| 1 | //===- RegAllocFast.cpp - A fast register allocator for debug code --------===// |
| 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 A block-local register allocator. No virtual register stays in a |
| 10 | /// register across a block boundary. A value live across one gets a stack slot: |
| 11 | /// spilled after its def and reloaded above its uses in each block, at the top |
| 12 | /// of the block or just after an intervening instruction that evicts it. |
| 13 | /// There is no dataflow liveness analysis, only a bounded scan of def and use |
| 14 | /// lists, and no live range splitting, interference graph or coalescer, only a |
| 15 | /// copy hint plus removal of COPYs that end up identity or dead. |
| 16 | /// |
| 17 | /// Each block is walked backwards: a use is the first reference reached and |
| 18 | /// acquires a register, a def is the last and releases one. |
| 19 | /// |
| 20 | /// Where the target enables it, TwoAddressInstructionPass is left out of the |
| 21 | /// pipeline: this pass lowers tied operands and expands REG_SEQUENCE and |
| 22 | /// INSERT_SUBREG itself. |
| 23 | // |
| 24 | //===----------------------------------------------------------------------===// |
| 25 | |
| 26 | #include "llvm/CodeGen/RegAllocFast.h" |
| 27 | #include "llvm/ADT/ArrayRef.h" |
| 28 | #include "llvm/ADT/DenseMap.h" |
| 29 | #include "llvm/ADT/IndexedMap.h" |
| 30 | #include "llvm/ADT/MapVector.h" |
| 31 | #include "llvm/ADT/SmallSet.h" |
| 32 | #include "llvm/ADT/SmallVector.h" |
| 33 | #include "llvm/ADT/SparseSet.h" |
| 34 | #include "llvm/ADT/Statistic.h" |
| 35 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 36 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 37 | #include "llvm/CodeGen/MachineFunction.h" |
| 38 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 39 | #include "llvm/CodeGen/MachineInstr.h" |
| 40 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 41 | #include "llvm/CodeGen/MachineOperand.h" |
| 42 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 43 | #include "llvm/CodeGen/RegAllocCommon.h" |
| 44 | #include "llvm/CodeGen/RegAllocRegistry.h" |
| 45 | #include "llvm/CodeGen/RegisterClassInfo.h" |
| 46 | #include "llvm/CodeGen/TargetInstrInfo.h" |
| 47 | #include "llvm/CodeGen/TargetOpcodes.h" |
| 48 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 49 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
| 50 | #include "llvm/InitializePasses.h" |
| 51 | #include "llvm/MC/MCRegisterInfo.h" |
| 52 | #include "llvm/Pass.h" |
| 53 | #include "llvm/Support/Debug.h" |
| 54 | #include "llvm/Support/ErrorHandling.h" |
| 55 | #include "llvm/Support/raw_ostream.h" |
| 56 | #include <cassert> |
| 57 | #include <tuple> |
| 58 | #include <vector> |
| 59 | |
| 60 | using namespace llvm; |
| 61 | |
| 62 | #define DEBUG_TYPE "regalloc" |
| 63 | |
| 64 | STATISTIC(NumStores, "Number of stores added" ); |
| 65 | STATISTIC(NumLoads, "Number of loads added" ); |
| 66 | STATISTIC(NumCoalesced, "Number of copies coalesced" ); |
| 67 | |
| 68 | static RegisterRegAlloc fastRegAlloc("fast" , "fast register allocator" , |
| 69 | createFastRegisterAllocator); |
| 70 | |
| 71 | namespace { |
| 72 | |
| 73 | /// Assign ascending index for instructions in machine basic block. The index |
| 74 | /// can be used to determine dominance between instructions in same MBB. |
| 75 | class InstrPosIndexes { |
| 76 | public: |
| 77 | void unsetInitialized() { IsInitialized = false; } |
| 78 | |
| 79 | void init(const MachineBasicBlock &MBB) { |
| 80 | CurMBB = &MBB; |
| 81 | Instr2PosIndex.clear(); |
| 82 | uint64_t LastIndex = 0; |
| 83 | for (const MachineInstr &MI : MBB) { |
| 84 | LastIndex += InstrDist; |
| 85 | Instr2PosIndex[&MI] = LastIndex; |
| 86 | } |
| 87 | } |
| 88 | |
| 89 | /// Set \p Index to index of \p MI. If \p MI is new inserted, it try to assign |
| 90 | /// index without affecting existing instruction's index. Return true if all |
| 91 | /// instructions index has been reassigned. |
| 92 | bool getIndex(const MachineInstr &MI, uint64_t &Index) { |
| 93 | if (!IsInitialized) { |
| 94 | init(MBB: *MI.getParent()); |
| 95 | IsInitialized = true; |
| 96 | Index = Instr2PosIndex.at(Val: &MI); |
| 97 | return true; |
| 98 | } |
| 99 | |
| 100 | assert(MI.getParent() == CurMBB && "MI is not in CurMBB" ); |
| 101 | auto It = Instr2PosIndex.find(Val: &MI); |
| 102 | if (It != Instr2PosIndex.end()) { |
| 103 | Index = It->second; |
| 104 | return false; |
| 105 | } |
| 106 | |
| 107 | // Distance is the number of consecutive unassigned instructions including |
| 108 | // MI. Start is the first instruction of them. End is the next of last |
| 109 | // instruction of them. |
| 110 | // e.g. |
