| 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 | //===- SIFixSGPRCopies.cpp - Remove potential VGPR => SGPR 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 | /// \file | |||
| 10 | /// Copies from VGPR to SGPR registers are illegal and the register coalescer | |||
| 11 | /// will sometimes generate these illegal copies in situations like this: | |||
| 12 | /// | |||
| 13 | /// Register Class <vsrc> is the union of <vgpr> and <sgpr> | |||
| 14 | /// | |||
| 15 | /// BB0: | |||
| 16 | /// %0 <sgpr> = SCALAR_INST | |||
| 17 | /// %1 <vsrc> = COPY %0 <sgpr> | |||
| 18 | /// ... | |||
| 19 | /// BRANCH %cond BB1, BB2 | |||
| 20 | /// BB1: | |||
| 21 | /// %2 <vgpr> = VECTOR_INST | |||
| 22 | /// %3 <vsrc> = COPY %2 <vgpr> | |||
| 23 | /// BB2: | |||
| 24 | /// %4 <vsrc> = PHI %1 <vsrc>, <%bb.0>, %3 <vrsc>, <%bb.1> | |||
| 25 | /// %5 <vgpr> = VECTOR_INST %4 <vsrc> | |||
| 26 | /// | |||
| 27 | /// | |||
| 28 | /// The coalescer will begin at BB0 and eliminate its copy, then the resulting | |||
| 29 | /// code will look like this: | |||
| 30 | /// | |||
| 31 | /// BB0: | |||
| 32 | /// %0 <sgpr> = SCALAR_INST | |||
| 33 | /// ... | |||
| 34 | /// BRANCH %cond BB1, BB2 | |||
| 35 | /// BB1: | |||
| 36 | /// %2 <vgpr> = VECTOR_INST | |||
| 37 | /// %3 <vsrc> = COPY %2 <vgpr> | |||
| 38 | /// BB2: | |||
| 39 | /// %4 <sgpr> = PHI %0 <sgpr>, <%bb.0>, %3 <vsrc>, <%bb.1> | |||
| 40 | /// %5 <vgpr> = VECTOR_INST %4 <sgpr> | |||
| 41 | /// | |||
| 42 | /// Now that the result of the PHI instruction is an SGPR, the register | |||
| 43 | /// allocator is now forced to constrain the register class of %3 to | |||
| 44 | /// <sgpr> so we end up with final code like this: | |||
| 45 | /// | |||
| 46 | /// BB0: | |||
| 47 | /// %0 <sgpr> = SCALAR_INST | |||
| 48 | /// ... | |||
| 49 | /// BRANCH %cond BB1, BB2 | |||
| 50 | /// BB1: | |||
| 51 | /// %2 <vgpr> = VECTOR_INST | |||
| 52 | /// %3 <sgpr> = COPY %2 <vgpr> | |||
| 53 | /// BB2: | |||
| 54 | /// %4 <sgpr> = PHI %0 <sgpr>, <%bb.0>, %3 <sgpr>, <%bb.1> | |||
| 55 | /// %5 <vgpr> = VECTOR_INST %4 <sgpr> | |||
| 56 | /// | |||
| 57 | /// Now this code contains an illegal copy from a VGPR to an SGPR. | |||
| 58 | /// | |||
| 59 | /// In order to avoid this problem, this pass searches for PHI instructions | |||
| 60 | /// which define a <vsrc> register and constrains its definition class to | |||
| 61 | /// <vgpr> if the user of the PHI's definition register is a vector instruction. | |||
| 62 | /// If the PHI's definition class is constrained to <vgpr> then the coalescer | |||
| 63 | /// will be unable to perform the COPY removal from the above example which | |||
| 64 | /// ultimately led to the creation of an illegal COPY. | |||
| 65 | //===----------------------------------------------------------------------===// | |||
| 66 | ||||
| 67 | #include "AMDGPU.h" | |||
| 68 | #include "GCNSubtarget.h" | |||
| 69 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" | |||
| 70 | #include "llvm/CodeGen/MachineDominators.h" | |||
| 71 | #include "llvm/InitializePasses.h" | |||
| 72 | #include "llvm/Target/TargetMachine.h" | |||
| 73 | ||||
| 74 | using namespace llvm; | |||
| 75 | ||||
| 76 | #define DEBUG_TYPE"si-fix-sgpr-copies" "si-fix-sgpr-copies" | |||
| 77 | ||||
| 78 | static cl::opt<bool> EnableM0Merge( | |||
| 79 | "amdgpu-enable-merge-m0", | |||
| 80 | cl::desc("Merge and hoist M0 initializations"), | |||
| 81 | cl::init(true)); | |||
| 82 | ||||
| 83 | namespace { | |||
| 84 | ||||
| 85 | class SIFixSGPRCopies : public MachineFunctionPass { | |||
| 86 | MachineDominatorTree *MDT; | |||
| 87 | ||||
| 88 | public: | |||
| 89 | static char ID; | |||
| 90 | ||||
| 91 | MachineRegisterInfo *MRI; | |||
| 92 | const SIRegisterInfo *TRI; | |||
| 93 | const SIInstrInfo *TII; | |||
| 94 | ||||
| 95 | SIFixSGPRCopies() : MachineFunctionPass(ID) {} | |||
| 96 | ||||
| 97 | bool runOnMachineFunction(MachineFunction &MF) override; | |||
| 98 | ||||
| 99 | MachineBasicBlock *processPHINode(MachineInstr &MI); | |||
| 100 | ||||
| 101 | StringRef getPassName() const override { return "SI Fix SGPR copies"; } | |||
| 102 | ||||
| 103 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
| 104 | AU.addRequired<MachineDominatorTree>(); | |||
| 105 | AU.addPreserved<MachineDominatorTree>(); | |||
| 106 | AU.setPreservesCFG(); | |||
| 107 | MachineFunctionPass::getAnalysisUsage(AU); | |||
| 108 | } | |||
| 109 | }; | |||
| 110 | ||||
| 111 | } // end anonymous namespace | |||
| 112 | ||||
| 113 | INITIALIZE_PASS_BEGIN(SIFixSGPRCopies, DEBUG_TYPE,static void *initializeSIFixSGPRCopiesPassOnce(PassRegistry & Registry) { | |||
| 114 | "SI Fix SGPR copies", false, false)static void *initializeSIFixSGPRCopiesPassOnce(PassRegistry & Registry) { | |||
