)0); | |||
| 463 | return this->Roots[0]; | |||
| 464 | } | |||
| 465 | ||||
| 466 | /// Find nearest common dominator basic block for basic block A and B. A and B | |||
| 467 | /// must have tree nodes. | |||
| 468 | NodeT *findNearestCommonDominator(NodeT *A, NodeT *B) const { | |||
| 469 | assert(A && B && "Pointers are not valid")((void)0); | |||
| 470 | assert(A->getParent() == B->getParent() &&((void)0) | |||
| 471 | "Two blocks are not in same function")((void)0); | |||
| 472 | ||||
| 473 | // If either A or B is a entry block then it is nearest common dominator | |||
| 474 | // (for forward-dominators). | |||
| 475 | if (!isPostDominator()) { | |||
| 476 | NodeT &Entry = A->getParent()->front(); | |||
| 477 | if (A == &Entry || B == &Entry) | |||
| 478 | return &Entry; | |||
| 479 | } | |||
| 480 | ||||
| 481 | DomTreeNodeBase<NodeT> *NodeA = getNode(A); | |||
| 482 | DomTreeNodeBase<NodeT> *NodeB = getNode(B); | |||
| 483 | assert(NodeA && "A must be in the tree")((void)0); | |||
| 484 | assert(NodeB && "B must be in the tree")((void)0); | |||
| 485 | ||||
| 486 | // Use level information to go up the tree until the levels match. Then | |||
| 487 | // continue going up til we arrive at the same node. | |||
| 488 | while (NodeA != NodeB) { | |||
| 489 | if (NodeA->getLevel() < NodeB->getLevel()) std::swap(NodeA, NodeB); | |||
| 490 | ||||
| 491 | NodeA = NodeA->IDom; | |||
| 492 | } | |||
| 493 | ||||
| 494 | return NodeA->getBlock(); | |||
| ||||
| 495 | } | |||
| 496 | ||||
| 497 | const NodeT *findNearestCommonDominator(const NodeT *A, | |||
| 498 | const NodeT *B) const { | |||
| 499 | // Cast away the const qualifiers here. This is ok since | |||
| 500 | // const is re-introduced on the return type. | |||
| 501 | return findNearestCommonDominator(const_cast<NodeT *>(A), | |||
| 502 | const_cast<NodeT *>(B)); | |||
| 503 | } | |||
| 504 | ||||
| 505 | bool isVirtualRoot(const DomTreeNodeBase<NodeT> *A) const { | |||
| 506 | return isPostDominator() && !A->getBlock(); | |||
| 507 | } | |||
| 508 | ||||
| 509 | //===--------------------------------------------------------------------===// | |||
| 510 | // API to update (Post)DominatorTree information based on modifications to | |||
| 511 | // the CFG... | |||
| 512 | ||||
| 513 | /// Inform the dominator tree about a sequence of CFG edge insertions and | |||
| 514 | /// deletions and perform a batch update on the tree. | |||
| 515 | /// | |||
| 516 | /// This function should be used when there were multiple CFG updates after | |||
| 517 | /// the last dominator tree update. It takes care of performing the updates | |||
| 518 | /// in sync with the CFG and optimizes away the redundant operations that | |||
| 519 | /// cancel each other. | |||
| 520 | /// The functions expects the sequence of updates to be balanced. Eg.: | |||
| 521 | /// - {{Insert, A, B}, {Delete, A, B}, {Insert, A, B}} is fine, because | |||
| 522 | /// logically it results in a single insertions. | |||
| 523 | /// - {{Insert, A, B}, {Insert, A, B}} is invalid, because it doesn't make | |||
| 524 | /// sense to insert the same edge twice. | |||
| 525 | /// | |||
| 526 | /// What's more, the functions assumes that it's safe to ask every node in the | |||
| 527 | /// CFG about its children and inverse children. This implies that deletions | |||
| 528 | /// of CFG edges must not delete the CFG nodes before calling this function. | |||
| 529 | /// | |||
| 530 | /// The applyUpdates function can reorder the updates and remove redundant | |||
| 531 | /// ones internally. The batch updater is also able to detect sequences of | |||
| 532 | /// zero and exactly one update -- it's optimized to do less work in these | |||
| 533 | /// cases. | |||
| 534 | /// | |||
| 535 | /// Note that for postdominators it automatically takes care of applying | |||
| 536 | /// updates on reverse edges internally (so there's no need to swap the | |||
| 537 | /// From and To pointers when constructing DominatorTree::UpdateType). | |||
| 538 | /// The type of updates is the same for DomTreeBase<T> and PostDomTreeBase<T> | |||
| 539 | /// with the same template parameter T. | |||
| 540 | /// | |||
| 541 | /// \param Updates An unordered sequence of updates to perform. The current | |||
| 542 | /// CFG and the reverse of these updates provides the pre-view of the CFG. | |||
| 543 | /// | |||
| 544 | void applyUpdates(ArrayRef<UpdateType> Updates) { | |||
| 545 | GraphDiff<NodePtr, IsPostDominator> PreViewCFG( | |||
| 546 | Updates, /*ReverseApplyUpdates=*/true); | |||
| 547 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, nullptr); | |||
| 548 | } | |||
| 549 | ||||
| 550 | /// \param Updates An unordered sequence of updates to perform. The current | |||
| 551 | /// CFG and the reverse of these updates provides the pre-view of the CFG. | |||
| 552 | /// \param PostViewUpdates An unordered sequence of update to perform in order | |||
| 553 | /// to obtain a post-view of the CFG. The DT will be updated assuming the | |||
| 554 | /// obtained PostViewCFG is the desired end state. | |||
| 555 | void applyUpdates(ArrayRef<UpdateType> Updates, | |||
| 556 | ArrayRef<UpdateType> PostViewUpdates) { | |||
| 557 | if (Updates.empty()) { | |||
| 558 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); | |||
| 559 | DomTreeBuilder::ApplyUpdates(*this, PostViewCFG, &PostViewCFG); | |||
| 560 | } else { | |||
| 561 | // PreViewCFG needs to merge Updates and PostViewCFG. The updates in | |||
| 562 | // Updates need to be reversed, and match the direction in PostViewCFG. | |||
| 563 | // The PostViewCFG is created with updates reversed (equivalent to changes | |||
| 564 | // made to the CFG), so the PreViewCFG needs all the updates reverse | |||
| 565 | // applied. | |||
| 566 | SmallVector<UpdateType> AllUpdates(Updates.begin(), Updates.end()); | |||
| 567 | append_range(AllUpdates, PostViewUpdates); | |||
| 568 | GraphDiff<NodePtr, IsPostDom> PreViewCFG(AllUpdates, | |||
| 569 | /*ReverseApplyUpdates=*/true); | |||
| 570 | GraphDiff<NodePtr, IsPostDom> PostViewCFG(PostViewUpdates); | |||
| 571 | DomTreeBuilder::ApplyUpdates(*this, PreViewCFG, &PostViewCFG); | |||
| 572 | } | |||
| 573 | } | |||
| 574 | ||||
| 575 | /// Inform the dominator tree about a CFG edge insertion and update the tree. | |||
| 576 | /// | |||
| 577 | /// This function has to be called just before or just after making the update | |||
| 578 | /// on the actual CFG. There cannot be any other updates that the dominator | |||
| 579 | /// tree doesn't know about. | |||
| 580 | /// | |||
| 581 | /// Note that for postdominators it automatically takes care of inserting | |||
| 582 | /// a reverse edge internally (so there's no need to swap the parameters). | |||
| 583 | /// | |||
| 584 | void insertEdge(NodeT *From, NodeT *To) { | |||
| 585 | assert(From)((void)0); | |||
| 586 | assert(To)((void)0); | |||
| 587 | assert(From->getParent() == Parent)((void)0); | |||
| 588 | assert(To->getParent() == Parent)((void)0); | |||
| 589 | DomTreeBuilder::InsertEdge(*this, From, To); | |||
| 590 | } | |||
| 591 | ||||
| 592 | /// Inform the dominator tree about a CFG edge deletion and update the tree. | |||
| 593 | /// | |||
| 594 | /// This function has to be called just after making the update on the actual | |||
| 595 | /// CFG. An internal functions checks if the edge doesn't exist in the CFG in | |||
| 596 | /// DEBUG mode. There cannot be any other updates that the | |||
| 597 | /// dominator tree doesn't know about. | |||
| 598 | /// | |||
| 599 | /// Note that for postdominators it automatically takes care of deleting | |||
| 600 | /// a reverse edge internally (so there's no need to swap the parameters). | |||
| 601 | /// | |||
| 602 | void deleteEdge(NodeT *From, NodeT *To) { | |||
| 603 | assert(From)((void)0); | |||
| 604 | assert(To)((void)0); | |||
| 605 | assert(From->getParent() == Parent)((void)0); | |||
| 606 | assert(To->getParent() == Parent)((void)0); | |||
| 607 | DomTreeBuilder::DeleteEdge(*this, From, To); | |||
| 608 | } | |||
| 609 | ||||
| 610 | /// Add a new node to the dominator tree information. | |||
| 611 | /// | |||
| 612 | /// This creates a new node as a child of DomBB dominator node, linking it | |||
| 613 | /// into the children list of the immediate dominator. | |||
| 614 | /// | |||
| 615 | /// \param BB New node in CFG. | |||
| 616 | /// \param DomBB CFG node that is dominator for BB. | |||
| 617 | /// \returns New dominator tree node that represents new CFG node. | |||
| 618 | /// | |||
| 619 | DomTreeNodeBase<NodeT> *addNewBlock(NodeT *BB, NodeT *DomBB) { | |||
| 620 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((void)0); | |||
| 621 | DomTreeNodeBase<NodeT> *IDomNode = getNode(DomBB); | |||
| 622 | assert(IDomNode && "Not immediate dominator specified for block!")((void)0); | |||
| 623 | DFSInfoValid = false; | |||
| 624 | return createChild(BB, IDomNode); | |||
| 625 | } | |||
| 626 | ||||
| 627 | /// Add a new node to the forward dominator tree and make it a new root. | |||
| 628 | /// | |||
| 629 | /// \param BB New node in CFG. | |||
| 630 | /// \returns New dominator tree node that represents new CFG node. | |||
| 631 | /// | |||
| 632 | DomTreeNodeBase<NodeT> *setNewRoot(NodeT *BB) { | |||