| 111 | // |Instruction| A | B | C | MI | D | E | |
| 112 | // | Index | 1024 | | | | | 2048 | |
| 113 | // |
| 114 | // In this case, B, C, MI, D are unassigned. Distance is 4, Start is B, End |
| 115 | // is E. |
| 116 | unsigned Distance = 1; |
| 117 | MachineBasicBlock::const_iterator Start = MI.getIterator(), |
| 118 | End = std::next(x: Start); |
| 119 | while (Start != CurMBB->begin() && |
| 120 | !Instr2PosIndex.count(Val: &*std::prev(x: Start))) { |
| 121 | --Start; |
| 122 | ++Distance; |
| 123 | } |
| 124 | while (End != CurMBB->end() && !Instr2PosIndex.count(Val: &*(End))) { |
| 125 | ++End; |
| 126 | ++Distance; |
| 127 | } |
| 128 | |
| 129 | // LastIndex is initialized to last used index prior to MI or zero. |
| 130 | // In previous example, LastIndex is 1024, EndIndex is 2048; |
| 131 | uint64_t LastIndex = |
| 132 | Start == CurMBB->begin() ? 0 : Instr2PosIndex.at(Val: &*std::prev(x: Start)); |
| 133 | uint64_t Step; |
| 134 | if (End == CurMBB->end()) |
| 135 | Step = static_cast<uint64_t>(InstrDist); |
| 136 | else { |
| 137 | // No instruction uses index zero. |
| 138 | uint64_t EndIndex = Instr2PosIndex.at(Val: &*End); |
| 139 | assert(EndIndex > LastIndex && "Index must be ascending order" ); |
| 140 | unsigned NumAvailableIndexes = EndIndex - LastIndex - 1; |
| 141 | // We want index gap between two adjacent MI is as same as possible. Given |
| 142 | // total A available indexes, D is number of consecutive unassigned |
| 143 | // instructions, S is the step. |
| 144 | // |<- S-1 -> MI <- S-1 -> MI <- A-S*D ->| |
| 145 | // There're S-1 available indexes between unassigned instruction and its |
| 146 | // predecessor. There're A-S*D available indexes between the last |
| 147 | // unassigned instruction and its successor. |
| 148 | // Ideally, we want |
| 149 | // S-1 = A-S*D |
| 150 | // then |
| 151 | // S = (A+1)/(D+1) |
| 152 | // An valid S must be integer greater than zero, so |
| 153 | // S <= (A+1)/(D+1) |
| 154 | // => |
| 155 | // A-S*D >= 0 |
| 156 | // That means we can safely use (A+1)/(D+1) as step. |
| 157 | // In previous example, Step is 204, Index of B, C, MI, D is 1228, 1432, |
| 158 | // 1636, 1840. |
| 159 | Step = (NumAvailableIndexes + 1) / (Distance + 1); |
| 160 | } |
| 161 | |
| 162 | // Reassign index for all instructions if number of new inserted |
| 163 | // instructions exceed slot or all instructions are new. |
| 164 | if (LLVM_UNLIKELY(!Step || (!LastIndex && Step == InstrDist))) { |
| 165 | init(MBB: *CurMBB); |
| 166 | Index = Instr2PosIndex.at(Val: &MI); |
| 167 | return true; |
| 168 | } |
| 169 | |
| 170 | for (auto I = Start; I != End; ++I) { |
| 171 | LastIndex += Step; |
| 172 | Instr2PosIndex[&*I] = LastIndex; |
| 173 | } |
| 174 | Index = Instr2PosIndex.at(Val: &MI); |
| 175 | return false; |
| 176 | } |
| 177 | |
| 178 | private: |
| 179 | bool IsInitialized = false; |
| 180 | enum { InstrDist = 1024 }; |
| 181 | const MachineBasicBlock *CurMBB = nullptr; |
| 182 | DenseMap<const MachineInstr *, uint64_t> Instr2PosIndex; |
| 183 | }; |
| 184 | |
| 185 | class RegAllocFastImpl { |
| 186 | public: |
| 187 | RegAllocFastImpl(const RegAllocFilterFunc F = nullptr, |
| 188 | bool ClearVirtRegs_ = true) |
| 189 | : ShouldAllocateRegisterImpl(F), StackSlotForVirtReg(-1), |
| 190 | ClearVirtRegs(ClearVirtRegs_) {} |
| 191 | |
| 192 | private: |
| 193 | MachineFrameInfo *MFI = nullptr; |
| 194 | MachineRegisterInfo *MRI = nullptr; |
| 195 | const TargetRegisterInfo *TRI = nullptr; |
| 196 | const TargetInstrInfo *TII = nullptr; |
| 197 | RegisterClassInfo RegClassInfo; |
| 198 | const RegAllocFilterFunc ShouldAllocateRegisterImpl; |
| 199 | |
| 200 | /// Tied operands reach this pass unrewritten (TwoAddressInstructionPass was |
| 201 | /// left out of the pipeline): lower them here. |
| 202 | bool LowerTiedOps = false; |
| 203 | |
| 204 | /// Basic block currently being allocated. |
| 205 | MachineBasicBlock *MBB = nullptr; |
| 206 | |
| 207 | /// Maps virtual regs to the frame index where these values are spilled. |
| 208 | IndexedMap<int, VirtReg2IndexFunctor> StackSlotForVirtReg; |
| 209 | |
| 210 | /// A virtual register live at the current point of the backward walk. |
| 211 | /// Created at its last reference, cleared only when the block is done. |
| 212 | struct LiveReg { |
| 213 | MachineInstr *LastUse = nullptr; ///< Last instr to use reg. |
| 214 | Register VirtReg; ///< Virtual register number. |
| 215 | MCRegister PhysReg; ///< Currently held here, 0 if none. |
| 216 | bool LiveOut = false; ///< May be live out; the def spills. |
| 217 | bool Reloaded = false; ///< Reloaded below; the def spills. |
| 218 | bool Error = false; ///< Could not allocate. |