| 115 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)initializeMachineDominatorTreePass(Registry); | |||
| 116 | INITIALIZE_PASS_END(SIFixSGPRCopies, DEBUG_TYPE,PassInfo *PI = new PassInfo( "SI Fix SGPR copies", "si-fix-sgpr-copies" , &SIFixSGPRCopies::ID, PassInfo::NormalCtor_t(callDefaultCtor <SIFixSGPRCopies>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeSIFixSGPRCopiesPassFlag ; void llvm::initializeSIFixSGPRCopiesPass(PassRegistry & Registry) { llvm::call_once(InitializeSIFixSGPRCopiesPassFlag , initializeSIFixSGPRCopiesPassOnce, std::ref(Registry)); } | |||
| 117 | "SI Fix SGPR copies", false, false)PassInfo *PI = new PassInfo( "SI Fix SGPR copies", "si-fix-sgpr-copies" , &SIFixSGPRCopies::ID, PassInfo::NormalCtor_t(callDefaultCtor <SIFixSGPRCopies>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeSIFixSGPRCopiesPassFlag ; void llvm::initializeSIFixSGPRCopiesPass(PassRegistry & Registry) { llvm::call_once(InitializeSIFixSGPRCopiesPassFlag , initializeSIFixSGPRCopiesPassOnce, std::ref(Registry)); } | |||
| 118 | ||||
| 119 | char SIFixSGPRCopies::ID = 0; | |||
| 120 | ||||
| 121 | char &llvm::SIFixSGPRCopiesID = SIFixSGPRCopies::ID; | |||
| 122 | ||||
| 123 | FunctionPass *llvm::createSIFixSGPRCopiesPass() { | |||
| 124 | return new SIFixSGPRCopies(); | |||
| 125 | } | |||
| 126 | ||||
| 127 | static bool hasVectorOperands(const MachineInstr &MI, | |||
| 128 | const SIRegisterInfo *TRI) { | |||
| 129 | const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo(); | |||
| 130 | for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { | |||
| 131 | if (!MI.getOperand(i).isReg() || !MI.getOperand(i).getReg().isVirtual()) | |||
| 132 | continue; | |||
| 133 | ||||
| 134 | if (TRI->hasVectorRegisters(MRI.getRegClass(MI.getOperand(i).getReg()))) | |||
| 135 | return true; | |||
| 136 | } | |||
| 137 | return false; | |||
| 138 | } | |||
| 139 | ||||
| 140 | static std::pair<const TargetRegisterClass *, const TargetRegisterClass *> | |||
| 141 | getCopyRegClasses(const MachineInstr &Copy, | |||
| 142 | const SIRegisterInfo &TRI, | |||
| 143 | const MachineRegisterInfo &MRI) { | |||
| 144 | Register DstReg = Copy.getOperand(0).getReg(); | |||
| 145 | Register SrcReg = Copy.getOperand(1).getReg(); | |||
| 146 | ||||
| 147 | const TargetRegisterClass *SrcRC = SrcReg.isVirtual() | |||
| 148 | ? MRI.getRegClass(SrcReg) | |||
| 149 | : TRI.getPhysRegClass(SrcReg); | |||
| 150 | ||||
| 151 | // We don't really care about the subregister here. | |||
| 152 | // SrcRC = TRI.getSubRegClass(SrcRC, Copy.getOperand(1).getSubReg()); | |||
| 153 | ||||
| 154 | const TargetRegisterClass *DstRC = DstReg.isVirtual() | |||
| 155 | ? MRI.getRegClass(DstReg) | |||
| 156 | : TRI.getPhysRegClass(DstReg); | |||
| 157 | ||||
| 158 | return std::make_pair(SrcRC, DstRC); | |||
| 159 | } | |||
| 160 | ||||
| 161 | static bool isVGPRToSGPRCopy(const TargetRegisterClass *SrcRC, | |||
| 162 | const TargetRegisterClass *DstRC, | |||
| 163 | const SIRegisterInfo &TRI) { | |||
| 164 | return SrcRC != &AMDGPU::VReg_1RegClass && TRI.isSGPRClass(DstRC) && | |||
| 165 | TRI.hasVectorRegisters(SrcRC); | |||
| 166 | } | |||
| 167 | ||||
| 168 | static bool isSGPRToVGPRCopy(const TargetRegisterClass *SrcRC, | |||
| 169 | const TargetRegisterClass *DstRC, | |||
| 170 | const SIRegisterInfo &TRI) { | |||
| 171 | return DstRC != &AMDGPU::VReg_1RegClass && TRI.isSGPRClass(SrcRC) && | |||
| 172 | TRI.hasVectorRegisters(DstRC); | |||
| 173 | } | |||
| 174 | ||||
| 175 | static bool tryChangeVGPRtoSGPRinCopy(MachineInstr &MI, | |||
| 176 | const SIRegisterInfo *TRI, | |||
| 177 | const SIInstrInfo *TII) { | |||
| 178 | MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo(); | |||
| 179 | auto &Src = MI.getOperand(1); | |||
| 180 | Register DstReg = MI.getOperand(0).getReg(); | |||
| 181 | Register SrcReg = Src.getReg(); | |||
| 182 | if (!SrcReg.isVirtual() || !DstReg.isVirtual()) | |||
| 183 | return false; | |||
| 184 | ||||
| 185 | for (const auto &MO : MRI.reg_nodbg_operands(DstReg)) { | |||
| 186 | const auto *UseMI = MO.getParent(); | |||
| 187 | if (UseMI == &MI) | |||
| 188 | continue; | |||
| 189 | if (MO.isDef() || UseMI->getParent() != MI.getParent() || | |||
| 190 | UseMI->getOpcode() <= TargetOpcode::GENERIC_OP_END) | |||
| 191 | return false; | |||
| 192 | ||||
| 193 | unsigned OpIdx = UseMI->getOperandNo(&MO); | |||
| 194 | if (OpIdx >= UseMI->getDesc().getNumOperands() || | |||
| 195 | !TII->isOperandLegal(*UseMI, OpIdx, &Src)) | |||
| 196 | return false; | |||
| 197 | } | |||
| 198 | // Change VGPR to SGPR destination. | |||
| 199 | MRI.setRegClass(DstReg, TRI->getEquivalentSGPRClass(MRI.getRegClass(DstReg))); | |||
| 200 | return true; | |||
| 201 | } | |||
| 202 | ||||
| 203 | // Distribute an SGPR->VGPR copy of a REG_SEQUENCE into a VGPR REG_SEQUENCE. | |||
| 204 | // | |||