| 633 | assert(getNode(BB) == nullptr && "Block already in dominator tree!")((void)0); | |||
| 634 | assert(!this->isPostDominator() &&((void)0) | |||
| 635 | "Cannot change root of post-dominator tree")((void)0); | |||
| 636 | DFSInfoValid = false; | |||
| 637 | DomTreeNodeBase<NodeT> *NewNode = createNode(BB); | |||
| 638 | if (Roots.empty()) { | |||
| 639 | addRoot(BB); | |||
| 640 | } else { | |||
| 641 | assert(Roots.size() == 1)((void)0); | |||
| 642 | NodeT *OldRoot = Roots.front(); | |||
| 643 | auto &OldNode = DomTreeNodes[OldRoot]; | |||
| 644 | OldNode = NewNode->addChild(std::move(DomTreeNodes[OldRoot])); | |||
| 645 | OldNode->IDom = NewNode; | |||
| 646 | OldNode->UpdateLevel(); | |||
| 647 | Roots[0] = BB; | |||
| 648 | } | |||
| 649 | return RootNode = NewNode; | |||
| 650 | } | |||
| 651 | ||||
| 652 | /// changeImmediateDominator - This method is used to update the dominator | |||
| 653 | /// tree information when a node's immediate dominator changes. | |||
| 654 | /// | |||
| 655 | void changeImmediateDominator(DomTreeNodeBase<NodeT> *N, | |||
| 656 | DomTreeNodeBase<NodeT> *NewIDom) { | |||
| 657 | assert(N && NewIDom && "Cannot change null node pointers!")((void)0); | |||
| 658 | DFSInfoValid = false; | |||
| 659 | N->setIDom(NewIDom); | |||
| 660 | } | |||
| 661 | ||||
| 662 | void changeImmediateDominator(NodeT *BB, NodeT *NewBB) { | |||
| 663 | changeImmediateDominator(getNode(BB), getNode(NewBB)); | |||
| 664 | } | |||
| 665 | ||||
| 666 | /// eraseNode - Removes a node from the dominator tree. Block must not | |||
| 667 | /// dominate any other blocks. Removes node from its immediate dominator's | |||
| 668 | /// children list. Deletes dominator node associated with basic block BB. | |||
| 669 | void eraseNode(NodeT *BB) { | |||
| 670 | DomTreeNodeBase<NodeT> *Node = getNode(BB); | |||
| 671 | assert(Node && "Removing node that isn't in dominator tree.")((void)0); | |||
| 672 | assert(Node->isLeaf() && "Node is not a leaf node.")((void)0); | |||
| 673 | ||||
| 674 | DFSInfoValid = false; | |||
| 675 | ||||
| 676 | // Remove node from immediate dominator's children list. | |||
| 677 | DomTreeNodeBase<NodeT> *IDom = Node->getIDom(); | |||
| 678 | if (IDom) { | |||
| 679 | const auto I = find(IDom->Children, Node); | |||
| 680 | assert(I != IDom->Children.end() &&((void)0) | |||
| 681 | "Not in immediate dominator children set!")((void)0); | |||
| 682 | // I am no longer your child... | |||
| 683 | IDom->Children.erase(I); | |||
| 684 | } | |||
| 685 | ||||
| 686 | DomTreeNodes.erase(BB); | |||
| 687 | ||||
| 688 | if (!IsPostDom) return; | |||
| 689 | ||||
| 690 | // Remember to update PostDominatorTree roots. | |||
| 691 | auto RIt = llvm::find(Roots, BB); | |||
| 692 | if (RIt != Roots.end()) { | |||
| 693 | std::swap(*RIt, Roots.back()); | |||
| 694 | Roots.pop_back(); | |||
| 695 | } | |||
| 696 | } | |||
| 697 | ||||
| 698 | /// splitBlock - BB is split and now it has one successor. Update dominator | |||
| 699 | /// tree to reflect this change. | |||
| 700 | void splitBlock(NodeT *NewBB) { | |||
| 701 | if (IsPostDominator) | |||
| 702 | Split<Inverse<NodeT *>>(NewBB); | |||
| 703 | else | |||
| 704 | Split<NodeT *>(NewBB); | |||
| 705 | } | |||
| 706 | ||||
| 707 | /// print - Convert to human readable form | |||
| 708 | /// | |||
| 709 | void print(raw_ostream &O) const { | |||
| 710 | O << "=============================--------------------------------\n"; | |||
| 711 | if (IsPostDominator) | |||
| 712 | O << "Inorder PostDominator Tree: "; | |||
| 713 | else | |||
| 714 | O << "Inorder Dominator Tree: "; | |||
| 715 | if (!DFSInfoValid) | |||
| 716 | O << "DFSNumbers invalid: " << SlowQueries << " slow queries."; | |||
| 717 | O << "\n"; | |||
| 718 | ||||
| 719 | // The postdom tree can have a null root if there are no returns. | |||
| 720 | if (getRootNode()) PrintDomTree<NodeT>(getRootNode(), O, 1); | |||
| 721 | O << "Roots: "; | |||
| 722 | for (const NodePtr Block : Roots) { | |||
| 723 | Block->printAsOperand(O, false); | |||
| 724 | O << " "; | |||
| 725 | } | |||
| 726 | O << "\n"; | |||
| 727 | } | |||
| 728 | ||||
| 729 | public: | |||
| 730 | /// updateDFSNumbers - Assign In and Out numbers to the nodes while walking | |||
| 731 | /// dominator tree in dfs order. | |||
| 732 | void updateDFSNumbers() const { | |||
| 733 | if (DFSInfoValid) { | |||
| 734 | SlowQueries = 0; | |||
| 735 | return; | |||
| 736 | } | |||
| 737 | ||||
| 738 | SmallVector<std::pair<const DomTreeNodeBase<NodeT> *, | |||
| 739 | typename DomTreeNodeBase<NodeT>::const_iterator>, | |||
| 740 | 32> WorkStack; | |||
| 741 | ||||
| 742 | const DomTreeNodeBase<NodeT> *ThisRoot = getRootNode(); | |||
| 743 | assert((!Parent || ThisRoot) && "Empty constructed DomTree")((void)0); | |||
| 744 | if (!ThisRoot) | |||
| 745 | return; | |||
| 746 | ||||
| 747 | // Both dominators and postdominators have a single root node. In the case | |||
| 748 | // case of PostDominatorTree, this node is a virtual root. | |||
| 749 | WorkStack.push_back({ThisRoot, ThisRoot->begin()}); | |||
| 750 | ||||
| 751 | unsigned DFSNum = 0; | |||
| 752 | ThisRoot->DFSNumIn = DFSNum++; | |||
| 753 | ||||
| 754 | while (!WorkStack.empty()) { | |||
| 755 | const DomTreeNodeBase<NodeT> *Node = WorkStack.back().first; | |||
| 756 | const auto ChildIt = WorkStack.back().second; | |||
| 757 | ||||
| 758 | // If we visited all of the children of this node, "recurse" back up the | |||
| 759 | // stack setting the DFOutNum. | |||
| 760 | if (ChildIt == Node->end()) { | |||
| 761 | Node->DFSNumOut = DFSNum++; | |||
| 762 | WorkStack.pop_back(); | |||
| 763 | } else { | |||
| 764 | // Otherwise, recursively visit this child. | |||
| 765 | const DomTreeNodeBase<NodeT> *Child = *ChildIt; | |||
| 766 | ++WorkStack.back().second; | |||
| 767 | ||||
| 768 | WorkStack.push_back({Child, Child->begin()}); | |||
| 769 | Child->DFSNumIn = DFSNum++; | |||
| 770 | } | |||
| 771 | } | |||
| 772 | ||||
| 773 | SlowQueries = 0; | |||
| 774 | DFSInfoValid = true; | |||
| 775 | } | |||
| 776 | ||||
| 777 | /// recalculate - compute a dominator tree for the given function | |||
| 778 | void recalculate(ParentType &Func) { | |||
| 779 | Parent = &Func; | |||
| 780 | DomTreeBuilder::Calculate(*this); | |||
| 781 | } | |||
| 782 | ||||
| 783 | void recalculate(ParentType &Func, ArrayRef<UpdateType> Updates) { | |||
| 784 | Parent = &Func; | |||
| 785 | DomTreeBuilder::CalculateWithUpdates(*this, Updates); | |||
| 786 | } | |||
| 787 | ||||
| 788 | /// verify - checks if the tree is correct. There are 3 level of verification: | |||
| 789 | /// - Full -- verifies if the tree is correct by making sure all the | |||
| 790 | /// properties (including the parent and the sibling property) | |||
| 791 | /// hold. | |||
| 792 | /// Takes O(N^3) time. | |||
| 793 | /// | |||
| 794 | /// - Basic -- checks if the tree is correct, but compares it to a freshly | |||
| 795 | /// constructed tree instead of checking the sibling property. | |||
| 796 | /// Takes O(N^2) time. | |||
| 797 | /// | |||
| 798 | /// - Fast -- checks basic tree structure and compares it with a freshly | |||
| 799 | /// constructed tree. | |||
| 800 | /// Takes O(N^2) time worst case, but is faster in practise (same | |||
| 801 | /// as tree construction). | |||
| 802 | bool verify(VerificationLevel VL = VerificationLevel::Full) const { | |||
| 803 | return DomTreeBuilder::Verify(*this, VL); | |||
| 804 | } | |||
| 805 | ||||
| 806 | void reset() { | |||
| 807 | DomTreeNodes.clear(); | |||
| 808 | Roots.clear(); | |||
| 809 | RootNode = nullptr; | |||
| 810 | Parent = nullptr; | |||
| 811 | DFSInfoValid = false; | |||
| 812 | SlowQueries = 0; | |||
| 813 | } | |||
| 814 | ||||
| 815 | protected: | |||
| 816 | void addRoot(NodeT *BB) { this->Roots.push_back(BB); } | |||
| 817 | ||||
| 818 | DomTreeNodeBase<NodeT> *createChild(NodeT *BB, DomTreeNodeBase<NodeT> *IDom) { | |||
| 819 | return (DomTreeNodes[BB] = IDom->addChild( | |||
| 820 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, IDom))) | |||
| 821 | .get(); | |||
| 822 | } | |||
| 823 | ||||
| 824 | DomTreeNodeBase<NodeT> *createNode(NodeT *BB) { | |||
| 825 | return (DomTreeNodes[BB] = | |||
| 826 | std::make_unique<DomTreeNodeBase<NodeT>>(BB, nullptr)) | |||
| 827 | .get(); | |||
| 828 | } | |||
| 829 | ||||
| 830 | // NewBB is split and now it has one successor. Update dominator tree to | |||
| 831 | // reflect this change. | |||
| 832 | template <class N> | |||
| 833 | void Split(typename GraphTraits<N>::NodeRef NewBB) { | |||
| 834 | using GraphT = GraphTraits<N>; | |||
| 835 | using NodeRef = typename GraphT::NodeRef; | |||
| 836 | assert(std::distance(GraphT::child_begin(NewBB),((void)0) | |||
| 837 | GraphT::child_end(NewBB)) == 1 &&((void)0) | |||
| 838 | "NewBB should have a single successor!")((void)0); | |||
| 839 | NodeRef NewBBSucc = *GraphT::child_begin(NewBB); | |||
| 840 | ||||
| 841 | SmallVector<NodeRef, 4> PredBlocks(children<Inverse<N>>(NewBB)); | |||
| 842 | ||||
| 843 | assert(!PredBlocks.empty() && "No predblocks?")((void)0); | |||
| 844 | ||||
| 845 | bool NewBBDominatesNewBBSucc = true; | |||
| 846 | for (auto Pred : children<Inverse<N>>(NewBBSucc)) { | |||