| 219 | |
| 220 | explicit LiveReg(Register VirtReg) : VirtReg(VirtReg) {} |
| 221 | explicit LiveReg() = default; |
| 222 | |
| 223 | unsigned getSparseSetIndex() const { return VirtReg.virtRegIndex(); } |
| 224 | }; |
| 225 | |
| 226 | using LiveRegMap = SparseSet<LiveReg, unsigned, identity, uint16_t>; |
| 227 | /// This map contains entries for each virtual register that is currently |
| 228 | /// available in a physical register. |
| 229 | LiveRegMap LiveVirtRegs; |
| 230 | |
| 231 | /// Stores assigned virtual registers present in the bundle MI. |
| 232 | DenseMap<Register, LiveReg> BundleVirtRegsMap; |
| 233 | |
| 234 | DenseMap<Register, SmallVector<MachineOperand *, 2>> LiveDbgValueMap; |
| 235 | /// List of DBG_VALUE that we encountered without the vreg being assigned |
| 236 | /// because they were placed after the last use of the vreg. |
| 237 | DenseMap<Register, SmallVector<MachineInstr *, 1>> DanglingDbgValues; |
| 238 | |
| 239 | /// Has a bit set for every virtual register for which it was determined |
| 240 | /// that it is alive across blocks. |
| 241 | BitVector MayLiveAcrossBlocks; |
| 242 | |
| 243 | /// What occupies a register unit. Registers interfere exactly when their |
| 244 | /// unit sets intersect, so overlap needs no alias walk. |
| 245 | enum RegUnitState { |
| 246 | /// Not in use; a register is allocatable iff all of its units are free. |
| 247 | regFree, |
| 248 | |
| 249 | /// Not available to the allocator and not a virtual register: a physreg |
| 250 | /// operand or a block live-out. Cannot be spilled. |
| 251 | regPreAssigned, |
| 252 | |
| 253 | /// Scratch marker: reloadAtBegin() stamps MBB.liveins() over the finished |
| 254 | /// map, and a virtual register left in a live-in register is not reloaded. |
| 255 | regLiveIn, |
| 256 | |
| 257 | /// Any other value is a virtual register number (>= VirtualRegFlag); |
| 258 | /// LiveVirtRegs holds the inverse mapping. |
| 259 | }; |
| 260 | |
| 261 | /// State of each register unit, indexed by MCRegUnit. |
| 262 | std::vector<unsigned> RegUnitStates; |
| 263 | |
| 264 | SmallVector<MachineInstr *, 32> Coalesced; |
| 265 | |
| 266 | /// Track register units that are used in the current instruction, and so |
| 267 | /// cannot be allocated. |
| 268 | /// |
| 269 | /// In the first phase (tied defs/early clobber), we consider also physical |
| 270 | /// uses, afterwards, we don't. If the lowest bit isn't set, it's a solely |
| 271 | /// physical use (markPhysRegUsedInInstr), otherwise, it's a normal use. To |
| 272 | /// avoid resetting the entire vector after every instruction, we track the |
| 273 | /// instruction "generation" in the remaining 31 bits -- this means, that if |
| 274 | /// UsedInInstr[Idx] < InstrGen, the register unit is unused. InstrGen is |
| 275 | /// never zero and always incremented by two. |
| 276 | /// |
| 277 | /// Don't allocate inline storage: the number of register units is typically |
| 278 | /// quite large (e.g., AArch64 > 100, X86 > 200, AMDGPU > 1000). |
| 279 | uint32_t InstrGen; |
| 280 | SmallVector<unsigned, 0> UsedInInstr; |
| 281 | |
| 282 | /// Register units defined by a non-dead physreg def of the current |
| 283 | /// instruction, indexed by MCRegUnit. Stamped with InstrGen like |
| 284 | /// UsedInInstr, so a unit is set if LiveDefUnits[Unit] == InstrGen. |
| 285 | SmallVector<uint32_t, 0> LiveDefUnits; |
| 286 | |
| 287 | SmallVector<unsigned, 8> DefOperandIndexes; |
| 288 | // Register masks attached to the current instruction. |
| 289 | SmallVector<const uint32_t *> RegMasks; |
| 290 | |
| 291 | // Assign index for each instruction to quickly determine dominance. |
| 292 | InstrPosIndexes PosIndexes; |
| 293 | |
| 294 | void setRegUnitState(MCRegUnit Unit, unsigned NewState); |
| 295 | unsigned getRegUnitState(MCRegUnit Unit) const; |
| 296 | |
| 297 | void setPhysRegState(MCRegister PhysReg, unsigned NewState); |
| 298 | bool isPhysRegFree(MCRegister PhysReg) const; |
| 299 | |
| 300 | /// Mark a physreg as used in this instruction. |
| 301 | void markRegUsedInInstr(MCRegister PhysReg) { |
| 302 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) |
| 303 | UsedInInstr[static_cast<unsigned>(Unit)] = InstrGen | 1; |
| 304 | } |
| 305 | |
| 306 | // Check if physreg is clobbered by instruction's regmask(s). |
| 307 | bool isClobberedByRegMasks(MCRegister PhysReg) const { |
| 308 | return llvm::any_of(Range: RegMasks, P: [PhysReg](const uint32_t *Mask) { |
| 309 | return MachineOperand::clobbersPhysReg(RegMask: Mask, PhysReg); |
| 310 | }); |
| 311 | } |
| 312 | |
| 313 | /// Check if a physreg or any of its aliases are used in this instruction. |