| 205 | // SGPRx = ... | |||
| 206 | // SGPRy = REG_SEQUENCE SGPRx, sub0 ... | |||
| 207 | // VGPRz = COPY SGPRy | |||
| 208 | // | |||
| 209 | // ==> | |||
| 210 | // | |||
| 211 | // VGPRx = COPY SGPRx | |||
| 212 | // VGPRz = REG_SEQUENCE VGPRx, sub0 | |||
| 213 | // | |||
| 214 | // This exposes immediate folding opportunities when materializing 64-bit | |||
| 215 | // immediates. | |||
| 216 | static bool foldVGPRCopyIntoRegSequence(MachineInstr &MI, | |||
| 217 | const SIRegisterInfo *TRI, | |||
| 218 | const SIInstrInfo *TII, | |||
| 219 | MachineRegisterInfo &MRI) { | |||
| 220 | assert(MI.isRegSequence())((void)0); | |||
| 221 | ||||
| 222 | Register DstReg = MI.getOperand(0).getReg(); | |||
| 223 | if (!TRI->isSGPRClass(MRI.getRegClass(DstReg))) | |||
| 224 | return false; | |||
| 225 | ||||
| 226 | if (!MRI.hasOneUse(DstReg)) | |||
| 227 | return false; | |||
| 228 | ||||
| 229 | MachineInstr &CopyUse = *MRI.use_instr_begin(DstReg); | |||
| 230 | if (!CopyUse.isCopy()) | |||
| 231 | return false; | |||
| 232 | ||||
| 233 | // It is illegal to have vreg inputs to a physreg defining reg_sequence. | |||
| 234 | if (CopyUse.getOperand(0).getReg().isPhysical()) | |||
| 235 | return false; | |||
| 236 | ||||
| 237 | const TargetRegisterClass *SrcRC, *DstRC; | |||
| 238 | std::tie(SrcRC, DstRC) = getCopyRegClasses(CopyUse, *TRI, MRI); | |||
| 239 | ||||
| 240 | if (!isSGPRToVGPRCopy(SrcRC, DstRC, *TRI)) | |||
| 241 | return false; | |||
| 242 | ||||
| 243 | if (tryChangeVGPRtoSGPRinCopy(CopyUse, TRI, TII)) | |||
| 244 | return true; | |||
| 245 | ||||
| 246 | // TODO: Could have multiple extracts? | |||
| 247 | unsigned SubReg = CopyUse.getOperand(1).getSubReg(); | |||
| 248 | if (SubReg != AMDGPU::NoSubRegister) | |||
| 249 | return false; | |||
| 250 | ||||
| 251 | MRI.setRegClass(DstReg, DstRC); | |||
| 252 | ||||
| 253 | // SGPRx = ... | |||
| 254 | // SGPRy = REG_SEQUENCE SGPRx, sub0 ... | |||
| 255 | // VGPRz = COPY SGPRy | |||
| 256 | ||||
| 257 | // => | |||
| 258 | // VGPRx = COPY SGPRx | |||
| 259 | // VGPRz = REG_SEQUENCE VGPRx, sub0 | |||
| 260 | ||||
| 261 | MI.getOperand(0).setReg(CopyUse.getOperand(0).getReg()); | |||
| 262 | bool IsAGPR = TRI->hasAGPRs(DstRC); | |||
| 263 | ||||
| 264 | for (unsigned I = 1, N = MI.getNumOperands(); I != N; I += 2) { | |||
| 265 | Register SrcReg = MI.getOperand(I).getReg(); | |||
| 266 | unsigned SrcSubReg = MI.getOperand(I).getSubReg(); | |||
| 267 | ||||
| 268 | const TargetRegisterClass *SrcRC = MRI.getRegClass(SrcReg); | |||
| 269 | assert(TRI->isSGPRClass(SrcRC) &&((void)0) | |||
| 270 | "Expected SGPR REG_SEQUENCE to only have SGPR inputs")((void)0); | |||
| 271 | ||||
| 272 | SrcRC = TRI->getSubRegClass(SrcRC, SrcSubReg); | |||
| 273 | const TargetRegisterClass *NewSrcRC = TRI->getEquivalentVGPRClass(SrcRC); | |||
| 274 | ||||
| 275 | Register TmpReg = MRI.createVirtualRegister(NewSrcRC); | |||
| 276 | ||||
| 277 | BuildMI(*MI.getParent(), &MI, MI.getDebugLoc(), TII->get(AMDGPU::COPY), | |||
| 278 | TmpReg) | |||
| 279 | .add(MI.getOperand(I)); | |||
| 280 | ||||
| 281 | if (IsAGPR) { | |||
| 282 | const TargetRegisterClass *NewSrcRC = TRI->getEquivalentAGPRClass(SrcRC); | |||
| 283 | Register TmpAReg = MRI.createVirtualRegister(NewSrcRC); | |||
| 284 | unsigned Opc = NewSrcRC == &AMDGPU::AGPR_32RegClass ? | |||
| 285 | AMDGPU::V_ACCVGPR_WRITE_B32_e64 : AMDGPU::COPY; | |||
| 286 | BuildMI(*MI.getParent(), &MI, MI.getDebugLoc(), TII->get(Opc), | |||
| 287 | TmpAReg) | |||
| 288 | .addReg(TmpReg, RegState::Kill); | |||
| 289 | TmpReg = TmpAReg; | |||
| 290 | } | |||
| 291 | ||||
| 292 | MI.getOperand(I).setReg(TmpReg); | |||
| 293 | } | |||
| 294 | ||||
| 295 | CopyUse.eraseFromParent(); | |||
| 296 | return true; | |||
| 297 | } | |||
| 298 | ||||
| 299 | static bool isSafeToFoldImmIntoCopy(const MachineInstr *Copy, | |||
| 300 | const MachineInstr *MoveImm, | |||
| 301 | const SIInstrInfo *TII, | |||
| 302 | unsigned &SMovOp, | |||
| 303 | int64_t &Imm) { | |||
| 304 | if (Copy->getOpcode() != AMDGPU::COPY) | |||
| 305 | return false; | |||
| 306 | ||||
| 307 | if (!MoveImm->isMoveImmediate()) | |||
| 308 | return false; | |||
| 309 | ||||
| 310 | const MachineOperand *ImmOp = | |||
| 311 | TII->getNamedOperand(*MoveImm, AMDGPU::OpName::src0); | |||
| 312 | if (!ImmOp->isImm()) | |||
| 313 | return false; | |||
| 314 | ||||
| 315 | // FIXME: Handle copies with sub-regs. | |||
| 316 | if (Copy->getOperand(0).getSubReg()) | |||
| 317 | return false; | |||
| 318 | ||||
| 319 | switch (MoveImm->getOpcode()) { | |||
| 320 | default: | |||
| 321 | return false; | |||
| 322 | case AMDGPU::V_MOV_B32_e32: | |||
| 323 | SMovOp = AMDGPU::S_MOV_B32; | |||
| 324 | break; | |||
| 325 | case AMDGPU::V_MOV_B64_PSEUDO: | |||
| 326 | SMovOp = AMDGPU::S_MOV_B64; | |||
| 327 | break; | |||