| 847 | if (Pred != NewBB && !dominates(NewBBSucc, Pred) && | |||
| 848 | isReachableFromEntry(Pred)) { | |||
| 849 | NewBBDominatesNewBBSucc = false; | |||
| 850 | break; | |||
| 851 | } | |||
| 852 | } | |||
| 853 | ||||
| 854 | // Find NewBB's immediate dominator and create new dominator tree node for | |||
| 855 | // NewBB. | |||
| 856 | NodeT *NewBBIDom = nullptr; | |||
| 857 | unsigned i = 0; | |||
| 858 | for (i = 0; i < PredBlocks.size(); ++i) | |||
| 859 | if (isReachableFromEntry(PredBlocks[i])) { | |||
| 860 | NewBBIDom = PredBlocks[i]; | |||
| 861 | break; | |||
| 862 | } | |||
| 863 | ||||
| 864 | // It's possible that none of the predecessors of NewBB are reachable; | |||
| 865 | // in that case, NewBB itself is unreachable, so nothing needs to be | |||
| 866 | // changed. | |||
| 867 | if (!NewBBIDom) return; | |||
| 868 | ||||
| 869 | for (i = i + 1; i < PredBlocks.size(); ++i) { | |||
| 870 | if (isReachableFromEntry(PredBlocks[i])) | |||
| 871 | NewBBIDom = findNearestCommonDominator(NewBBIDom, PredBlocks[i]); | |||
| 872 | } | |||
| 873 | ||||
| 874 | // Create the new dominator tree node... and set the idom of NewBB. | |||
| 875 | DomTreeNodeBase<NodeT> *NewBBNode = addNewBlock(NewBB, NewBBIDom); | |||
| 876 | ||||
| 877 | // If NewBB strictly dominates other blocks, then it is now the immediate | |||
| 878 | // dominator of NewBBSucc. Update the dominator tree as appropriate. | |||
| 879 | if (NewBBDominatesNewBBSucc) { | |||
| 880 | DomTreeNodeBase<NodeT> *NewBBSuccNode = getNode(NewBBSucc); | |||
| 881 | changeImmediateDominator(NewBBSuccNode, NewBBNode); | |||
| 882 | } | |||
| 883 | } | |||
| 884 | ||||
| 885 | private: | |||
| 886 | bool dominatedBySlowTreeWalk(const DomTreeNodeBase<NodeT> *A, | |||
| 887 | const DomTreeNodeBase<NodeT> *B) const { | |||
| 888 | assert(A != B)((void)0); | |||
| 889 | assert(isReachableFromEntry(B))((void)0); | |||
| 890 | assert(isReachableFromEntry(A))((void)0); | |||
| 891 | ||||
| 892 | const unsigned ALevel = A->getLevel(); | |||
| 893 | const DomTreeNodeBase<NodeT> *IDom; | |||
| 894 | ||||
| 895 | // Don't walk nodes above A's subtree. When we reach A's level, we must | |||
| 896 | // either find A or be in some other subtree not dominated by A. | |||
| 897 | while ((IDom = B->getIDom()) != nullptr && IDom->getLevel() >= ALevel) | |||
| 898 | B = IDom; // Walk up the tree | |||
| 899 | ||||
| 900 | return B == A; | |||
| 901 | } | |||
| 902 | ||||
| 903 | /// Wipe this tree's state without releasing any resources. | |||
| 904 | /// | |||
| 905 | /// This is essentially a post-move helper only. It leaves the object in an | |||
| 906 | /// assignable and destroyable state, but otherwise invalid. | |||
| 907 | void wipe() { | |||
| 908 | DomTreeNodes.clear(); | |||
| 909 | RootNode = nullptr; | |||
| 910 | Parent = nullptr; | |||
| 911 | } | |||
| 912 | }; | |||
| 913 | ||||
| 914 | template <typename T> | |||
| 915 | using DomTreeBase = DominatorTreeBase<T, false>; | |||
| 916 | ||||
| 917 | template <typename T> | |||
| 918 | using PostDomTreeBase = DominatorTreeBase<T, true>; | |||
| 919 | ||||
| 920 | // These two functions are declared out of line as a workaround for building | |||
| 921 | // with old (< r147295) versions of clang because of pr11642. | |||
| 922 | template <typename NodeT, bool IsPostDom> | |||
| 923 | bool DominatorTreeBase<NodeT, IsPostDom>::dominates(const NodeT *A, | |||
| 924 | const NodeT *B) const { | |||
| 925 | if (A == B) | |||
| 926 | return true; | |||
| 927 | ||||
| 928 | // Cast away the const qualifiers here. This is ok since | |||
| 929 | // this function doesn't actually return the values returned | |||
| 930 | // from getNode. | |||
| 931 | return dominates(getNode(const_cast<NodeT *>(A)), | |||
| 932 | getNode(const_cast<NodeT *>(B))); | |||
| 933 | } | |||
| 934 | template <typename NodeT, bool IsPostDom> | |||
| 935 | bool DominatorTreeBase<NodeT, IsPostDom>::properlyDominates( | |||
| 936 | const NodeT *A, const NodeT *B) const { | |||
| 937 | if (A == B) | |||
| 938 | return false; | |||
| 939 | ||||
| 940 | // Cast away the const qualifiers here. This is ok since | |||
| 941 | // this function doesn't actually return the values returned | |||
| 942 | // from getNode. | |||
| 943 | return dominates(getNode(const_cast<NodeT *>(A)), | |||
| 944 | getNode(const_cast<NodeT *>(B))); | |||
| 945 | } | |||
| 946 | ||||
| 947 | } // end namespace llvm | |||
| 948 | ||||
| 949 | #endif // LLVM_SUPPORT_GENERICDOMTREE_H |
| 1 | //===- llvm/ADT/DenseMap.h - Dense probed hash table ------------*- 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 DenseMap class. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #ifndef LLVM_ADT_DENSEMAP_H |
| 14 | #define LLVM_ADT_DENSEMAP_H |
| 15 | |
| 16 | #include "llvm/ADT/DenseMapInfo.h" |
| 17 | #include "llvm/ADT/EpochTracker.h" |
| 18 | #include "llvm/Support/AlignOf.h" |
| 19 | #include "llvm/Support/Compiler.h" |
| 20 | #include "llvm/Support/MathExtras.h" |
| 21 | #include "llvm/Support/MemAlloc.h" |
| 22 | #include "llvm/Support/ReverseIteration.h" |
| 23 | #include "llvm/Support/type_traits.h" |
| 24 | #include <algorithm> |
| 25 | #include <cassert> |
| 26 | #include <cstddef> |
| 27 | #include <cstring> |
| 28 | #include <initializer_list> |
| 29 | #include <iterator> |
| 30 | #include <new> |
| 31 | #include <type_traits> |
| 32 | #include <utility> |
| 33 | |
| 34 | namespace llvm { |
| 35 | |
| 36 | namespace detail { |
| 37 | |
| 38 | // We extend a pair to allow users to override the bucket type with their own |
| 39 | // implementation without requiring two members. |
| 40 | template <typename KeyT, typename ValueT> |
| 41 | struct DenseMapPair : public std::pair<KeyT, ValueT> { |
| 42 | using std::pair<KeyT, ValueT>::pair; |
| 43 | |
| 44 | KeyT &getFirst() { return std::pair<KeyT, ValueT>::first; } |
| 45 | const KeyT &getFirst() const { return std::pair<KeyT, ValueT>::first; } |
| 46 | ValueT &getSecond() { return std::pair<KeyT, ValueT>::second; } |
| 47 | const ValueT &getSecond() const { return std::pair<KeyT, ValueT>::second; } |
| 48 | }; |
| 49 | |
| 50 | } // end namespace detail |
| 51 | |
| 52 | template <typename KeyT, typename ValueT, |
| 53 | typename KeyInfoT = DenseMapInfo<KeyT>, |
| 54 | typename Bucket = llvm::detail::DenseMapPair<KeyT, ValueT>, |
| 55 | bool IsConst = false> |
| 56 | class DenseMapIterator; |
| 57 | |
| 58 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
| 59 | typename BucketT> |
| 60 | class DenseMapBase : public DebugEpochBase { |
| 61 | template <typename T> |
| 62 | using const_arg_type_t = typename const_pointer_or_const_ref<T>::type; |
| 63 | |
| 64 | public: |
| 65 | using size_type = unsigned; |
| 66 | using key_type = KeyT; |
| 67 | using mapped_type = ValueT; |
| 68 | using value_type = BucketT; |
| 69 | |
| 70 | using iterator = DenseMapIterator<KeyT, ValueT, KeyInfoT, BucketT>; |
| 71 | using const_iterator = |
| 72 | DenseMapIterator<KeyT, ValueT, KeyInfoT, BucketT, true>; |
| 73 | |
| 74 | inline iterator begin() { |
| 75 | // When the map is empty, avoid the overhead of advancing/retreating past |
| 76 | // empty buckets. |
| 77 | if (empty()) |
| 78 | return end(); |
| 79 | if (shouldReverseIterate<KeyT>()) |
| 80 | return makeIterator(getBucketsEnd() - 1, getBuckets(), *this); |
| 81 | return makeIterator(getBuckets(), getBucketsEnd(), *this); |
| 82 | } |
| 83 | inline iterator end() { |
| 84 | return makeIterator(getBucketsEnd(), getBucketsEnd(), *this, true); |
| 85 | } |
| 86 | inline const_iterator begin() const { |
| 87 | if (empty()) |
| 88 | return end(); |
| 89 | if (shouldReverseIterate<KeyT>()) |
| 90 | return makeConstIterator(getBucketsEnd() - 1, getBuckets(), *this); |
| 91 | return makeConstIterator(getBuckets(), getBucketsEnd(), *this); |
| 92 | } |
| 93 | inline const_iterator end() const { |
| 94 | return makeConstIterator(getBucketsEnd(), getBucketsEnd(), *this, true); |
| 95 | } |
| 96 | |
| 97 | LLVM_NODISCARD[[clang::warn_unused_result]] bool empty() const { |
| 98 | return getNumEntries() == 0; |
| 99 | } |
| 100 | unsigned size() const { return getNumEntries(); } |
| 101 | |
| 102 | /// Grow the densemap so that it can contain at least \p NumEntries items |
| 103 | /// before resizing again. |
| 104 | void reserve(size_type NumEntries) { |
| 105 | auto NumBuckets = getMinBucketToReserveForEntries(NumEntries); |
| 106 | incrementEpoch(); |
| 107 | if (NumBuckets > getNumBuckets()) |
| 108 | grow(NumBuckets); |
| 109 | } |
| 110 | |
| 111 | void clear() { |
| 112 | incrementEpoch(); |
| 113 | if (getNumEntries() == 0 && getNumTombstones() == 0) return; |
| 114 | |
| 115 | // If the capacity of the array is huge, and the # elements used is small, |
| 116 | // shrink the array. |
| 117 | if (getNumEntries() * 4 < getNumBuckets() && getNumBuckets() > 64) { |
| 118 | shrink_and_clear(); |
| 119 | return; |
| 120 | } |
| 121 | |
| 122 | const KeyT EmptyKey = getEmptyKey(), TombstoneKey = getTombstoneKey(); |
| 123 | if (std::is_trivially_destructible<ValueT>::value) { |
| 124 | // Use a simpler loop when values don't need destruction. |
| 125 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) |
| 126 | P->getFirst() = EmptyKey; |
| 127 | } else { |