| 314 | bool isRegUsedInInstr(MCRegister PhysReg, bool LookAtPhysRegUses) const { |
| 315 | if (LookAtPhysRegUses && isClobberedByRegMasks(PhysReg)) |
| 316 | return true; |
| 317 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) |
| 318 | if (UsedInInstr[static_cast<unsigned>(Unit)] >= |
| 319 | (InstrGen | !LookAtPhysRegUses)) |
| 320 | return true; |
| 321 | return false; |
| 322 | } |
| 323 | |
| 324 | /// Mark physical register as being used in a register use operand. |
| 325 | /// This is only used by the special livethrough handling code. |
| 326 | void markPhysRegUsedInInstr(MCRegister PhysReg) { |
| 327 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
| 328 | assert(UsedInInstr[static_cast<unsigned>(Unit)] <= InstrGen && |
| 329 | "non-phys use before phys use?" ); |
| 330 | UsedInInstr[static_cast<unsigned>(Unit)] = InstrGen; |
| 331 | } |
| 332 | } |
| 333 | |
| 334 | /// Remove mark of physical register being used in the instruction. |
| 335 | void unmarkRegUsedInInstr(MCRegister PhysReg) { |
| 336 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) |
| 337 | UsedInInstr[static_cast<unsigned>(Unit)] = 0; |
| 338 | } |
| 339 | |
| 340 | /// Record that a non-dead def of the current instruction keeps every register |
| 341 | /// unit of \p PhysReg live. |
| 342 | void markLiveDefUnits(MCRegister PhysReg) { |
| 343 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) |
| 344 | LiveDefUnits[static_cast<unsigned>(Unit)] = InstrGen; |
| 345 | } |
| 346 | |
| 347 | /// Check if every register unit of \p PhysReg is defined by a non-dead def of |
| 348 | /// the current instruction. |
| 349 | bool hasLiveDefUnits(MCRegister PhysReg) const { |
| 350 | return all_of(Range: TRI->regunits(Reg: PhysReg), P: [this](MCRegUnit Unit) { |
| 351 | return LiveDefUnits[static_cast<unsigned>(Unit)] == InstrGen; |
| 352 | }); |
| 353 | } |
| 354 | |
| 355 | enum : unsigned { |
| 356 | spillClean = 50, |
| 357 | spillDirty = 100, |
| 358 | spillPrefBonus = 20, |
| 359 | spillImpossible = ~0u |
| 360 | }; |
| 361 | |
| 362 | public: |
| 363 | bool ClearVirtRegs; |
| 364 | |
| 365 | bool runOnMachineFunction(MachineFunction &MF); |
| 366 | |
| 367 | private: |
| 368 | void allocateBasicBlock(MachineBasicBlock &MBB); |
| 369 | void expandSubregPseudo(MachineInstr &MI); |
| 370 | |
| 371 | void addRegClassDefCounts(MutableArrayRef<unsigned> RegClassDefCounts, |
| 372 | Register Reg) const; |
| 373 | |
| 374 | void findAndSortDefOperandIndexes(const MachineInstr &MI); |
| 375 | |
| 376 | void allocateInstruction(MachineInstr &MI); |
| 377 | void handleDebugValue(MachineInstr &MI); |
| 378 | void handleBundle(MachineInstr &MI); |
| 379 | |
| 380 | bool usePhysReg(MachineInstr &MI, MCRegister PhysReg); |
| 381 | bool definePhysReg(MachineInstr &MI, MCRegister PhysReg); |
| 382 | bool displacePhysReg(MachineInstr &MI, MCRegister PhysReg); |
| 383 | void freePhysReg(MCRegister PhysReg); |
| 384 | |
| 385 | unsigned calcSpillCost(MCPhysReg PhysReg) const; |
| 386 | |
| 387 | LiveRegMap::iterator findLiveVirtReg(Register VirtReg) { |
| 388 | return LiveVirtRegs.find(Key: VirtReg.virtRegIndex()); |
| 389 | } |
| 390 | |
| 391 | LiveRegMap::const_iterator findLiveVirtReg(Register VirtReg) const { |
| 392 | return LiveVirtRegs.find(Key: VirtReg.virtRegIndex()); |
| 393 | } |
| 394 | |
| 395 | void assignVirtToPhysReg(MachineInstr &MI, LiveReg &, MCRegister PhysReg); |
| 396 | void allocVirtReg(MachineInstr &MI, LiveReg &LR, Register Hint, |
| 397 | bool LookAtPhysRegUses = false); |
| 398 | void allocVirtRegUndef(MachineOperand &MO); |
| 399 | void assignDanglingDebugValues(MachineInstr &Def, Register VirtReg, |
| 400 | MCRegister Reg); |
| 401 | bool defineLiveThroughVirtReg(MachineInstr &MI, unsigned OpNum, |
| 402 | Register VirtReg); |
| 403 | bool defineVirtReg(MachineInstr &MI, unsigned OpNum, Register VirtReg, |
| 404 | bool LookAtPhysRegUses = false); |
| 405 | bool useVirtReg(MachineInstr &MI, MachineOperand &MO, Register VirtReg); |
| 406 | bool lowerTiedUse(MachineInstr &MI, MachineOperand &MO, LiveReg &LR); |
| 407 | |
| 408 | MCPhysReg getErrorAssignment(const LiveReg &LR, MachineInstr &MI, |
| 409 | const TargetRegisterClass &RC); |
| 410 | |
| 411 | MachineBasicBlock::iterator |
| 412 | getMBBBeginInsertionPoint(MachineBasicBlock &MBB, |
| 413 | SmallSet<Register, 2> &PrologLiveIns) const; |
| 414 | |
| 415 | void reloadAtBegin(MachineBasicBlock &MBB); |
| 416 | bool setPhysReg(MachineInstr &MI, MachineOperand &MO, |
| 417 | const LiveReg &Assignment); |
| 418 | |
| 419 | Register traceCopies(Register VirtReg) const; |
| 420 | Register traceCopyChain(Register Reg) const; |
| 421 | |
| 422 | bool shouldAllocateRegister(const Register Reg) const; |