| 328 | } | |||
| 329 | Imm = ImmOp->getImm(); | |||
| 330 | return true; | |||
| 331 | } | |||
| 332 | ||||
| 333 | template <class UnaryPredicate> | |||
| 334 | bool searchPredecessors(const MachineBasicBlock *MBB, | |||
| 335 | const MachineBasicBlock *CutOff, | |||
| 336 | UnaryPredicate Predicate) { | |||
| 337 | if (MBB == CutOff) | |||
| 338 | return false; | |||
| 339 | ||||
| 340 | DenseSet<const MachineBasicBlock *> Visited; | |||
| 341 | SmallVector<MachineBasicBlock *, 4> Worklist(MBB->predecessors()); | |||
| 342 | ||||
| 343 | while (!Worklist.empty()) { | |||
| 344 | MachineBasicBlock *MBB = Worklist.pop_back_val(); | |||
| 345 | ||||
| 346 | if (!Visited.insert(MBB).second) | |||
| 347 | continue; | |||
| 348 | if (MBB == CutOff) | |||
| 349 | continue; | |||
| 350 | if (Predicate(MBB)) | |||
| 351 | return true; | |||
| 352 | ||||
| 353 | Worklist.append(MBB->pred_begin(), MBB->pred_end()); | |||
| 354 | } | |||
| 355 | ||||
| 356 | return false; | |||
| 357 | } | |||
| 358 | ||||
| 359 | // Checks if there is potential path From instruction To instruction. | |||
| 360 | // If CutOff is specified and it sits in between of that path we ignore | |||
| 361 | // a higher portion of the path and report it is not reachable. | |||
| 362 | static bool isReachable(const MachineInstr *From, | |||
| 363 | const MachineInstr *To, | |||
| 364 | const MachineBasicBlock *CutOff, | |||
| 365 | MachineDominatorTree &MDT) { | |||
| 366 | if (MDT.dominates(From, To)) | |||
| 367 | return true; | |||
| 368 | ||||
| 369 | const MachineBasicBlock *MBBFrom = From->getParent(); | |||
| 370 | const MachineBasicBlock *MBBTo = To->getParent(); | |||
| 371 | ||||
| 372 | // Do predecessor search. | |||
| 373 | // We should almost never get here since we do not usually produce M0 stores | |||
| 374 | // other than -1. | |||
| 375 | return searchPredecessors(MBBTo, CutOff, [MBBFrom] | |||
| 376 | (const MachineBasicBlock *MBB) { return MBB == MBBFrom; }); | |||
| 377 | } | |||
| 378 | ||||
| 379 | // Return the first non-prologue instruction in the block. | |||
| 380 | static MachineBasicBlock::iterator | |||
| 381 | getFirstNonPrologue(MachineBasicBlock *MBB, const TargetInstrInfo *TII) { | |||
| 382 | MachineBasicBlock::iterator I = MBB->getFirstNonPHI(); | |||
| 383 | while (I != MBB->end() && TII->isBasicBlockPrologue(*I)) | |||
| 384 | ++I; | |||
| 385 | ||||
| 386 | return I; | |||
| 387 | } | |||
| 388 | ||||
| 389 | // Hoist and merge identical SGPR initializations into a common predecessor. | |||
| 390 | // This is intended to combine M0 initializations, but can work with any | |||
| 391 | // SGPR. A VGPR cannot be processed since we cannot guarantee vector | |||
| 392 | // executioon. | |||
| 393 | static bool hoistAndMergeSGPRInits(unsigned Reg, | |||
| 394 | const MachineRegisterInfo &MRI, | |||
| 395 | const TargetRegisterInfo *TRI, | |||
| 396 | MachineDominatorTree &MDT, | |||
| 397 | const TargetInstrInfo *TII) { | |||
| 398 | // List of inits by immediate value. | |||
| 399 | using InitListMap = std::map<unsigned, std::list<MachineInstr *>>; | |||
| 400 | InitListMap Inits; | |||
| 401 | // List of clobbering instructions. | |||
| 402 | SmallVector<MachineInstr*, 8> Clobbers; | |||
| 403 | // List of instructions marked for deletion. | |||
| 404 | SmallSet<MachineInstr*, 8> MergedInstrs; | |||
| 405 | ||||
| 406 | bool Changed = false; | |||
| 407 | ||||
| 408 | for (auto &MI : MRI.def_instructions(Reg)) { | |||
| 409 | MachineOperand *Imm = nullptr; | |||
| 410 | for (auto &MO : MI.operands()) { | |||
| 411 | if ((MO.isReg() && ((MO.isDef() && MO.getReg() != Reg) || !MO.isDef())) || | |||
| 412 | (!MO.isImm() && !MO.isReg()) || (MO.isImm() && Imm)) { | |||
| 413 | Imm = nullptr; | |||
| 414 | break; | |||
| 415 | } else if (MO.isImm()) | |||
| 416 | Imm = &MO; | |||
| 417 | } | |||
| 418 | if (Imm) | |||
| 419 | Inits[Imm->getImm()].push_front(&MI); | |||
| 420 | else | |||
| 421 | Clobbers.push_back(&MI); | |||
| 422 | } | |||
| 423 | ||||
| 424 | for (auto &Init : Inits) { | |||
| 425 | auto &Defs = Init.second; | |||
| 426 | ||||
| 427 | for (auto I1 = Defs.begin(), E = Defs.end(); I1 != E; ) { | |||
| 428 | MachineInstr *MI1 = *I1; | |||
| 429 | ||||
| 430 | for (auto I2 = std::next(I1); I2 != E; ) { | |||
| 431 | MachineInstr *MI2 = *I2; | |||
| 432 | ||||
| 433 | // Check any possible interference | |||
| 434 | auto interferes = [&](MachineBasicBlock::iterator From, | |||
| 435 | MachineBasicBlock::iterator To) -> bool { | |||
| 436 | ||||
| 437 | assert(MDT.dominates(&*To, &*From))((void)0); | |||
| 438 | ||||
| 439 | auto interferes = [&MDT, From, To](MachineInstr* &Clobber) -> bool { | |||
| 440 | const MachineBasicBlock *MBBFrom = From->getParent(); | |||
| 441 | const MachineBasicBlock *MBBTo = To->getParent(); | |||