| 128 | unsigned NumEntries = getNumEntries(); |
| 129 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) { |
| 130 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey)) { |
| 131 | if (!KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) { |
| 132 | P->getSecond().~ValueT(); |
| 133 | --NumEntries; |
| 134 | } |
| 135 | P->getFirst() = EmptyKey; |
| 136 | } |
| 137 | } |
| 138 | assert(NumEntries == 0 && "Node count imbalance!")((void)0); |
| 139 | } |
| 140 | setNumEntries(0); |
| 141 | setNumTombstones(0); |
| 142 | } |
| 143 | |
| 144 | /// Return 1 if the specified key is in the map, 0 otherwise. |
| 145 | size_type count(const_arg_type_t<KeyT> Val) const { |
| 146 | const BucketT *TheBucket; |
| 147 | return LookupBucketFor(Val, TheBucket) ? 1 : 0; |
| 148 | } |
| 149 | |
| 150 | iterator find(const_arg_type_t<KeyT> Val) { |
| 151 | BucketT *TheBucket; |
| 152 | if (LookupBucketFor(Val, TheBucket)) |
| 153 | return makeIterator(TheBucket, |
| 154 | shouldReverseIterate<KeyT>() ? getBuckets() |
| 155 | : getBucketsEnd(), |
| 156 | *this, true); |
| 157 | return end(); |
| 158 | } |
| 159 | const_iterator find(const_arg_type_t<KeyT> Val) const { |
| 160 | const BucketT *TheBucket; |
| 161 | if (LookupBucketFor(Val, TheBucket)) |
| 162 | return makeConstIterator(TheBucket, |
| 163 | shouldReverseIterate<KeyT>() ? getBuckets() |
| 164 | : getBucketsEnd(), |
| 165 | *this, true); |
| 166 | return end(); |
| 167 | } |
| 168 | |
| 169 | /// Alternate version of find() which allows a different, and possibly |
| 170 | /// less expensive, key type. |
| 171 | /// The DenseMapInfo is responsible for supplying methods |
| 172 | /// getHashValue(LookupKeyT) and isEqual(LookupKeyT, KeyT) for each key |
| 173 | /// type used. |
| 174 | template<class LookupKeyT> |
| 175 | iterator find_as(const LookupKeyT &Val) { |
| 176 | BucketT *TheBucket; |
| 177 | if (LookupBucketFor(Val, TheBucket)) |
| 178 | return makeIterator(TheBucket, |
| 179 | shouldReverseIterate<KeyT>() ? getBuckets() |
| 180 | : getBucketsEnd(), |
| 181 | *this, true); |
| 182 | return end(); |
| 183 | } |
| 184 | template<class LookupKeyT> |
| 185 | const_iterator find_as(const LookupKeyT &Val) const { |
| 186 | const BucketT *TheBucket; |
| 187 | if (LookupBucketFor(Val, TheBucket)) |
| 188 | return makeConstIterator(TheBucket, |
| 189 | shouldReverseIterate<KeyT>() ? getBuckets() |
| 190 | : getBucketsEnd(), |
| 191 | *this, true); |
| 192 | return end(); |
| 193 | } |
| 194 | |
| 195 | /// lookup - Return the entry for the specified key, or a default |
| 196 | /// constructed value if no such entry exists. |
| 197 | ValueT lookup(const_arg_type_t<KeyT> Val) const { |
| 198 | const BucketT *TheBucket; |
| 199 | if (LookupBucketFor(Val, TheBucket)) |
| 200 | return TheBucket->getSecond(); |
| 201 | return ValueT(); |
| 202 | } |
| 203 | |
| 204 | // Inserts key,value pair into the map if the key isn't already in the map. |
| 205 | // If the key is already in the map, it returns false and doesn't update the |
| 206 | // value. |
| 207 | std::pair<iterator, bool> insert(const std::pair<KeyT, ValueT> &KV) { |
| 208 | return try_emplace(KV.first, KV.second); |
| 209 | } |
| 210 | |
| 211 | // Inserts key,value pair into the map if the key isn't already in the map. |
| 212 | // If the key is already in the map, it returns false and doesn't update the |
| 213 | // value. |
| 214 | std::pair<iterator, bool> insert(std::pair<KeyT, ValueT> &&KV) { |
| 215 | return try_emplace(std::move(KV.first), std::move(KV.second)); |
| 216 | } |
| 217 | |
| 218 | // Inserts key,value pair into the map if the key isn't already in the map. |
| 219 | // The value is constructed in-place if the key is not in the map, otherwise |
| 220 | // it is not moved. |
| 221 | template <typename... Ts> |
| 222 | std::pair<iterator, bool> try_emplace(KeyT &&Key, Ts &&... Args) { |
| 223 | BucketT *TheBucket; |
| 224 | if (LookupBucketFor(Key, TheBucket)) |
| 225 | return std::make_pair(makeIterator(TheBucket, |
| 226 | shouldReverseIterate<KeyT>() |
| 227 | ? getBuckets() |
| 228 | : getBucketsEnd(), |
| 229 | *this, true), |
| 230 | false); // Already in map. |
| 231 | |
| 232 | // Otherwise, insert the new element. |
| 233 | TheBucket = |
| 234 | InsertIntoBucket(TheBucket, std::move(Key), std::forward<Ts>(Args)...); |
| 235 | return std::make_pair(makeIterator(TheBucket, |
| 236 | shouldReverseIterate<KeyT>() |
| 237 | ? getBuckets() |
| 238 | : getBucketsEnd(), |
| 239 | *this, true), |
| 240 | true); |
| 241 | } |
| 242 | |
| 243 | // Inserts key,value pair into the map if the key isn't already in the map. |
| 244 | // The value is constructed in-place if the key is not in the map, otherwise |
| 245 | // it is not moved. |
| 246 | template <typename... Ts> |
| 247 | std::pair<iterator, bool> try_emplace(const KeyT &Key, Ts &&... Args) { |
| 248 | BucketT *TheBucket; |
| 249 | if (LookupBucketFor(Key, TheBucket)) |
| 250 | return std::make_pair(makeIterator(TheBucket, |
| 251 | shouldReverseIterate<KeyT>() |
| 252 | ? getBuckets() |
| 253 | : getBucketsEnd(), |
| 254 | *this, true), |
| 255 | false); // Already in map. |
| 256 | |
| 257 | // Otherwise, insert the new element. |
| 258 | TheBucket = InsertIntoBucket(TheBucket, Key, std::forward<Ts>(Args)...); |
| 259 | return std::make_pair(makeIterator(TheBucket, |
| 260 | shouldReverseIterate<KeyT>() |
| 261 | ? getBuckets() |
| 262 | : getBucketsEnd(), |
| 263 | *this, true), |
| 264 | true); |
| 265 | } |
| 266 | |
| 267 | /// Alternate version of insert() which allows a different, and possibly |
| 268 | /// less expensive, key type. |
| 269 | /// The DenseMapInfo is responsible for supplying methods |
| 270 | /// getHashValue(LookupKeyT) and isEqual(LookupKeyT, KeyT) for each key |
| 271 | /// type used. |
| 272 | template <typename LookupKeyT> |
| 273 | std::pair<iterator, bool> insert_as(std::pair<KeyT, ValueT> &&KV, |
| 274 | const LookupKeyT &Val) { |
| 275 | BucketT *TheBucket; |
| 276 | if (LookupBucketFor(Val, TheBucket)) |
| 277 | return std::make_pair(makeIterator(TheBucket, |
| 278 | shouldReverseIterate<KeyT>() |
| 279 | ? getBuckets() |
| 280 | : getBucketsEnd(), |
| 281 | *this, true), |
| 282 | false); // Already in map. |
| 283 | |
| 284 | // Otherwise, insert the new element. |
| 285 | TheBucket = InsertIntoBucketWithLookup(TheBucket, std::move(KV.first), |
| 286 | std::move(KV.second), Val); |
| 287 | return std::make_pair(makeIterator(TheBucket, |
| 288 | shouldReverseIterate<KeyT>() |
| 289 | ? getBuckets() |
| 290 | : getBucketsEnd(), |
| 291 | *this, true), |
| 292 | true); |
| 293 | } |
| 294 | |
| 295 | /// insert - Range insertion of pairs. |
| 296 | template<typename InputIt> |
| 297 | void insert(InputIt I, InputIt E) { |
| 298 | for (; I != E; ++I) |
| 299 | insert(*I); |
| 300 | } |
| 301 | |
| 302 | bool erase(const KeyT &Val) { |
| 303 | BucketT *TheBucket; |
| 304 | if (!LookupBucketFor(Val, TheBucket)) |
| 305 | return false; // not in map. |
| 306 | |
| 307 | TheBucket->getSecond().~ValueT(); |
| 308 | TheBucket->getFirst() = getTombstoneKey(); |
| 309 | decrementNumEntries(); |
| 310 | incrementNumTombstones(); |
| 311 | return true; |
| 312 | } |
| 313 | void erase(iterator I) { |
| 314 | BucketT *TheBucket = &*I; |
| 315 | TheBucket->getSecond().~ValueT(); |
| 316 | TheBucket->getFirst() = getTombstoneKey(); |
| 317 | decrementNumEntries(); |
| 318 | incrementNumTombstones(); |
| 319 | } |
| 320 | |
| 321 | value_type& FindAndConstruct(const KeyT &Key) { |
| 322 | BucketT *TheBucket; |
| 323 | if (LookupBucketFor(Key, TheBucket)) |
| 324 | return *TheBucket; |
| 325 | |
| 326 | return *InsertIntoBucket(TheBucket, Key); |
| 327 | } |
| 328 | |
| 329 | ValueT &operator[](const KeyT &Key) { |
| 330 | return FindAndConstruct(Key).second; |
| 331 | } |
| 332 | |
| 333 | value_type& FindAndConstruct(KeyT &&Key) { |
| 334 | BucketT *TheBucket; |
| 335 | if (LookupBucketFor(Key, TheBucket)) |
| 336 | return *TheBucket; |
| 337 | |
| 338 | return *InsertIntoBucket(TheBucket, std::move(Key)); |
| 339 | } |
| 340 | |
| 341 | ValueT &operator[](KeyT &&Key) { |
| 342 | return FindAndConstruct(std::move(Key)).second; |
| 343 | } |
| 344 | |
| 345 | /// isPointerIntoBucketsArray - Return true if the specified pointer points |
| 346 | /// somewhere into the DenseMap's array of buckets (i.e. either to a key or |
| 347 | /// value in the DenseMap). |
| 348 | bool isPointerIntoBucketsArray(const void *Ptr) const { |
| 349 | return Ptr >= getBuckets() && Ptr < getBucketsEnd(); |
| 350 | } |
| 351 | |
| 352 | /// getPointerIntoBucketsArray() - Return an opaque pointer into the buckets |
| 353 | /// array. In conjunction with the previous method, this can be used to |
| 354 | /// determine whether an insertion caused the DenseMap to reallocate. |
| 355 | const void *getPointerIntoBucketsArray() const { return getBuckets(); } |
| 356 | |
| 357 | protected: |
| 358 | DenseMapBase() = default; |
| 359 | |
| 360 | void destroyAll() { |
| 361 | if (getNumBuckets() == 0) // Nothing to do. |
| 362 | return; |
| 363 | |
| 364 | const KeyT EmptyKey = getEmptyKey(), TombstoneKey = getTombstoneKey(); |
| 365 | for (BucketT *P = getBuckets(), *E = getBucketsEnd(); P != E; ++P) { |