| 423 | int getStackSpaceFor(Register VirtReg); |
| 424 | void spill(MachineBasicBlock::iterator Before, Register VirtReg, |
| 425 | MCRegister AssignedReg, bool Kill, bool LiveOut); |
| 426 | void reload(MachineBasicBlock::iterator Before, Register VirtReg, |
| 427 | MCRegister PhysReg); |
| 428 | |
| 429 | bool mayLiveOut(Register VirtReg); |
| 430 | bool mayLiveIn(Register VirtReg); |
| 431 | |
| 432 | bool mayBeSpillFromInlineAsmBr(const MachineInstr &MI) const; |
| 433 | |
| 434 | void dumpState() const; |
| 435 | }; |
| 436 | |
| 437 | class RegAllocFast : public MachineFunctionPass { |
| 438 | RegAllocFastImpl Impl; |
| 439 | |
| 440 | public: |
| 441 | static char ID; |
| 442 | |
| 443 | RegAllocFast(const RegAllocFilterFunc F = nullptr, bool ClearVirtRegs_ = true) |
| 444 | : MachineFunctionPass(ID), Impl(F, ClearVirtRegs_) {} |
| 445 | |
| 446 | bool runOnMachineFunction(MachineFunction &MF) override { |
| 447 | return Impl.runOnMachineFunction(MF); |
| 448 | } |
| 449 | |
| 450 | StringRef getPassName() const override { return "Fast Register Allocator" ; } |
| 451 | |
| 452 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 453 | AU.setPreservesCFG(); |
| 454 | MachineFunctionPass::getAnalysisUsage(AU); |
| 455 | } |
| 456 | |
| 457 | MachineFunctionProperties getRequiredProperties() const override { |
| 458 | return MachineFunctionProperties().setNoPHIs(); |
| 459 | } |
| 460 | |
| 461 | MachineFunctionProperties getSetProperties() const override { |
| 462 | MachineFunctionProperties P; |
| 463 | if (Impl.ClearVirtRegs) |
| 464 | P.setNoVRegs().setTiedOpsRewritten(); |
| 465 | return P; |
| 466 | } |
| 467 | |
| 468 | MachineFunctionProperties getClearedProperties() const override { |
| 469 | return MachineFunctionProperties().setIsSSA(); |
| 470 | } |
| 471 | }; |
| 472 | |
| 473 | } // end anonymous namespace |
| 474 | |
| 475 | char RegAllocFast::ID = 0; |
| 476 | |
| 477 | INITIALIZE_PASS(RegAllocFast, "regallocfast" , "Fast Register Allocator" , false, |
| 478 | false) |
| 479 | |
| 480 | bool RegAllocFastImpl::shouldAllocateRegister(const Register Reg) const { |
| 481 | assert(Reg.isVirtual()); |
| 482 | if (!ShouldAllocateRegisterImpl) |
| 483 | return true; |
| 484 | |
| 485 | return ShouldAllocateRegisterImpl(*TRI, *MRI, Reg); |
| 486 | } |
| 487 | |
| 488 | void RegAllocFastImpl::setRegUnitState(MCRegUnit Unit, unsigned NewState) { |
| 489 | RegUnitStates[static_cast<unsigned>(Unit)] = NewState; |
| 490 | } |
| 491 | |
| 492 | unsigned RegAllocFastImpl::getRegUnitState(MCRegUnit Unit) const { |
| 493 | return RegUnitStates[static_cast<unsigned>(Unit)]; |
| 494 | } |
| 495 | |
| 496 | void RegAllocFastImpl::setPhysRegState(MCRegister PhysReg, unsigned NewState) { |
| 497 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) |
| 498 | setRegUnitState(Unit, NewState); |
| 499 | } |
| 500 | |
| 501 | bool RegAllocFastImpl::isPhysRegFree(MCRegister PhysReg) const { |
| 502 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
| 503 | if (getRegUnitState(Unit) != regFree) |
| 504 | return false; |
| 505 | } |
| 506 | return true; |
| 507 | } |
| 508 | |
| 509 | /// This allocates space for the specified virtual register to be held on the |
| 510 | /// stack. |
| 511 | int RegAllocFastImpl::getStackSpaceFor(Register VirtReg) { |
| 512 | // Find the location Reg would belong... |
| 513 | int SS = StackSlotForVirtReg[VirtReg]; |
| 514 | // Already has space allocated? |
| 515 | if (SS != -1) |
| 516 | return SS; |
| 517 | |
| 518 | // Allocate a new stack object for this spill location... |
| 519 | const TargetRegisterClass &RC = *MRI->getRegClass(Reg: VirtReg); |
| 520 | unsigned Size = TRI->getSpillSize(RC); |
| 521 | Align Alignment = TRI->getSpillAlign(RC); |
| 522 | |
| 523 | const MachineFunction &MF = MRI->getMF(); |
| 524 | auto &ST = MF.getSubtarget(); |
| 525 | Align CurrentAlign = ST.getFrameLowering()->getStackAlign(); |
| 526 | if (Alignment > CurrentAlign && !TRI->canRealignStack(MF)) |
| 527 | Alignment = CurrentAlign; |
| 528 | |
| 529 | int FrameIdx = |
| 530 | MFI->CreateSpillStackObject(Size, Alignment, StackID: TRI->getSpillStackID(RC)); |
| 531 | |
| 532 | // Assign the slot. |
| 533 | StackSlotForVirtReg[VirtReg] = FrameIdx; |
| 534 | return FrameIdx; |
| 535 | } |
| 536 | |
| 537 | static bool dominates(InstrPosIndexes &PosIndexes, const MachineInstr &A, |
| 538 | const MachineInstr &B) { |
| 539 | uint64_t IndexA, IndexB; |
| 540 | PosIndexes.getIndex(MI: A, Index&: IndexA); |
| 541 | // getIndex() returns true when it renumbered the block, invalidating IndexA. |
| 542 | if (LLVM_UNLIKELY(PosIndexes.getIndex(B, IndexB))) |
| 543 | PosIndexes.getIndex(MI: A, Index&: IndexA); |