| 442 | bool MayClobberFrom = isReachable(Clobber, &*From, MBBTo, MDT); | |||
| 443 | bool MayClobberTo = isReachable(Clobber, &*To, MBBTo, MDT); | |||
| 444 | if (!MayClobberFrom && !MayClobberTo) | |||
| 445 | return false; | |||
| 446 | if ((MayClobberFrom && !MayClobberTo) || | |||
| 447 | (!MayClobberFrom && MayClobberTo)) | |||
| 448 | return true; | |||
| 449 | // Both can clobber, this is not an interference only if both are | |||
| 450 | // dominated by Clobber and belong to the same block or if Clobber | |||
| 451 | // properly dominates To, given that To >> From, so it dominates | |||
| 452 | // both and located in a common dominator. | |||
| 453 | return !((MBBFrom == MBBTo && | |||
| 454 | MDT.dominates(Clobber, &*From) && | |||
| 455 | MDT.dominates(Clobber, &*To)) || | |||
| 456 | MDT.properlyDominates(Clobber->getParent(), MBBTo)); | |||
| 457 | }; | |||
| 458 | ||||
| 459 | return (llvm::any_of(Clobbers, interferes)) || | |||
| 460 | (llvm::any_of(Inits, [&](InitListMap::value_type &C) { | |||
| 461 | return C.first != Init.first && | |||
| 462 | llvm::any_of(C.second, interferes); | |||
| 463 | })); | |||
| 464 | }; | |||
| 465 | ||||
| 466 | if (MDT.dominates(MI1, MI2)) { | |||
| 467 | if (!interferes(MI2, MI1)) { | |||
| 468 | LLVM_DEBUG(dbgs()do { } while (false) | |||
| 469 | << "Erasing from "do { } while (false) | |||
| 470 | << printMBBReference(*MI2->getParent()) << " " << *MI2)do { } while (false); | |||
| 471 | MergedInstrs.insert(MI2); | |||
| 472 | Changed = true; | |||
| 473 | ++I2; | |||
| 474 | continue; | |||
| 475 | } | |||
| 476 | } else if (MDT.dominates(MI2, MI1)) { | |||
| 477 | if (!interferes(MI1, MI2)) { | |||
| 478 | LLVM_DEBUG(dbgs()do { } while (false) | |||
| 479 | << "Erasing from "do { } while (false) | |||
| 480 | << printMBBReference(*MI1->getParent()) << " " << *MI1)do { } while (false); | |||
| 481 | MergedInstrs.insert(MI1); | |||
| 482 | Changed = true; | |||
| 483 | ++I1; | |||
| 484 | break; | |||
| 485 | } | |||
| 486 | } else { | |||
| 487 | auto *MBB = MDT.findNearestCommonDominator(MI1->getParent(), | |||
| 488 | MI2->getParent()); | |||
| 489 | if (!MBB) { | |||
| 490 | ++I2; | |||
| 491 | continue; | |||
| 492 | } | |||
| 493 | ||||
| 494 | MachineBasicBlock::iterator I = getFirstNonPrologue(MBB, TII); | |||
| 495 | if (!interferes(MI1, I) && !interferes(MI2, I)) { | |||
| 496 | LLVM_DEBUG(dbgs()do { } while (false) | |||
| 497 | << "Erasing from "do { } while (false) | |||
| 498 | << printMBBReference(*MI1->getParent()) << " " << *MI1do { } while (false) | |||
| 499 | << "and moving from "do { } while (false) | |||
| 500 | << printMBBReference(*MI2->getParent()) << " to "do { } while (false) | |||
| 501 | << printMBBReference(*I->getParent()) << " " << *MI2)do { } while (false); | |||
| 502 | I->getParent()->splice(I, MI2->getParent(), MI2); | |||
| 503 | MergedInstrs.insert(MI1); | |||
| 504 | Changed = true; | |||
| 505 | ++I1; | |||
| 506 | break; | |||
| 507 | } | |||
| 508 | } | |||
| 509 | ++I2; | |||
| 510 | } | |||
| 511 | ++I1; | |||
| 512 | } | |||
| 513 | } | |||
| 514 | ||||
| 515 | // Remove initializations that were merged into another. | |||
| 516 | for (auto &Init : Inits) { | |||
| 517 | auto &Defs = Init.second; | |||
| 518 | auto I = Defs.begin(); | |||
| 519 | while (I != Defs.end()) { | |||
| 520 | if (MergedInstrs.count(*I)) { | |||
| 521 | (*I)->eraseFromParent(); | |||
| 522 | I = Defs.erase(I); | |||
| 523 | } else | |||
| 524 | ++I; | |||
| 525 | } | |||
| 526 | } | |||
| 527 | ||||
| 528 | // Try to schedule SGPR initializations as early as possible in the MBB. | |||
| 529 | for (auto &Init : Inits) { | |||
| 530 | auto &Defs = Init.second; | |||
| 531 | for (auto MI : Defs) { | |||
| 532 | auto MBB = MI->getParent(); | |||
| 533 | MachineInstr &BoundaryMI = *getFirstNonPrologue(MBB, TII); | |||
| 534 | MachineBasicBlock::reverse_iterator B(BoundaryMI); | |||
| 535 | // Check if B should actually be a boundary. If not set the previous | |||
| 536 | // instruction as the boundary instead. | |||
| 537 | if (!TII->isBasicBlockPrologue(*B)) | |||
| 538 | B++; | |||
| 539 | ||||
| 540 | auto R = std::next(MI->getReverseIterator()); | |||
| 541 | const unsigned Threshold = 50; | |||
| 542 | // Search until B or Threshold for a place to insert the initialization. | |||
| 543 | for (unsigned I = 0; R != B && I < Threshold; ++R, ++I) | |||
| 544 | if (R->readsRegister(Reg, TRI) || R->definesRegister(Reg, TRI) || | |||
| 545 | TII->isSchedulingBoundary(*R, MBB, *MBB->getParent())) | |||
| 546 | break; | |||
| 547 | ||||
| 548 | // Move to directly after R. | |||
| 549 | if (&*--R != MI) | |||
| 550 | MBB->splice(*R, MBB, MI); | |||
| 551 | } | |||
| 552 | } | |||
| 553 | ||||
| 554 | if (Changed) | |||
| 555 | MRI.clearKillFlags(Reg); | |||