| 366 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey) && |
| 367 | !KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) |
| 368 | P->getSecond().~ValueT(); |
| 369 | P->getFirst().~KeyT(); |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | void initEmpty() { |
| 374 | setNumEntries(0); |
| 375 | setNumTombstones(0); |
| 376 | |
| 377 | assert((getNumBuckets() & (getNumBuckets()-1)) == 0 &&((void)0) |
| 378 | "# initial buckets must be a power of two!")((void)0); |
| 379 | const KeyT EmptyKey = getEmptyKey(); |
| 380 | for (BucketT *B = getBuckets(), *E = getBucketsEnd(); B != E; ++B) |
| 381 | ::new (&B->getFirst()) KeyT(EmptyKey); |
| 382 | } |
| 383 | |
| 384 | /// Returns the number of buckets to allocate to ensure that the DenseMap can |
| 385 | /// accommodate \p NumEntries without need to grow(). |
| 386 | unsigned getMinBucketToReserveForEntries(unsigned NumEntries) { |
| 387 | // Ensure that "NumEntries * 4 < NumBuckets * 3" |
| 388 | if (NumEntries == 0) |
| 389 | return 0; |
| 390 | // +1 is required because of the strict equality. |
| 391 | // For example if NumEntries is 48, we need to return 401. |
| 392 | return NextPowerOf2(NumEntries * 4 / 3 + 1); |
| 393 | } |
| 394 | |
| 395 | void moveFromOldBuckets(BucketT *OldBucketsBegin, BucketT *OldBucketsEnd) { |
| 396 | initEmpty(); |
| 397 | |
| 398 | // Insert all the old elements. |
| 399 | const KeyT EmptyKey = getEmptyKey(); |
| 400 | const KeyT TombstoneKey = getTombstoneKey(); |
| 401 | for (BucketT *B = OldBucketsBegin, *E = OldBucketsEnd; B != E; ++B) { |
| 402 | if (!KeyInfoT::isEqual(B->getFirst(), EmptyKey) && |
| 403 | !KeyInfoT::isEqual(B->getFirst(), TombstoneKey)) { |
| 404 | // Insert the key/value into the new table. |
| 405 | BucketT *DestBucket; |
| 406 | bool FoundVal = LookupBucketFor(B->getFirst(), DestBucket); |
| 407 | (void)FoundVal; // silence warning. |
| 408 | assert(!FoundVal && "Key already in new map?")((void)0); |
| 409 | DestBucket->getFirst() = std::move(B->getFirst()); |
| 410 | ::new (&DestBucket->getSecond()) ValueT(std::move(B->getSecond())); |
| 411 | incrementNumEntries(); |
| 412 | |
| 413 | // Free the value. |
| 414 | B->getSecond().~ValueT(); |
| 415 | } |
| 416 | B->getFirst().~KeyT(); |
| 417 | } |
| 418 | } |
| 419 | |
| 420 | template <typename OtherBaseT> |
| 421 | void copyFrom( |
| 422 | const DenseMapBase<OtherBaseT, KeyT, ValueT, KeyInfoT, BucketT> &other) { |
| 423 | assert(&other != this)((void)0); |
| 424 | assert(getNumBuckets() == other.getNumBuckets())((void)0); |
| 425 | |
| 426 | setNumEntries(other.getNumEntries()); |
| 427 | setNumTombstones(other.getNumTombstones()); |
| 428 | |
| 429 | if (std::is_trivially_copyable<KeyT>::value && |
| 430 | std::is_trivially_copyable<ValueT>::value) |
| 431 | memcpy(reinterpret_cast<void *>(getBuckets()), other.getBuckets(), |
| 432 | getNumBuckets() * sizeof(BucketT)); |
| 433 | else |
| 434 | for (size_t i = 0; i < getNumBuckets(); ++i) { |
| 435 | ::new (&getBuckets()[i].getFirst()) |
| 436 | KeyT(other.getBuckets()[i].getFirst()); |
| 437 | if (!KeyInfoT::isEqual(getBuckets()[i].getFirst(), getEmptyKey()) && |
| 438 | !KeyInfoT::isEqual(getBuckets()[i].getFirst(), getTombstoneKey())) |
| 439 | ::new (&getBuckets()[i].getSecond()) |
| 440 | ValueT(other.getBuckets()[i].getSecond()); |
| 441 | } |
| 442 | } |
| 443 | |
| 444 | static unsigned getHashValue(const KeyT &Val) { |
| 445 | return KeyInfoT::getHashValue(Val); |
| 446 | } |
| 447 | |
| 448 | template<typename LookupKeyT> |
| 449 | static unsigned getHashValue(const LookupKeyT &Val) { |
| 450 | return KeyInfoT::getHashValue(Val); |
| 451 | } |
| 452 | |
| 453 | static const KeyT getEmptyKey() { |
| 454 | static_assert(std::is_base_of<DenseMapBase, DerivedT>::value, |
| 455 | "Must pass the derived type to this template!"); |
| 456 | return KeyInfoT::getEmptyKey(); |
| 457 | } |
| 458 | |
| 459 | static const KeyT getTombstoneKey() { |
| 460 | return KeyInfoT::getTombstoneKey(); |
| 461 | } |
| 462 | |
| 463 | private: |
| 464 | iterator makeIterator(BucketT *P, BucketT *E, |
| 465 | DebugEpochBase &Epoch, |
| 466 | bool NoAdvance=false) { |
| 467 | if (shouldReverseIterate<KeyT>()) { |
| 468 | BucketT *B = P == getBucketsEnd() ? getBuckets() : P + 1; |
| 469 | return iterator(B, E, Epoch, NoAdvance); |
| 470 | } |
| 471 | return iterator(P, E, Epoch, NoAdvance); |
| 472 | } |
| 473 | |
| 474 | const_iterator makeConstIterator(const BucketT *P, const BucketT *E, |
| 475 | const DebugEpochBase &Epoch, |
| 476 | const bool NoAdvance=false) const { |
| 477 | if (shouldReverseIterate<KeyT>()) { |
| 478 | const BucketT *B = P == getBucketsEnd() ? getBuckets() : P + 1; |
| 479 | return const_iterator(B, E, Epoch, NoAdvance); |
| 480 | } |
| 481 | return const_iterator(P, E, Epoch, NoAdvance); |
| 482 | } |
| 483 | |
| 484 | unsigned getNumEntries() const { |
| 485 | return static_cast<const DerivedT *>(this)->getNumEntries(); |
| 486 | } |
| 487 | |
| 488 | void setNumEntries(unsigned Num) { |
| 489 | static_cast<DerivedT *>(this)->setNumEntries(Num); |
| 490 | } |
| 491 | |
| 492 | void incrementNumEntries() { |
| 493 | setNumEntries(getNumEntries() + 1); |
| 494 | } |
| 495 | |
| 496 | void decrementNumEntries() { |
| 497 | setNumEntries(getNumEntries() - 1); |
| 498 | } |
| 499 | |
| 500 | unsigned getNumTombstones() const { |
| 501 | return static_cast<const DerivedT *>(this)->getNumTombstones(); |
| 502 | } |
| 503 | |
| 504 | void setNumTombstones(unsigned Num) { |
| 505 | static_cast<DerivedT *>(this)->setNumTombstones(Num); |
| 506 | } |
| 507 | |
| 508 | void incrementNumTombstones() { |
| 509 | setNumTombstones(getNumTombstones() + 1); |
| 510 | } |
| 511 | |
| 512 | void decrementNumTombstones() { |
| 513 | setNumTombstones(getNumTombstones() - 1); |
| 514 | } |
| 515 | |
| 516 | const BucketT *getBuckets() const { |
| 517 | return static_cast<const DerivedT *>(this)->getBuckets(); |
| 518 | } |
| 519 | |
| 520 | BucketT *getBuckets() { |
| 521 | return static_cast<DerivedT *>(this)->getBuckets(); |
| 522 | } |
| 523 | |
| 524 | unsigned getNumBuckets() const { |
| 525 | return static_cast<const DerivedT *>(this)->getNumBuckets(); |
| 526 | } |
| 527 | |
| 528 | BucketT *getBucketsEnd() { |
| 529 | return getBuckets() + getNumBuckets(); |
| 530 | } |
| 531 | |
| 532 | const BucketT *getBucketsEnd() const { |
| 533 | return getBuckets() + getNumBuckets(); |
| 534 | } |
| 535 | |
| 536 | void grow(unsigned AtLeast) { |
| 537 | static_cast<DerivedT *>(this)->grow(AtLeast); |
| 538 | } |
| 539 | |
| 540 | void shrink_and_clear() { |
| 541 | static_cast<DerivedT *>(this)->shrink_and_clear(); |
| 542 | } |
| 543 | |
| 544 | template <typename KeyArg, typename... ValueArgs> |
| 545 | BucketT *InsertIntoBucket(BucketT *TheBucket, KeyArg &&Key, |
| 546 | ValueArgs &&... Values) { |
| 547 | TheBucket = InsertIntoBucketImpl(Key, Key, TheBucket); |
| 548 | |
| 549 | TheBucket->getFirst() = std::forward<KeyArg>(Key); |
| 550 | ::new (&TheBucket->getSecond()) ValueT(std::forward<ValueArgs>(Values)...); |
| 551 | return TheBucket; |
| 552 | } |
| 553 | |
| 554 | template <typename LookupKeyT> |
| 555 | BucketT *InsertIntoBucketWithLookup(BucketT *TheBucket, KeyT &&Key, |
| 556 | ValueT &&Value, LookupKeyT &Lookup) { |
| 557 | TheBucket = InsertIntoBucketImpl(Key, Lookup, TheBucket); |
| 558 | |
| 559 | TheBucket->getFirst() = std::move(Key); |
| 560 | ::new (&TheBucket->getSecond()) ValueT(std::move(Value)); |
| 561 | return TheBucket; |
| 562 | } |
| 563 | |
| 564 | template <typename LookupKeyT> |
| 565 | BucketT *InsertIntoBucketImpl(const KeyT &Key, const LookupKeyT &Lookup, |
| 566 | BucketT *TheBucket) { |
| 567 | incrementEpoch(); |
| 568 | |
| 569 | // If the load of the hash table is more than 3/4, or if fewer than 1/8 of |
| 570 | // the buckets are empty (meaning that many are filled with tombstones), |
| 571 | // grow the table. |
| 572 | // |
| 573 | // The later case is tricky. For example, if we had one empty bucket with |
| 574 | // tons of tombstones, failing lookups (e.g. for insertion) would have to |
| 575 | // probe almost the entire table until it found the empty bucket. If the |
| 576 | // table completely filled with tombstones, no lookup would ever succeed, |
| 577 | // causing infinite loops in lookup. |
| 578 | unsigned NewNumEntries = getNumEntries() + 1; |
| 579 | unsigned NumBuckets = getNumBuckets(); |
| 580 | if (LLVM_UNLIKELY(NewNumEntries * 4 >= NumBuckets * 3)__builtin_expect((bool)(NewNumEntries * 4 >= NumBuckets * 3 ), false)) { |
| 581 | this->grow(NumBuckets * 2); |
| 582 | LookupBucketFor(Lookup, TheBucket); |
| 583 | NumBuckets = getNumBuckets(); |
| 584 | } else if (LLVM_UNLIKELY(NumBuckets-(NewNumEntries+getNumTombstones()) <=__builtin_expect((bool)(NumBuckets-(NewNumEntries+getNumTombstones ()) <= NumBuckets/8), false) |
| 585 | NumBuckets/8)__builtin_expect((bool)(NumBuckets-(NewNumEntries+getNumTombstones ()) <= NumBuckets/8), false)) { |
| 586 | this->grow(NumBuckets); |
| 587 | LookupBucketFor(Lookup, TheBucket); |
| 588 | } |
| 589 | assert(TheBucket)((void)0); |
| 590 | |
| 591 | // Only update the state after we've grown our bucket space appropriately |
| 592 | // so that when growing buckets we have self-consistent entry count. |