| 544 | return IndexA < IndexB; |
| 545 | } |
| 546 | |
| 547 | /// Returns true if \p MI is a spill of a live-in physical register in a block |
| 548 | /// targeted by an INLINEASM_BR. Such spills must precede reloads of live-in |
| 549 | /// virtual registers, so that we do not reload from an uninitialized stack |
| 550 | /// slot. |
| 551 | bool RegAllocFastImpl::mayBeSpillFromInlineAsmBr(const MachineInstr &MI) const { |
| 552 | int FI; |
| 553 | auto *MBB = MI.getParent(); |
| 554 | if (MBB->isInlineAsmBrIndirectTarget() && TII->isStoreToStackSlot(MI, FrameIndex&: FI) && |
| 555 | MFI->isSpillSlotObjectIndex(ObjectIdx: FI)) |
| 556 | for (const auto &Op : MI.operands()) |
| 557 | if (Op.isReg() && Op.getReg().isValid() && MBB->isLiveIn(Reg: Op.getReg())) |
| 558 | return true; |
| 559 | return false; |
| 560 | } |
| 561 | |
| 562 | /// Returns false if \p VirtReg is known to not live out of the current block. |
| 563 | bool RegAllocFastImpl::mayLiveOut(Register VirtReg) { |
| 564 | if (MayLiveAcrossBlocks.test(Idx: VirtReg.virtRegIndex())) { |
| 565 | // Cannot be live-out if there are no successors. |
| 566 | return !MBB->succ_empty(); |
| 567 | } |
| 568 | |
| 569 | const MachineInstr *SelfLoopDef = nullptr; |
| 570 | |
| 571 | // If this block loops back to itself, it is necessary to check whether the |
| 572 | // use comes after the def. |
| 573 | if (MBB->isSuccessor(MBB)) { |
| 574 | // Find the first def in the self loop MBB. |
| 575 | for (const MachineInstr &DefInst : MRI->def_instructions(Reg: VirtReg)) { |
| 576 | if (DefInst.getParent() != MBB) { |
| 577 | MayLiveAcrossBlocks.set(VirtReg.virtRegIndex()); |
| 578 | return true; |
| 579 | } else { |
| 580 | if (!SelfLoopDef || dominates(PosIndexes, A: DefInst, B: *SelfLoopDef)) |
| 581 | SelfLoopDef = &DefInst; |
| 582 | } |
| 583 | } |
| 584 | if (!SelfLoopDef) { |
| 585 | MayLiveAcrossBlocks.set(VirtReg.virtRegIndex()); |
| 586 | return true; |
| 587 | } |
| 588 | } |
| 589 | |
| 590 | // See if the first \p Limit uses of the register are all in the current |
| 591 | // block. |
| 592 | static const unsigned Limit = 8; |
| 593 | unsigned C = 0; |
| 594 | for (const MachineInstr &UseInst : MRI->use_nodbg_instructions(Reg: VirtReg)) { |
| 595 | if (UseInst.getParent() != MBB || ++C >= Limit) { |
| 596 | MayLiveAcrossBlocks.set(VirtReg.virtRegIndex()); |
| 597 | // Cannot be live-out if there are no successors. |
| 598 | return !MBB->succ_empty(); |
| 599 | } |
| 600 | |
| 601 | if (SelfLoopDef) { |
| 602 | // Try to handle some simple cases to avoid spilling and reloading every |
| 603 | // value inside a self looping block. |
| 604 | if (SelfLoopDef == &UseInst || |
| 605 | !dominates(PosIndexes, A: *SelfLoopDef, B: UseInst)) { |
| 606 | MayLiveAcrossBlocks.set(VirtReg.virtRegIndex()); |
| 607 | return true; |
| 608 | } |
| 609 | } |
| 610 | } |
| 611 | |
| 612 | return false; |
| 613 | } |
| 614 | |
| 615 | /// Returns false if \p VirtReg is known to not be live into the current block. |
| 616 | bool RegAllocFastImpl::mayLiveIn(Register VirtReg) { |
| 617 | if (MayLiveAcrossBlocks.test(Idx: VirtReg.virtRegIndex())) |
| 618 | return !MBB->pred_empty(); |
| 619 | |
| 620 | // See if the first \p Limit def of the register are all in the current block. |
| 621 | static const unsigned Limit = 8; |
| 622 | unsigned C = 0; |
| 623 | for (const MachineInstr &DefInst : MRI->def_instructions(Reg: VirtReg)) { |
| 624 | if (DefInst.getParent() != MBB || ++C >= Limit) { |
| 625 | MayLiveAcrossBlocks.set(VirtReg.virtRegIndex()); |
| 626 | return !MBB->pred_empty(); |
| 627 | } |
| 628 | } |
| 629 | |
| 630 | return false; |
| 631 | } |
| 632 | |
| 633 | /// Insert spill instruction for \p AssignedReg before \p Before. Update |
| 634 | /// DBG_VALUEs with \p VirtReg operands with the stack slot. |
| 635 | void RegAllocFastImpl::spill(MachineBasicBlock::iterator Before, |
| 636 | Register VirtReg, MCRegister AssignedReg, |
| 637 | bool Kill, bool LiveOut) { |
| 638 | LLVM_DEBUG(dbgs() << "Spilling " << printReg(VirtReg, TRI) << " in " |
| 639 | << printReg(AssignedReg, TRI)); |
| 640 | int FI = getStackSpaceFor(VirtReg); |
| 641 | LLVM_DEBUG(dbgs() << " to stack slot #" << FI << '\n'); |
| 642 | |
| 643 | const TargetRegisterClass &RC = *MRI->getRegClass(Reg: VirtReg); |
| 644 | TII->storeRegToStackSlot(MBB&: *MBB, MI: Before, SrcReg: AssignedReg, isKill: Kill, FrameIndex: FI, RC: &RC, VReg: VirtReg); |
| 645 | ++NumStores; |
| 646 | |
| 647 | MachineBasicBlock::iterator FirstTerm = MBB->getFirstTerminator(); |
| 648 | |
| 649 | // When we spill a virtual register, we will have spill instructions behind |
| 650 | // every definition of it, meaning we can switch all the DBG_VALUEs over |