| 556 | ||||
| 557 | return Changed; | |||
| 558 | } | |||
| 559 | ||||
| 560 | bool SIFixSGPRCopies::runOnMachineFunction(MachineFunction &MF) { | |||
| 561 | // Only need to run this in SelectionDAG path. | |||
| 562 | if (MF.getProperties().hasProperty( | |||
| ||||
| 563 | MachineFunctionProperties::Property::Selected)) | |||
| 564 | return false; | |||
| 565 | ||||
| 566 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); | |||
| 567 | MRI = &MF.getRegInfo(); | |||
| 568 | TRI = ST.getRegisterInfo(); | |||
| 569 | TII = ST.getInstrInfo(); | |||
| 570 | MDT = &getAnalysis<MachineDominatorTree>(); | |||
| 571 | ||||
| 572 | for (MachineFunction::iterator BI = MF.begin(), BE = MF.end(); | |||
| 573 | BI != BE; ++BI) { | |||
| 574 | MachineBasicBlock *MBB = &*BI; | |||
| 575 | for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; | |||
| 576 | ++I) { | |||
| 577 | MachineInstr &MI = *I; | |||
| 578 | ||||
| 579 | switch (MI.getOpcode()) { | |||
| 580 | default: | |||
| 581 | continue; | |||
| 582 | case AMDGPU::COPY: | |||
| 583 | case AMDGPU::WQM: | |||
| 584 | case AMDGPU::STRICT_WQM: | |||
| 585 | case AMDGPU::SOFT_WQM: | |||
| 586 | case AMDGPU::STRICT_WWM: { | |||
| 587 | Register DstReg = MI.getOperand(0).getReg(); | |||
| 588 | ||||
| 589 | const TargetRegisterClass *SrcRC, *DstRC; | |||
| 590 | std::tie(SrcRC, DstRC) = getCopyRegClasses(MI, *TRI, *MRI); | |||
| 591 | ||||
| 592 | if (!DstReg.isVirtual()) { | |||
| 593 | // If the destination register is a physical register there isn't | |||
| 594 | // really much we can do to fix this. | |||
| 595 | // Some special instructions use M0 as an input. Some even only use | |||
| 596 | // the first lane. Insert a readfirstlane and hope for the best. | |||
| 597 | if (DstReg == AMDGPU::M0 && TRI->hasVectorRegisters(SrcRC)) { | |||
| 598 | Register TmpReg | |||
| 599 | = MRI->createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass); | |||
| 600 | ||||
| 601 | BuildMI(*MBB, MI, MI.getDebugLoc(), | |||
| 602 | TII->get(AMDGPU::V_READFIRSTLANE_B32), TmpReg) | |||
| 603 | .add(MI.getOperand(1)); | |||
| 604 | MI.getOperand(1).setReg(TmpReg); | |||
| 605 | } | |||
| 606 | ||||
| 607 | continue; | |||
| 608 | } | |||
| 609 | ||||
| 610 | if (isVGPRToSGPRCopy(SrcRC, DstRC, *TRI)) { | |||
| 611 | Register SrcReg = MI.getOperand(1).getReg(); | |||
| 612 | if (!SrcReg.isVirtual()) { | |||
| 613 | MachineBasicBlock *NewBB = TII->moveToVALU(MI, MDT); | |||
| 614 | if (NewBB && NewBB != MBB) { | |||
| 615 | MBB = NewBB; | |||
| 616 | E = MBB->end(); | |||
| 617 | BI = MachineFunction::iterator(MBB); | |||
| 618 | BE = MF.end(); | |||
| 619 | } | |||
| 620 | assert((!NewBB || NewBB == I->getParent()) &&((void)0) | |||
| 621 | "moveToVALU did not return the right basic block")((void)0); | |||
| 622 | break; | |||
| 623 | } | |||
| 624 | ||||
| 625 | MachineInstr *DefMI = MRI->getVRegDef(SrcReg); | |||
| 626 | unsigned SMovOp; | |||
| 627 | int64_t Imm; | |||
| 628 | // If we are just copying an immediate, we can replace the copy with | |||
| 629 | // s_mov_b32. | |||
| 630 | if (isSafeToFoldImmIntoCopy(&MI, DefMI, TII, SMovOp, Imm)) { | |||
| 631 | MI.getOperand(1).ChangeToImmediate(Imm); | |||
| 632 | MI.addImplicitDefUseOperands(MF); | |||
| 633 | MI.setDesc(TII->get(SMovOp)); | |||
| 634 | break; | |||
| 635 | } | |||
| 636 | MachineBasicBlock *NewBB = TII->moveToVALU(MI, MDT); | |||
| 637 | if (NewBB && NewBB != MBB) { | |||
| 638 | MBB = NewBB; | |||
| 639 | E = MBB->end(); | |||
| 640 | BI = MachineFunction::iterator(MBB); | |||
| 641 | BE = MF.end(); | |||
| 642 | } | |||
| 643 | assert((!NewBB || NewBB == I->getParent()) &&((void)0) | |||
| 644 | "moveToVALU did not return the right basic block")((void)0); | |||
| 645 | } else if (isSGPRToVGPRCopy(SrcRC, DstRC, *TRI)) { | |||
| 646 | tryChangeVGPRtoSGPRinCopy(MI, TRI, TII); | |||
| 647 | } | |||
| 648 | ||||
| 649 | break; | |||
| 650 | } | |||
| 651 | case AMDGPU::PHI: { | |||
| 652 | MachineBasicBlock *NewBB = processPHINode(MI); | |||
| 653 | if (NewBB && NewBB != MBB) { | |||
| 654 | MBB = NewBB; | |||
| 655 | E = MBB->end(); | |||
| 656 | BI = MachineFunction::iterator(MBB); | |||
| 657 | BE = MF.end(); | |||
| 658 | } | |||
| 659 | assert((!NewBB || NewBB == I->getParent()) &&((void)0) | |||
| 660 | "moveToVALU did not return the right basic block")((void)0); | |||
| 661 | break; | |||
| 662 | } | |||
| 663 | case AMDGPU::REG_SEQUENCE: { | |||
| 664 | if (TRI->hasVectorRegisters(TII->getOpRegClass(MI, 0)) || | |||
| 665 | !hasVectorOperands(MI, TRI)) { | |||
| 666 | foldVGPRCopyIntoRegSequence(MI, TRI, TII, *MRI); | |||
| 667 | continue; | |||
| 668 | } | |||
| 669 | ||||
| 670 | LLVM_DEBUG(dbgs() << "Fixing REG_SEQUENCE: " << MI)do { } while (false); | |||
| 671 | ||||
| 672 | MachineBasicBlock *NewBB = TII->moveToVALU(MI, MDT); | |||