| 593 | incrementNumEntries(); |
| 594 | |
| 595 | // If we are writing over a tombstone, remember this. |
| 596 | const KeyT EmptyKey = getEmptyKey(); |
| 597 | if (!KeyInfoT::isEqual(TheBucket->getFirst(), EmptyKey)) |
| 598 | decrementNumTombstones(); |
| 599 | |
| 600 | return TheBucket; |
| 601 | } |
| 602 | |
| 603 | /// LookupBucketFor - Lookup the appropriate bucket for Val, returning it in |
| 604 | /// FoundBucket. If the bucket contains the key and a value, this returns |
| 605 | /// true, otherwise it returns a bucket with an empty marker or tombstone and |
| 606 | /// returns false. |
| 607 | template<typename LookupKeyT> |
| 608 | bool LookupBucketFor(const LookupKeyT &Val, |
| 609 | const BucketT *&FoundBucket) const { |
| 610 | const BucketT *BucketsPtr = getBuckets(); |
| 611 | const unsigned NumBuckets = getNumBuckets(); |
| 612 | |
| 613 | if (NumBuckets == 0) { |
| 614 | FoundBucket = nullptr; |
| 615 | return false; |
| 616 | } |
| 617 | |
| 618 | // FoundTombstone - Keep track of whether we find a tombstone while probing. |
| 619 | const BucketT *FoundTombstone = nullptr; |
| 620 | const KeyT EmptyKey = getEmptyKey(); |
| 621 | const KeyT TombstoneKey = getTombstoneKey(); |
| 622 | assert(!KeyInfoT::isEqual(Val, EmptyKey) &&((void)0) |
| 623 | !KeyInfoT::isEqual(Val, TombstoneKey) &&((void)0) |
| 624 | "Empty/Tombstone value shouldn't be inserted into map!")((void)0); |
| 625 | |
| 626 | unsigned BucketNo = getHashValue(Val) & (NumBuckets-1); |
| 627 | unsigned ProbeAmt = 1; |
| 628 | while (true) { |
| 629 | const BucketT *ThisBucket = BucketsPtr + BucketNo; |
| 630 | // Found Val's bucket? If so, return it. |
| 631 | if (LLVM_LIKELY(KeyInfoT::isEqual(Val, ThisBucket->getFirst()))__builtin_expect((bool)(KeyInfoT::isEqual(Val, ThisBucket-> getFirst())), true)) { |
| 632 | FoundBucket = ThisBucket; |
| 633 | return true; |
| 634 | } |
| 635 | |
| 636 | // If we found an empty bucket, the key doesn't exist in the set. |
| 637 | // Insert it and return the default value. |
| 638 | if (LLVM_LIKELY(KeyInfoT::isEqual(ThisBucket->getFirst(), EmptyKey))__builtin_expect((bool)(KeyInfoT::isEqual(ThisBucket->getFirst (), EmptyKey)), true)) { |
| 639 | // If we've already seen a tombstone while probing, fill it in instead |
| 640 | // of the empty bucket we eventually probed to. |
| 641 | FoundBucket = FoundTombstone ? FoundTombstone : ThisBucket; |
| 642 | return false; |
| 643 | } |
| 644 | |
| 645 | // If this is a tombstone, remember it. If Val ends up not in the map, we |
| 646 | // prefer to return it than something that would require more probing. |
| 647 | if (KeyInfoT::isEqual(ThisBucket->getFirst(), TombstoneKey) && |
| 648 | !FoundTombstone) |
| 649 | FoundTombstone = ThisBucket; // Remember the first tombstone found. |
| 650 | |
| 651 | // Otherwise, it's a hash collision or a tombstone, continue quadratic |
| 652 | // probing. |
| 653 | BucketNo += ProbeAmt++; |
| 654 | BucketNo &= (NumBuckets-1); |
| 655 | } |
| 656 | } |
| 657 | |
| 658 | template <typename LookupKeyT> |
| 659 | bool LookupBucketFor(const LookupKeyT &Val, BucketT *&FoundBucket) { |
| 660 | const BucketT *ConstFoundBucket; |
| 661 | bool Result = const_cast<const DenseMapBase *>(this) |
| 662 | ->LookupBucketFor(Val, ConstFoundBucket); |
| 663 | FoundBucket = const_cast<BucketT *>(ConstFoundBucket); |
| 664 | return Result; |
| 665 | } |
| 666 | |
| 667 | public: |
| 668 | /// Return the approximate size (in bytes) of the actual map. |
| 669 | /// This is just the raw memory used by DenseMap. |
| 670 | /// If entries are pointers to objects, the size of the referenced objects |
| 671 | /// are not included. |
| 672 | size_t getMemorySize() const { |
| 673 | return getNumBuckets() * sizeof(BucketT); |
| 674 | } |
| 675 | }; |
| 676 | |
| 677 | /// Equality comparison for DenseMap. |
| 678 | /// |
| 679 | /// Iterates over elements of LHS confirming that each (key, value) pair in LHS |
| 680 | /// is also in RHS, and that no additional pairs are in RHS. |
| 681 | /// Equivalent to N calls to RHS.find and N value comparisons. Amortized |
| 682 | /// complexity is linear, worst case is O(N^2) (if every hash collides). |
| 683 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
| 684 | typename BucketT> |
| 685 | bool operator==( |
| 686 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &LHS, |
| 687 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &RHS) { |
| 688 | if (LHS.size() != RHS.size()) |
| 689 | return false; |
| 690 | |
| 691 | for (auto &KV : LHS) { |
| 692 | auto I = RHS.find(KV.first); |
| 693 | if (I == RHS.end() || I->second != KV.second) |
| 694 | return false; |
| 695 | } |
| 696 | |
| 697 | return true; |
| 698 | } |
| 699 | |
| 700 | /// Inequality comparison for DenseMap. |
| 701 | /// |
| 702 | /// Equivalent to !(LHS == RHS). See operator== for performance notes. |
| 703 | template <typename DerivedT, typename KeyT, typename ValueT, typename KeyInfoT, |
| 704 | typename BucketT> |
| 705 | bool operator!=( |
| 706 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &LHS, |
| 707 | const DenseMapBase<DerivedT, KeyT, ValueT, KeyInfoT, BucketT> &RHS) { |
| 708 | return !(LHS == RHS); |
| 709 | } |
| 710 | |
| 711 | template <typename KeyT, typename ValueT, |
| 712 | typename KeyInfoT = DenseMapInfo<KeyT>, |
| 713 | typename BucketT = llvm::detail::DenseMapPair<KeyT, ValueT>> |
| 714 | class DenseMap : public DenseMapBase<DenseMap<KeyT, ValueT, KeyInfoT, BucketT>, |
| 715 | KeyT, ValueT, KeyInfoT, BucketT> { |
| 716 | friend class DenseMapBase<DenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
| 717 | |
| 718 | // Lift some types from the dependent base class into this class for |
| 719 | // simplicity of referring to them. |
| 720 | using BaseT = DenseMapBase<DenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
| 721 | |
| 722 | BucketT *Buckets; |
| 723 | unsigned NumEntries; |
| 724 | unsigned NumTombstones; |
| 725 | unsigned NumBuckets; |
| 726 | |
| 727 | public: |
| 728 | /// Create a DenseMap with an optional \p InitialReserve that guarantee that |
| 729 | /// this number of elements can be inserted in the map without grow() |
| 730 | explicit DenseMap(unsigned InitialReserve = 0) { init(InitialReserve); } |
| 731 | |
| 732 | DenseMap(const DenseMap &other) : BaseT() { |
| 733 | init(0); |
| 734 | copyFrom(other); |
| 735 | } |
| 736 | |
| 737 | DenseMap(DenseMap &&other) : BaseT() { |
| 738 | init(0); |
| 739 | swap(other); |
| 740 | } |
| 741 | |
| 742 | template<typename InputIt> |
| 743 | DenseMap(const InputIt &I, const InputIt &E) { |
| 744 | init(std::distance(I, E)); |
| 745 | this->insert(I, E); |
| 746 | } |
| 747 | |
| 748 | DenseMap(std::initializer_list<typename BaseT::value_type> Vals) { |
| 749 | init(Vals.size()); |
| 750 | this->insert(Vals.begin(), Vals.end()); |
| 751 | } |
| 752 | |
| 753 | ~DenseMap() { |
| 754 | this->destroyAll(); |
| 755 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
| 756 | } |
| 757 | |
| 758 | void swap(DenseMap& RHS) { |
| 759 | this->incrementEpoch(); |
| 760 | RHS.incrementEpoch(); |
| 761 | std::swap(Buckets, RHS.Buckets); |
| 762 | std::swap(NumEntries, RHS.NumEntries); |
| 763 | std::swap(NumTombstones, RHS.NumTombstones); |
| 764 | std::swap(NumBuckets, RHS.NumBuckets); |
| 765 | } |
| 766 | |
| 767 | DenseMap& operator=(const DenseMap& other) { |
| 768 | if (&other != this) |
| 769 | copyFrom(other); |
| 770 | return *this; |
| 771 | } |
| 772 | |
| 773 | DenseMap& operator=(DenseMap &&other) { |
| 774 | this->destroyAll(); |
| 775 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
| 776 | init(0); |
| 777 | swap(other); |
| 778 | return *this; |
| 779 | } |
| 780 | |
| 781 | void copyFrom(const DenseMap& other) { |
| 782 | this->destroyAll(); |
| 783 | deallocate_buffer(Buckets, sizeof(BucketT) * NumBuckets, alignof(BucketT)); |
| 784 | if (allocateBuckets(other.NumBuckets)) { |
| 785 | this->BaseT::copyFrom(other); |
| 786 | } else { |
| 787 | NumEntries = 0; |
| 788 | NumTombstones = 0; |
| 789 | } |
| 790 | } |
| 791 | |
| 792 | void init(unsigned InitNumEntries) { |
| 793 | auto InitBuckets = BaseT::getMinBucketToReserveForEntries(InitNumEntries); |
| 794 | if (allocateBuckets(InitBuckets)) { |
| 795 | this->BaseT::initEmpty(); |
| 796 | } else { |
| 797 | NumEntries = 0; |
| 798 | NumTombstones = 0; |
| 799 | } |
| 800 | } |
| 801 | |
| 802 | void grow(unsigned AtLeast) { |
| 803 | unsigned OldNumBuckets = NumBuckets; |
| 804 | BucketT *OldBuckets = Buckets; |
| 805 | |
| 806 | allocateBuckets(std::max<unsigned>(64, static_cast<unsigned>(NextPowerOf2(AtLeast-1)))); |
| 807 | assert(Buckets)((void)0); |
| 808 | if (!OldBuckets) { |
| 809 | this->BaseT::initEmpty(); |
| 810 | return; |
| 811 | } |
| 812 | |
| 813 | this->moveFromOldBuckets(OldBuckets, OldBuckets+OldNumBuckets); |
| 814 | |
| 815 | // Free the old table. |
| 816 | deallocate_buffer(OldBuckets, sizeof(BucketT) * OldNumBuckets, |
| 817 | alignof(BucketT)); |
| 818 | } |
| 819 | |
| 820 | void shrink_and_clear() { |
| 821 | unsigned OldNumBuckets = NumBuckets; |
| 822 | unsigned OldNumEntries = NumEntries; |
| 823 | this->destroyAll(); |
| 824 | |
| 825 | // Reduce the number of buckets. |