| 651 | // to just reference the stack slot. |
| 652 | SmallVectorImpl<MachineOperand *> &LRIDbgOperands = LiveDbgValueMap[VirtReg]; |
| 653 | SmallMapVector<MachineInstr *, SmallVector<const MachineOperand *>, 2> |
| 654 | SpilledOperandsMap; |
| 655 | for (MachineOperand *MO : LRIDbgOperands) |
| 656 | SpilledOperandsMap[MO->getParent()].push_back(Elt: MO); |
| 657 | for (const auto &MISpilledOperands : SpilledOperandsMap) { |
| 658 | MachineInstr &DBG = *MISpilledOperands.first; |
| 659 | // We don't have enough support for tracking operands of DBG_VALUE_LISTs. |
| 660 | if (DBG.isDebugValueList()) |
| 661 | continue; |
| 662 | MachineInstr *NewDV = buildDbgValueForSpill( |
| 663 | BB&: *MBB, I: Before, Orig: *MISpilledOperands.first, FrameIndex: FI, SpilledOperands: MISpilledOperands.second); |
| 664 | assert(NewDV->getParent() == MBB && "dangling parent pointer" ); |
| 665 | (void)NewDV; |
| 666 | LLVM_DEBUG(dbgs() << "Inserting debug info due to spill:\n" << *NewDV); |
| 667 | |
| 668 | if (LiveOut) { |
| 669 | // We need to insert a DBG_VALUE at the end of the block if the spill slot |
| 670 | // is live out, but there is another use of the value after the |
| 671 | // spill. This will allow LiveDebugValues to see the correct live out |
| 672 | // value to propagate to the successors. |
| 673 | MachineInstr *ClonedDV = MBB->getParent()->CloneMachineInstr(Orig: NewDV); |
| 674 | MBB->insert(I: FirstTerm, MI: ClonedDV); |
| 675 | LLVM_DEBUG(dbgs() << "Cloning debug info due to live out spill\n" ); |
| 676 | } |
| 677 | |
| 678 | // Rewrite unassigned dbg_values to use the stack slot. |
| 679 | // TODO We can potentially do this for list debug values as well if we know |
| 680 | // how the dbg_values are getting unassigned. |
| 681 | if (DBG.isNonListDebugValue()) { |
| 682 | MachineOperand &MO = DBG.getDebugOperand(Index: 0); |
| 683 | if (MO.isReg() && !MO.getReg()) { |
| 684 | updateDbgValueForSpill(Orig&: DBG, FrameIndex: FI, Reg: Register()); |
| 685 | } |
| 686 | } |
| 687 | } |
| 688 | // Now this register is spilled there is should not be any DBG_VALUE |
| 689 | // pointing to this register because they are all pointing to spilled value |
| 690 | // now. |
| 691 | LRIDbgOperands.clear(); |
| 692 | } |
| 693 | |
| 694 | /// Insert reload instruction for \p PhysReg before \p Before. |
| 695 | void RegAllocFastImpl::reload(MachineBasicBlock::iterator Before, |
| 696 | Register VirtReg, MCRegister PhysReg) { |
| 697 | LLVM_DEBUG(dbgs() << "Reloading " << printReg(VirtReg, TRI) << " into " |
| 698 | << printReg(PhysReg, TRI) << '\n'); |
| 699 | int FI = getStackSpaceFor(VirtReg); |
| 700 | const TargetRegisterClass &RC = *MRI->getRegClass(Reg: VirtReg); |
| 701 | TII->loadRegFromStackSlot(MBB&: *MBB, MI: Before, DestReg: PhysReg, FrameIndex: FI, RC: &RC, VReg: VirtReg); |
| 702 | ++NumLoads; |
| 703 | } |
| 704 | |
| 705 | /// Get basic block begin insertion point. |
| 706 | /// This is not just MBB.begin() because surprisingly we have EH_LABEL |
| 707 | /// instructions marking the begin of a basic block. This means we must insert |
| 708 | /// new instructions after such labels... |
| 709 | MachineBasicBlock::iterator RegAllocFastImpl::getMBBBeginInsertionPoint( |
| 710 | MachineBasicBlock &MBB, SmallSet<Register, 2> &PrologLiveIns) const { |
| 711 | MachineBasicBlock::iterator I = MBB.begin(); |
| 712 | while (I != MBB.end()) { |
| 713 | if (I->isLabel()) { |
| 714 | ++I; |
| 715 | continue; |
| 716 | } |
| 717 | |
| 718 | // Skip prologues and inlineasm_br spills to place reloads afterwards. |
| 719 | if (!TII->isBasicBlockPrologue(MI: *I) && !mayBeSpillFromInlineAsmBr(MI: *I)) |
| 720 | break; |
| 721 | |
| 722 | // However if a prolog instruction reads a register that needs to be |
| 723 | // reloaded, the reload should be inserted before the prolog. |
| 724 | for (MachineOperand &MO : I->operands()) { |
| 725 | if (MO.isReg()) |
| 726 | PrologLiveIns.insert(V: MO.getReg()); |
| 727 | } |
| 728 | |
| 729 | ++I; |
| 730 | } |
| 731 | |
| 732 | return I; |
| 733 | } |
| 734 | |
| 735 | /// Reload all currently assigned virtual registers. |
| 736 | void RegAllocFastImpl::reloadAtBegin(MachineBasicBlock &MBB) { |
| 737 | if (LiveVirtRegs.empty()) |
| 738 | return; |
| 739 | |
| 740 | // Mark live-in registers so the loop below skips reloads into them. The |
| 741 | // virtual register mappings this overwrites are not needed anymore. |
| 742 | for (MachineBasicBlock::RegisterMaskPair P : MBB.liveins()) |
| 743 | setPhysRegState(PhysReg: P.PhysReg, NewState: regLiveIn); |
| 744 | |
| 745 | SmallSet<Register, 2> PrologLiveIns; |
| 746 | |
| 747 | // The LiveRegMap is keyed by an unsigned (the virtreg number), so the order |