| 673 | if (NewBB && NewBB != MBB) { | |||
| 674 | MBB = NewBB; | |||
| 675 | E = MBB->end(); | |||
| 676 | BI = MachineFunction::iterator(MBB); | |||
| 677 | BE = MF.end(); | |||
| 678 | } | |||
| 679 | assert((!NewBB || NewBB == I->getParent()) &&((void)0) | |||
| 680 | "moveToVALU did not return the right basic block")((void)0); | |||
| 681 | break; | |||
| 682 | } | |||
| 683 | case AMDGPU::INSERT_SUBREG: { | |||
| 684 | const TargetRegisterClass *DstRC, *Src0RC, *Src1RC; | |||
| 685 | DstRC = MRI->getRegClass(MI.getOperand(0).getReg()); | |||
| 686 | Src0RC = MRI->getRegClass(MI.getOperand(1).getReg()); | |||
| 687 | Src1RC = MRI->getRegClass(MI.getOperand(2).getReg()); | |||
| 688 | if (TRI->isSGPRClass(DstRC) && | |||
| 689 | (TRI->hasVectorRegisters(Src0RC) || | |||
| 690 | TRI->hasVectorRegisters(Src1RC))) { | |||
| 691 | LLVM_DEBUG(dbgs() << " Fixing INSERT_SUBREG: " << MI)do { } while (false); | |||
| 692 | MachineBasicBlock *NewBB = TII->moveToVALU(MI, MDT); | |||
| 693 | if (NewBB && NewBB != MBB) { | |||
| 694 | MBB = NewBB; | |||
| 695 | E = MBB->end(); | |||
| 696 | BI = MachineFunction::iterator(MBB); | |||
| 697 | BE = MF.end(); | |||
| 698 | } | |||
| 699 | assert((!NewBB || NewBB == I->getParent()) &&((void)0) | |||
| 700 | "moveToVALU did not return the right basic block")((void)0); | |||
| 701 | } | |||
| 702 | break; | |||
| 703 | } | |||
| 704 | case AMDGPU::V_WRITELANE_B32: { | |||
| 705 | // Some architectures allow more than one constant bus access without | |||
| 706 | // SGPR restriction | |||
| 707 | if (ST.getConstantBusLimit(MI.getOpcode()) != 1) | |||
| 708 | break; | |||
| 709 | ||||
| 710 | // Writelane is special in that it can use SGPR and M0 (which would | |||
| 711 | // normally count as using the constant bus twice - but in this case it | |||
| 712 | // is allowed since the lane selector doesn't count as a use of the | |||
| 713 | // constant bus). However, it is still required to abide by the 1 SGPR | |||
| 714 | // rule. Apply a fix here as we might have multiple SGPRs after | |||
| 715 | // legalizing VGPRs to SGPRs | |||
| 716 | int Src0Idx = | |||
| 717 | AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::src0); | |||
| 718 | int Src1Idx = | |||
| 719 | AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::src1); | |||
| 720 | MachineOperand &Src0 = MI.getOperand(Src0Idx); | |||
| 721 | MachineOperand &Src1 = MI.getOperand(Src1Idx); | |||
| 722 | ||||
| 723 | // Check to see if the instruction violates the 1 SGPR rule | |||
| 724 | if ((Src0.isReg() && TRI->isSGPRReg(*MRI, Src0.getReg()) && | |||
| 725 | Src0.getReg() != AMDGPU::M0) && | |||
| 726 | (Src1.isReg() && TRI->isSGPRReg(*MRI, Src1.getReg()) && | |||
| 727 | Src1.getReg() != AMDGPU::M0)) { | |||
| 728 | ||||
| 729 | // Check for trivially easy constant prop into one of the operands | |||
| 730 | // If this is the case then perform the operation now to resolve SGPR | |||
| 731 | // issue. If we don't do that here we will always insert a mov to m0 | |||
| 732 | // that can't be resolved in later operand folding pass | |||
| 733 | bool Resolved = false; | |||
| 734 | for (MachineOperand *MO : {&Src0, &Src1}) { | |||
| 735 | if (MO->getReg().isVirtual()) { | |||
| 736 | MachineInstr *DefMI = MRI->getVRegDef(MO->getReg()); | |||
| 737 | if (DefMI && TII->isFoldableCopy(*DefMI)) { | |||
| 738 | const MachineOperand &Def = DefMI->getOperand(0); | |||
| 739 | if (Def.isReg() && | |||
| 740 | MO->getReg() == Def.getReg() && | |||
| 741 | MO->getSubReg() == Def.getSubReg()) { | |||
| 742 | const MachineOperand &Copied = DefMI->getOperand(1); | |||
| 743 | if (Copied.isImm() && | |||
| 744 | TII->isInlineConstant(APInt(64, Copied.getImm(), true))) { | |||
| 745 | MO->ChangeToImmediate(Copied.getImm()); | |||
| 746 | Resolved = true; | |||
| 747 | break; | |||
| 748 | } | |||
| 749 | } | |||
| 750 | } | |||
| 751 | } | |||
| 752 | } | |||
| 753 | ||||
| 754 | if (!Resolved) { | |||
| 755 | // Haven't managed to resolve by replacing an SGPR with an immediate | |||
| 756 | // Move src1 to be in M0 | |||
| 757 | BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), | |||
| 758 | TII->get(AMDGPU::COPY), AMDGPU::M0) | |||
| 759 | .add(Src1); | |||
| 760 | Src1.ChangeToRegister(AMDGPU::M0, false); | |||
| 761 | } | |||
| 762 | } | |||
| 763 | break; | |||
| 764 | } | |||
| 765 | } | |||
| 766 | } | |||
| 767 | } | |||
| 768 | ||||
| 769 | if (MF.getTarget().getOptLevel() > CodeGenOpt::None && EnableM0Merge) | |||
| 770 | hoistAndMergeSGPRInits(AMDGPU::M0, *MRI, TRI, *MDT, TII); | |||
| 771 | ||||
| 772 | return true; | |||
| 773 | } | |||
| 774 | ||||
| 775 | MachineBasicBlock *SIFixSGPRCopies::processPHINode(MachineInstr &MI) { | |||
| 776 | unsigned numVGPRUses = 0; | |||
| 777 | bool AllAGPRUses = true; | |||