| 826 | unsigned NewNumBuckets = 0; |
| 827 | if (OldNumEntries) |
| 828 | NewNumBuckets = std::max(64, 1 << (Log2_32_Ceil(OldNumEntries) + 1)); |
| 829 | if (NewNumBuckets == NumBuckets) { |
| 830 | this->BaseT::initEmpty(); |
| 831 | return; |
| 832 | } |
| 833 | |
| 834 | deallocate_buffer(Buckets, sizeof(BucketT) * OldNumBuckets, |
| 835 | alignof(BucketT)); |
| 836 | init(NewNumBuckets); |
| 837 | } |
| 838 | |
| 839 | private: |
| 840 | unsigned getNumEntries() const { |
| 841 | return NumEntries; |
| 842 | } |
| 843 | |
| 844 | void setNumEntries(unsigned Num) { |
| 845 | NumEntries = Num; |
| 846 | } |
| 847 | |
| 848 | unsigned getNumTombstones() const { |
| 849 | return NumTombstones; |
| 850 | } |
| 851 | |
| 852 | void setNumTombstones(unsigned Num) { |
| 853 | NumTombstones = Num; |
| 854 | } |
| 855 | |
| 856 | BucketT *getBuckets() const { |
| 857 | return Buckets; |
| 858 | } |
| 859 | |
| 860 | unsigned getNumBuckets() const { |
| 861 | return NumBuckets; |
| 862 | } |
| 863 | |
| 864 | bool allocateBuckets(unsigned Num) { |
| 865 | NumBuckets = Num; |
| 866 | if (NumBuckets == 0) { |
| 867 | Buckets = nullptr; |
| 868 | return false; |
| 869 | } |
| 870 | |
| 871 | Buckets = static_cast<BucketT *>( |
| 872 | allocate_buffer(sizeof(BucketT) * NumBuckets, alignof(BucketT))); |
| 873 | return true; |
| 874 | } |
| 875 | }; |
| 876 | |
| 877 | template <typename KeyT, typename ValueT, unsigned InlineBuckets = 4, |
| 878 | typename KeyInfoT = DenseMapInfo<KeyT>, |
| 879 | typename BucketT = llvm::detail::DenseMapPair<KeyT, ValueT>> |
| 880 | class SmallDenseMap |
| 881 | : public DenseMapBase< |
| 882 | SmallDenseMap<KeyT, ValueT, InlineBuckets, KeyInfoT, BucketT>, KeyT, |
| 883 | ValueT, KeyInfoT, BucketT> { |
| 884 | friend class DenseMapBase<SmallDenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
| 885 | |
| 886 | // Lift some types from the dependent base class into this class for |
| 887 | // simplicity of referring to them. |
| 888 | using BaseT = DenseMapBase<SmallDenseMap, KeyT, ValueT, KeyInfoT, BucketT>; |
| 889 | |
| 890 | static_assert(isPowerOf2_64(InlineBuckets), |
| 891 | "InlineBuckets must be a power of 2."); |
| 892 | |
| 893 | unsigned Small : 1; |
| 894 | unsigned NumEntries : 31; |
| 895 | unsigned NumTombstones; |
| 896 | |
| 897 | struct LargeRep { |
| 898 | BucketT *Buckets; |
| 899 | unsigned NumBuckets; |
| 900 | }; |
| 901 | |
| 902 | /// A "union" of an inline bucket array and the struct representing |
| 903 | /// a large bucket. This union will be discriminated by the 'Small' bit. |
| 904 | AlignedCharArrayUnion<BucketT[InlineBuckets], LargeRep> storage; |
| 905 | |
| 906 | public: |
| 907 | explicit SmallDenseMap(unsigned NumInitBuckets = 0) { |
| 908 | init(NumInitBuckets); |
| 909 | } |
| 910 | |
| 911 | SmallDenseMap(const SmallDenseMap &other) : BaseT() { |
| 912 | init(0); |
| 913 | copyFrom(other); |
| 914 | } |
| 915 | |
| 916 | SmallDenseMap(SmallDenseMap &&other) : BaseT() { |
| 917 | init(0); |
| 918 | swap(other); |
| 919 | } |
| 920 | |
| 921 | template<typename InputIt> |
| 922 | SmallDenseMap(const InputIt &I, const InputIt &E) { |
| 923 | init(NextPowerOf2(std::distance(I, E))); |
| 924 | this->insert(I, E); |
| 925 | } |
| 926 | |
| 927 | SmallDenseMap(std::initializer_list<typename BaseT::value_type> Vals) |
| 928 | : SmallDenseMap(Vals.begin(), Vals.end()) {} |
| 929 | |
| 930 | ~SmallDenseMap() { |
| 931 | this->destroyAll(); |
| 932 | deallocateBuckets(); |
| 933 | } |
| 934 | |
| 935 | void swap(SmallDenseMap& RHS) { |
| 936 | unsigned TmpNumEntries = RHS.NumEntries; |
| 937 | RHS.NumEntries = NumEntries; |
| 938 | NumEntries = TmpNumEntries; |
| 939 | std::swap(NumTombstones, RHS.NumTombstones); |
| 940 | |
| 941 | const KeyT EmptyKey = this->getEmptyKey(); |
| 942 | const KeyT TombstoneKey = this->getTombstoneKey(); |
| 943 | if (Small && RHS.Small) { |
| 944 | // If we're swapping inline bucket arrays, we have to cope with some of |
| 945 | // the tricky bits of DenseMap's storage system: the buckets are not |
| 946 | // fully initialized. Thus we swap every key, but we may have |
| 947 | // a one-directional move of the value. |
| 948 | for (unsigned i = 0, e = InlineBuckets; i != e; ++i) { |
| 949 | BucketT *LHSB = &getInlineBuckets()[i], |
| 950 | *RHSB = &RHS.getInlineBuckets()[i]; |
| 951 | bool hasLHSValue = (!KeyInfoT::isEqual(LHSB->getFirst(), EmptyKey) && |
| 952 | !KeyInfoT::isEqual(LHSB->getFirst(), TombstoneKey)); |
| 953 | bool hasRHSValue = (!KeyInfoT::isEqual(RHSB->getFirst(), EmptyKey) && |
| 954 | !KeyInfoT::isEqual(RHSB->getFirst(), TombstoneKey)); |
| 955 | if (hasLHSValue && hasRHSValue) { |
| 956 | // Swap together if we can... |
| 957 | std::swap(*LHSB, *RHSB); |
| 958 | continue; |
| 959 | } |
| 960 | // Swap separately and handle any asymmetry. |
| 961 | std::swap(LHSB->getFirst(), RHSB->getFirst()); |
| 962 | if (hasLHSValue) { |
| 963 | ::new (&RHSB->getSecond()) ValueT(std::move(LHSB->getSecond())); |
| 964 | LHSB->getSecond().~ValueT(); |
| 965 | } else if (hasRHSValue) { |
| 966 | ::new (&LHSB->getSecond()) ValueT(std::move(RHSB->getSecond())); |
| 967 | RHSB->getSecond().~ValueT(); |
| 968 | } |
| 969 | } |
| 970 | return; |
| 971 | } |
| 972 | if (!Small && !RHS.Small) { |
| 973 | std::swap(getLargeRep()->Buckets, RHS.getLargeRep()->Buckets); |
| 974 | std::swap(getLargeRep()->NumBuckets, RHS.getLargeRep()->NumBuckets); |
| 975 | return; |
| 976 | } |
| 977 | |
| 978 | SmallDenseMap &SmallSide = Small ? *this : RHS; |
| 979 | SmallDenseMap &LargeSide = Small ? RHS : *this; |
| 980 | |
| 981 | // First stash the large side's rep and move the small side across. |
| 982 | LargeRep TmpRep = std::move(*LargeSide.getLargeRep()); |
| 983 | LargeSide.getLargeRep()->~LargeRep(); |
| 984 | LargeSide.Small = true; |
| 985 | // This is similar to the standard move-from-old-buckets, but the bucket |
| 986 | // count hasn't actually rotated in this case. So we have to carefully |
| 987 | // move construct the keys and values into their new locations, but there |
| 988 | // is no need to re-hash things. |
| 989 | for (unsigned i = 0, e = InlineBuckets; i != e; ++i) { |
| 990 | BucketT *NewB = &LargeSide.getInlineBuckets()[i], |
| 991 | *OldB = &SmallSide.getInlineBuckets()[i]; |
| 992 | ::new (&NewB->getFirst()) KeyT(std::move(OldB->getFirst())); |
| 993 | OldB->getFirst().~KeyT(); |
| 994 | if (!KeyInfoT::isEqual(NewB->getFirst(), EmptyKey) && |
| 995 | !KeyInfoT::isEqual(NewB->getFirst(), TombstoneKey)) { |
| 996 | ::new (&NewB->getSecond()) ValueT(std::move(OldB->getSecond())); |
| 997 | OldB->getSecond().~ValueT(); |
| 998 | } |
| 999 | } |
| 1000 | |
| 1001 | // The hard part of moving the small buckets across is done, just move |
| 1002 | // the TmpRep into its new home. |
| 1003 | SmallSide.Small = false; |
| 1004 | new (SmallSide.getLargeRep()) LargeRep(std::move(TmpRep)); |
| 1005 | } |
| 1006 | |
| 1007 | SmallDenseMap& operator=(const SmallDenseMap& other) { |
| 1008 | if (&other != this) |
| 1009 | copyFrom(other); |
| 1010 | return *this; |
| 1011 | } |
| 1012 | |
| 1013 | SmallDenseMap& operator=(SmallDenseMap &&other) { |
| 1014 | this->destroyAll(); |
| 1015 | deallocateBuckets(); |
| 1016 | init(0); |
| 1017 | swap(other); |
| 1018 | return *this; |
| 1019 | } |
| 1020 | |
| 1021 | void copyFrom(const SmallDenseMap& other) { |
| 1022 | this->destroyAll(); |
| 1023 | deallocateBuckets(); |
| 1024 | Small = true; |
| 1025 | if (other.getNumBuckets() > InlineBuckets) { |
| 1026 | Small = false; |
| 1027 | new (getLargeRep()) LargeRep(allocateBuckets(other.getNumBuckets())); |
| 1028 | } |
| 1029 | this->BaseT::copyFrom(other); |
| 1030 | } |
| 1031 | |
| 1032 | void init(unsigned InitBuckets) { |
| 1033 | Small = true; |
| 1034 | if (InitBuckets > InlineBuckets) { |
| 1035 | Small = false; |
| 1036 | new (getLargeRep()) LargeRep(allocateBuckets(InitBuckets)); |
| 1037 | } |
| 1038 | this->BaseT::initEmpty(); |
| 1039 | } |
| 1040 | |
| 1041 | void grow(unsigned AtLeast) { |
| 1042 | if (AtLeast > InlineBuckets) |
| 1043 | AtLeast = std::max<unsigned>(64, NextPowerOf2(AtLeast-1)); |
| 1044 | |
| 1045 | if (Small) { |
| 1046 | // First move the inline buckets into a temporary storage. |
| 1047 | AlignedCharArrayUnion<BucketT[InlineBuckets]> TmpStorage; |
| 1048 | BucketT *TmpBegin = reinterpret_cast<BucketT *>(&TmpStorage); |
| 1049 | BucketT *TmpEnd = TmpBegin; |
| 1050 | |
| 1051 | // Loop over the buckets, moving non-empty, non-tombstones into the |
| 1052 | // temporary storage. Have the loop move the TmpEnd forward as it goes. |
| 1053 | const KeyT EmptyKey = this->getEmptyKey(); |
| 1054 | const KeyT TombstoneKey = this->getTombstoneKey(); |
| 1055 | for (BucketT *P = getBuckets(), *E = P + InlineBuckets; P != E; ++P) { |
| 1056 | if (!KeyInfoT::isEqual(P->getFirst(), EmptyKey) && |
| 1057 | !KeyInfoT::isEqual(P->getFirst(), TombstoneKey)) { |
| 1058 | assert(size_t(TmpEnd - TmpBegin) < InlineBuckets &&((void)0) |
| 1059 | "Too many inline buckets!")((void)0); |
| 1060 | ::new (&TmpEnd->getFirst()) KeyT(std::move(P->getFirst())); |
| 1061 | ::new (&TmpEnd->getSecond()) ValueT(std::move(P->getSecond())); |
| 1062 | ++TmpEnd; |
| 1063 | P->getSecond().~ValueT(); |
| 1064 | } |
| 1065 | P->getFirst().~KeyT(); |
| 1066 | } |
| 1067 | |
| 1068 | // AtLeast == InlineBuckets can happen if there are many tombstones, |
| 1069 | // and grow() is used to remove them. Usually we always switch to the |
| 1070 | // large rep here. |
| 1071 | if (AtLeast > InlineBuckets) { |
| 1072 | Small = false; |
| 1073 | new (getLargeRep()) LargeRep(allocateBuckets(AtLeast)); |
| 1074 | } |
| 1075 | this->moveFromOldBuckets(TmpBegin, TmpEnd); |
| 1076 | return; |
| 1077 | } |
| 1078 | |
| 1079 | LargeRep OldRep = std::move(*getLargeRep()); |
| 1080 | getLargeRep()->~LargeRep(); |
| 1081 | if (AtLeast <= InlineBuckets) { |
| 1082 | Small = true; |
| 1083 | } else { |
| 1084 | new (getLargeRep()) LargeRep(allocateBuckets(AtLeast)); |
| 1085 | } |
| 1086 | |
| 1087 | this->moveFromOldBuckets(OldRep.Buckets, OldRep.Buckets+OldRep.NumBuckets); |
| 1088 | |
| 1089 | // Free the old table. |
| 1090 | deallocate_buffer(OldRep.Buckets, sizeof(BucketT) * OldRep.NumBuckets, |
| 1091 | alignof(BucketT)); |
| 1092 | } |
| 1093 | |
| 1094 | void shrink_and_clear() { |
| 1095 | unsigned OldSize = this->size(); |
| 1096 | this->destroyAll(); |
| 1097 | |
| 1098 | // Reduce the number of buckets. |
| 1099 | unsigned NewNumBuckets = 0; |
| 1100 | if (OldSize) { |
| 1101 | NewNumBuckets = 1 << (Log2_32_Ceil(OldSize) + 1); |
| 1102 | if (NewNumBuckets > InlineBuckets && NewNumBuckets < 64u) |
| 1103 | NewNumBuckets = 64; |
| 1104 | } |
| 1105 | if ((Small && NewNumBuckets <= InlineBuckets) || |
| 1106 | (!Small && NewNumBuckets == getLargeRep()->NumBuckets)) { |
| 1107 | this->BaseT::initEmpty(); |
| 1108 | return; |
| 1109 | } |
| 1110 | |
| 1111 | deallocateBuckets(); |
| 1112 | init(NewNumBuckets); |
| 1113 | } |
| 1114 | |
| 1115 | private: |
| 1116 | unsigned getNumEntries() const { |
| 1117 | return NumEntries; |
| 1118 | } |
| 1119 | |
| 1120 | void setNumEntries(unsigned Num) { |
| 1121 | // NumEntries is hardcoded to be 31 bits wide. |
| 1122 | assert(Num < (1U << 31) && "Cannot support more than 1<<31 entries")((void)0); |
| 1123 | NumEntries = Num; |
| 1124 | } |
| 1125 | |
| 1126 | unsigned getNumTombstones() const { |
| 1127 | return NumTombstones; |
| 1128 | } |
| 1129 | |
| 1130 | void setNumTombstones(unsigned Num) { |
| 1131 | NumTombstones = Num; |
| 1132 | } |
| 1133 | |
| 1134 | const BucketT *getInlineBuckets() const { |
| 1135 | assert(Small)((void)0); |
| 1136 | // Note that this cast does not violate aliasing rules as we assert that |
| 1137 | // the memory's dynamic type is the small, inline bucket buffer, and the |
| 1138 | // 'storage' is a POD containing a char buffer. |
| 1139 | return reinterpret_cast<const BucketT *>(&storage); |
| 1140 | } |
| 1141 | |
| 1142 | BucketT *getInlineBuckets() { |
| 1143 | return const_cast<BucketT *>( |
| 1144 | const_cast<const SmallDenseMap *>(this)->getInlineBuckets()); |
| 1145 | } |
| 1146 | |
| 1147 | const LargeRep *getLargeRep() const { |
| 1148 | assert(!Small)((void)0); |
| 1149 | // Note, same rule about aliasing as with getInlineBuckets. |
| 1150 | return reinterpret_cast<const LargeRep *>(&storage); |
| 1151 | } |
| 1152 | |
| 1153 | LargeRep *getLargeRep() { |
| 1154 | return const_cast<LargeRep *>( |
| 1155 | const_cast<const SmallDenseMap *>(this)->getLargeRep()); |
| 1156 | } |
| 1157 | |
| 1158 | const BucketT *getBuckets() const { |
| 1159 | return Small ? getInlineBuckets() : getLargeRep()->Buckets; |
| 1160 | } |
| 1161 | |
| 1162 | BucketT *getBuckets() { |
| 1163 | return const_cast<BucketT *>( |
| 1164 | const_cast<const SmallDenseMap *>(this)->getBuckets()); |
| 1165 | } |
| 1166 | |
| 1167 | unsigned getNumBuckets() const { |
| 1168 | return Small ? InlineBuckets : getLargeRep()->NumBuckets; |
| 1169 | } |
| 1170 | |
| 1171 | void deallocateBuckets() { |
| 1172 | if (Small) |
| 1173 | return; |
| 1174 | |
| 1175 | deallocate_buffer(getLargeRep()->Buckets, |
| 1176 | sizeof(BucketT) * getLargeRep()->NumBuckets, |
| 1177 | alignof(BucketT)); |
| 1178 | getLargeRep()->~LargeRep(); |
| 1179 | } |
| 1180 | |
| 1181 | LargeRep allocateBuckets(unsigned Num) { |
| 1182 | assert(Num > InlineBuckets && "Must allocate more buckets than are inline")((void)0); |
| 1183 | LargeRep Rep = {static_cast<BucketT *>(allocate_buffer( |
| 1184 | sizeof(BucketT) * Num, alignof(BucketT))), |
| 1185 | Num}; |
| 1186 | return Rep; |
| 1187 | } |
| 1188 | }; |
| 1189 | |
| 1190 | template <typename KeyT, typename ValueT, typename KeyInfoT, typename Bucket, |
| 1191 | bool IsConst> |
| 1192 | class DenseMapIterator : DebugEpochBase::HandleBase { |
| 1193 | friend class DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, true>; |
| 1194 | friend class DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, false>; |
| 1195 | |
| 1196 | public: |
| 1197 | using difference_type = ptrdiff_t; |
| 1198 | using value_type = |
| 1199 | typename std::conditional<IsConst, const Bucket, Bucket>::type; |
| 1200 | using pointer = value_type *; |
| 1201 | using reference = value_type &; |
| 1202 | using iterator_category = std::forward_iterator_tag; |
| 1203 | |
| 1204 | private: |
| 1205 | pointer Ptr = nullptr; |
| 1206 | pointer End = nullptr; |
| 1207 | |
| 1208 | public: |
| 1209 | DenseMapIterator() = default; |
| 1210 | |
| 1211 | DenseMapIterator(pointer Pos, pointer E, const DebugEpochBase &Epoch, |
| 1212 | bool NoAdvance = false) |
| 1213 | : DebugEpochBase::HandleBase(&Epoch), Ptr(Pos), End(E) { |
| 1214 | assert(isHandleInSync() && "invalid construction!")((void)0); |
| 1215 | |
| 1216 | if (NoAdvance) return; |
| 1217 | if (shouldReverseIterate<KeyT>()) { |
| 1218 | RetreatPastEmptyBuckets(); |
| 1219 | return; |
| 1220 | } |
| 1221 | AdvancePastEmptyBuckets(); |
| 1222 | } |
| 1223 | |
| 1224 | // Converting ctor from non-const iterators to const iterators. SFINAE'd out |
| 1225 | // for const iterator destinations so it doesn't end up as a user defined copy |
| 1226 | // constructor. |
| 1227 | template <bool IsConstSrc, |
| 1228 | typename = std::enable_if_t<!IsConstSrc && IsConst>> |
| 1229 | DenseMapIterator( |
| 1230 | const DenseMapIterator<KeyT, ValueT, KeyInfoT, Bucket, IsConstSrc> &I) |
| 1231 | : DebugEpochBase::HandleBase(I), Ptr(I.Ptr), End(I.End) {} |
| 1232 | |
| 1233 | reference operator*() const { |
| 1234 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
| 1235 | assert(Ptr != End && "dereferencing end() iterator")((void)0); |
| 1236 | if (shouldReverseIterate<KeyT>()) |
| 1237 | return Ptr[-1]; |
| 1238 | return *Ptr; |
| 1239 | } |
| 1240 | pointer operator->() const { |
| 1241 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
| 1242 | assert(Ptr != End && "dereferencing end() iterator")((void)0); |
| 1243 | if (shouldReverseIterate<KeyT>()) |
| 1244 | return &(Ptr[-1]); |
| 1245 | return Ptr; |
| 1246 | } |
| 1247 | |
| 1248 | friend bool operator==(const DenseMapIterator &LHS, |
| 1249 | const DenseMapIterator &RHS) { |
| 1250 | assert((!LHS.Ptr || LHS.isHandleInSync()) && "handle not in sync!")((void)0); |
| 1251 | assert((!RHS.Ptr || RHS.isHandleInSync()) && "handle not in sync!")((void)0); |
| 1252 | assert(LHS.getEpochAddress() == RHS.getEpochAddress() &&((void)0) |
| 1253 | "comparing incomparable iterators!")((void)0); |
| 1254 | return LHS.Ptr == RHS.Ptr; |
| 1255 | } |
| 1256 | |
| 1257 | friend bool operator!=(const DenseMapIterator &LHS, |
| 1258 | const DenseMapIterator &RHS) { |
| 1259 | return !(LHS == RHS); |
| 1260 | } |
| 1261 | |
| 1262 | inline DenseMapIterator& operator++() { // Preincrement |
| 1263 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
| 1264 | assert(Ptr != End && "incrementing end() iterator")((void)0); |
| 1265 | if (shouldReverseIterate<KeyT>()) { |
| 1266 | --Ptr; |
| 1267 | RetreatPastEmptyBuckets(); |
| 1268 | return *this; |
| 1269 | } |
| 1270 | ++Ptr; |
| 1271 | AdvancePastEmptyBuckets(); |
| 1272 | return *this; |
| 1273 | } |
| 1274 | DenseMapIterator operator++(int) { // Postincrement |
| 1275 | assert(isHandleInSync() && "invalid iterator access!")((void)0); |
| 1276 | DenseMapIterator tmp = *this; ++*this; return tmp; |
| 1277 | } |
| 1278 | |
| 1279 | private: |
| 1280 | void AdvancePastEmptyBuckets() { |
| 1281 | assert(Ptr <= End)((void)0); |
| 1282 | const KeyT Empty = KeyInfoT::getEmptyKey(); |
| 1283 | const KeyT Tombstone = KeyInfoT::getTombstoneKey(); |
| 1284 | |
| 1285 | while (Ptr != End && (KeyInfoT::isEqual(Ptr->getFirst(), Empty) || |
| 1286 | KeyInfoT::isEqual(Ptr->getFirst(), Tombstone))) |
| 1287 | ++Ptr; |
| 1288 | } |
| 1289 | |
| 1290 | void RetreatPastEmptyBuckets() { |
| 1291 | assert(Ptr >= End)((void)0); |
| 1292 | const KeyT Empty = KeyInfoT::getEmptyKey(); |
| 1293 | const KeyT Tombstone = KeyInfoT::getTombstoneKey(); |
| 1294 | |
| 1295 | while (Ptr != End && (KeyInfoT::isEqual(Ptr[-1].getFirst(), Empty) || |
| 1296 | KeyInfoT::isEqual(Ptr[-1].getFirst(), Tombstone))) |
| 1297 | --Ptr; |
| 1298 | } |
| 1299 | }; |
| 1300 | |
| 1301 | template <typename KeyT, typename ValueT, typename KeyInfoT> |
| 1302 | inline size_t capacity_in_bytes(const DenseMap<KeyT, ValueT, KeyInfoT> &X) { |
| 1303 | return X.getMemorySize(); |
| 1304 | } |
| 1305 | |
| 1306 | } // end namespace llvm |
| 1307 | |
| 1308 | #endif // LLVM_ADT_DENSEMAP_H |