| 778 | SetVector<const MachineInstr *> worklist; | |||
| 779 | SmallSet<const MachineInstr *, 4> Visited; | |||
| 780 | SetVector<MachineInstr *> PHIOperands; | |||
| 781 | MachineBasicBlock *CreatedBB = nullptr; | |||
| 782 | worklist.insert(&MI); | |||
| 783 | Visited.insert(&MI); | |||
| 784 | while (!worklist.empty()) { | |||
| 785 | const MachineInstr *Instr = worklist.pop_back_val(); | |||
| 786 | Register Reg = Instr->getOperand(0).getReg(); | |||
| 787 | for (const auto &Use : MRI->use_operands(Reg)) { | |||
| 788 | const MachineInstr *UseMI = Use.getParent(); | |||
| 789 | AllAGPRUses &= (UseMI->isCopy() && | |||
| 790 | TRI->isAGPR(*MRI, UseMI->getOperand(0).getReg())) || | |||
| 791 | TRI->isAGPR(*MRI, Use.getReg()); | |||
| 792 | if (UseMI->isCopy() || UseMI->isRegSequence()) { | |||
| 793 | if (UseMI->isCopy() && | |||
| 794 | UseMI->getOperand(0).getReg().isPhysical() && | |||
| 795 | !TRI->isSGPRReg(*MRI, UseMI->getOperand(0).getReg())) { | |||
| 796 | numVGPRUses++; | |||
| 797 | } | |||
| 798 | if (Visited.insert(UseMI).second) | |||
| 799 | worklist.insert(UseMI); | |||
| 800 | ||||
| 801 | continue; | |||
| 802 | } | |||
| 803 | ||||
| 804 | if (UseMI->isPHI()) { | |||
| 805 | const TargetRegisterClass *UseRC = MRI->getRegClass(Use.getReg()); | |||
| 806 | if (!TRI->isSGPRReg(*MRI, Use.getReg()) && | |||
| 807 | UseRC != &AMDGPU::VReg_1RegClass) | |||
| 808 | numVGPRUses++; | |||
| 809 | continue; | |||
| 810 | } | |||
| 811 | ||||
| 812 | const TargetRegisterClass *OpRC = | |||
| 813 | TII->getOpRegClass(*UseMI, UseMI->getOperandNo(&Use)); | |||
| 814 | if (!TRI->isSGPRClass(OpRC) && OpRC != &AMDGPU::VS_32RegClass && | |||
| 815 | OpRC != &AMDGPU::VS_64RegClass) { | |||
| 816 | numVGPRUses++; | |||
| 817 | } | |||
| 818 | } | |||
| 819 | } | |||
| 820 | ||||
| 821 | Register PHIRes = MI.getOperand(0).getReg(); | |||
| 822 | const TargetRegisterClass *RC0 = MRI->getRegClass(PHIRes); | |||
| 823 | if (AllAGPRUses && numVGPRUses && !TRI->hasAGPRs(RC0)) { | |||
| 824 | LLVM_DEBUG(dbgs() << "Moving PHI to AGPR: " << MI)do { } while (false); | |||
| 825 | MRI->setRegClass(PHIRes, TRI->getEquivalentAGPRClass(RC0)); | |||
| 826 | for (unsigned I = 1, N = MI.getNumOperands(); I != N; I += 2) { | |||
| 827 | MachineInstr *DefMI = MRI->getVRegDef(MI.getOperand(I).getReg()); | |||
| 828 | if (DefMI && DefMI->isPHI()) | |||
| 829 | PHIOperands.insert(DefMI); | |||
| 830 | } | |||
| 831 | } | |||
| 832 | ||||
| 833 | bool hasVGPRInput = false; | |||
| 834 | for (unsigned i = 1; i < MI.getNumOperands(); i += 2) { | |||
| 835 | Register InputReg = MI.getOperand(i).getReg(); | |||
| 836 | MachineInstr *Def = MRI->getVRegDef(InputReg); | |||
| 837 | if (TRI->isVectorRegister(*MRI, InputReg)) { | |||
| 838 | if (Def->isCopy()) { | |||
| 839 | Register SrcReg = Def->getOperand(1).getReg(); | |||
| 840 | const TargetRegisterClass *RC = | |||
| 841 | TRI->getRegClassForReg(*MRI, SrcReg); | |||
| 842 | if (TRI->isSGPRClass(RC)) | |||
| 843 | continue; | |||
| 844 | } | |||
| 845 | hasVGPRInput = true; | |||
| 846 | break; | |||
| 847 | } | |||
| 848 | else if (Def->isCopy() && | |||
| 849 | TRI->isVectorRegister(*MRI, Def->getOperand(1).getReg())) { | |||
| 850 | Register SrcReg = Def->getOperand(1).getReg(); | |||
| 851 | MachineInstr *SrcDef = MRI->getVRegDef(SrcReg); | |||
| 852 | unsigned SMovOp; | |||
| 853 | int64_t Imm; | |||
| 854 | if (!isSafeToFoldImmIntoCopy(Def, SrcDef, TII, SMovOp, Imm)) { | |||
| 855 | hasVGPRInput = true; | |||
| 856 | break; | |||
| 857 | } else { | |||
| 858 | // Formally, if we did not do this right away | |||
| 859 | // it would be done on the next iteration of the | |||
| 860 | // runOnMachineFunction main loop. But why not if we can? | |||
| 861 | MachineFunction *MF = MI.getParent()->getParent(); | |||
| 862 | Def->getOperand(1).ChangeToImmediate(Imm); | |||
| 863 | Def->addImplicitDefUseOperands(*MF); | |||
| 864 | Def->setDesc(TII->get(SMovOp)); | |||
| 865 | } | |||
| 866 | } | |||
| 867 | } | |||
| 868 | ||||
| 869 | if ((!TRI->isVectorRegister(*MRI, PHIRes) && | |||
| 870 | RC0 != &AMDGPU::VReg_1RegClass) && | |||
| 871 | (hasVGPRInput || numVGPRUses > 1)) { | |||
| 872 | LLVM_DEBUG(dbgs() << "Fixing PHI: " << MI)do { } while (false); | |||
| 873 | CreatedBB = TII->moveToVALU(MI); | |||
| 874 | } | |||
| 875 | else { | |||
| 876 | LLVM_DEBUG(dbgs() << "Legalizing PHI: " << MI)do { } while (false); | |||
| 877 | TII->legalizeOperands(MI, MDT); | |||
| 878 | } | |||
| 879 | ||||
| 880 | // Propagate register class back to PHI operands which are PHI themselves. | |||
| 881 | while (!PHIOperands.empty()) { | |||
| 882 | processPHINode(*PHIOperands.pop_back_val()); | |||
| 883 | } | |||
| 884 | return CreatedBB; | |||
| 885 | } |
| 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 |