| 748 | // of spilling here is deterministic, if arbitrary. |
| 749 | MachineBasicBlock::iterator InsertBefore = |
| 750 | getMBBBeginInsertionPoint(MBB, PrologLiveIns); |
| 751 | for (const LiveReg &LR : LiveVirtRegs) { |
| 752 | MCRegister PhysReg = LR.PhysReg; |
| 753 | if (!PhysReg || LR.Error) |
| 754 | continue; |
| 755 | |
| 756 | MCRegUnit FirstUnit = *TRI->regunits(Reg: PhysReg).begin(); |
| 757 | if (getRegUnitState(Unit: FirstUnit) == regLiveIn) |
| 758 | continue; |
| 759 | |
| 760 | assert(&MBB != &MBB.getParent()->front() && |
| 761 | "no reload in start block. Missing vreg def?" ); |
| 762 | |
| 763 | if (PrologLiveIns.count(V: PhysReg)) { |
| 764 | // FIXME: Theoretically this should use an insert point skipping labels |
| 765 | // but I'm not sure how labels should interact with prolog instruction |
| 766 | // that need reloads. |
| 767 | reload(Before: MBB.begin(), VirtReg: LR.VirtReg, PhysReg); |
| 768 | } else |
| 769 | reload(Before: InsertBefore, VirtReg: LR.VirtReg, PhysReg); |
| 770 | } |
| 771 | LiveVirtRegs.clear(); |
| 772 | } |
| 773 | |
| 774 | /// Handle the direct use of a physical register. Displace whatever occupies it |
| 775 | /// and mark it pre-assigned: backwards, a use means live from here upward. |
| 776 | /// Returns false if nothing was displaced, so the use is a kill. This may add |
| 777 | /// implicit kills to MO->getParent() and invalidate MO. |
| 778 | bool RegAllocFastImpl::usePhysReg(MachineInstr &MI, MCRegister Reg) { |
| 779 | assert(Reg.isPhysical() && "expected physreg" ); |
| 780 | bool displacedAny = displacePhysReg(MI, PhysReg: Reg); |
| 781 | setPhysRegState(PhysReg: Reg, NewState: regPreAssigned); |
| 782 | markRegUsedInInstr(PhysReg: Reg); |
| 783 | return displacedAny; |
| 784 | } |
| 785 | |
| 786 | /// Displace whatever holds \p Reg and reserve it, so a virtual register def |
| 787 | /// cannot land on a register this instruction already writes. Released in the |
| 788 | /// free-def-operands step, after the uses for an early clobber, or by |
| 789 | /// lowerTiedUse(); if the instruction also reads \p Reg it ends up reserved |
| 790 | /// for the code above. |
| 791 | bool RegAllocFastImpl::definePhysReg(MachineInstr &MI, MCRegister Reg) { |
| 792 | bool displacedAny = displacePhysReg(MI, PhysReg: Reg); |
| 793 | setPhysRegState(PhysReg: Reg, NewState: regPreAssigned); |
| 794 | return displacedAny; |
| 795 | } |
| 796 | |
| 797 | /// Mark PhysReg as reserved or free after spilling any virtregs. This is very |
| 798 | /// similar to defineVirtReg except the physreg is reserved instead of |
| 799 | /// allocated. |
| 800 | bool RegAllocFastImpl::displacePhysReg(MachineInstr &MI, MCRegister PhysReg) { |
| 801 | bool displacedAny = false; |
| 802 | |
| 803 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
| 804 | switch (unsigned VirtReg = getRegUnitState(Unit)) { |
| 805 | default: { |
| 806 | LiveRegMap::iterator LRI = findLiveVirtReg(VirtReg); |
| 807 | assert(LRI != LiveVirtRegs.end() && "datastructures in sync" ); |
| 808 | MachineBasicBlock::iterator ReloadBefore = |
| 809 | std::next(x: (MachineBasicBlock::iterator)MI.getIterator()); |
| 810 | while (mayBeSpillFromInlineAsmBr(MI: *ReloadBefore)) |
| 811 | ++ReloadBefore; |
| 812 | reload(Before: ReloadBefore, VirtReg, PhysReg: LRI->PhysReg); |
| 813 | |
| 814 | setPhysRegState(PhysReg: LRI->PhysReg, NewState: regFree); |
| 815 | LRI->PhysReg = MCRegister(); |
| 816 | LRI->Reloaded = true; |
| 817 | displacedAny = true; |
| 818 | break; |
| 819 | } |
| 820 | case regPreAssigned: |
| 821 | setRegUnitState(Unit, NewState: regFree); |
| 822 | displacedAny = true; |
| 823 | break; |
| 824 | case regFree: |
| 825 | break; |
| 826 | } |
| 827 | } |
| 828 | return displacedAny; |
| 829 | } |
| 830 | |
| 831 | void RegAllocFastImpl::freePhysReg(MCRegister PhysReg) { |
| 832 | LLVM_DEBUG(dbgs() << "Freeing " << printReg(PhysReg, TRI) << ':'); |
| 833 | |
| 834 | MCRegUnit FirstUnit = *TRI->regunits(Reg: PhysReg).begin(); |
| 835 | switch (unsigned VirtReg = getRegUnitState(Unit: FirstUnit)) { |
| 836 | case regFree: |
| 837 | LLVM_DEBUG(dbgs() << '\n'); |
| 838 | return; |
| 839 | case regPreAssigned: |
| 840 | LLVM_DEBUG(dbgs() << '\n'); |
| 841 | setPhysRegState(PhysReg, NewState: regFree); |
| 842 | return; |
| 843 | default: { |
| 844 | LiveRegMap::iterator LRI = findLiveVirtReg(VirtReg); |
| 845 | assert(LRI != LiveVirtRegs.end()); |
| 846 | LLVM_DEBUG(dbgs() << ' ' << printReg(LRI->VirtReg, TRI) << '\n'); |
| 847 | setPhysRegState(PhysReg: LRI->PhysReg, NewState: regFree); |
| 848 | LRI->PhysReg = MCRegister(); |
| 849 | } |
| 850 | return; |
| 851 | } |
| 852 | } |
| 853 | |
| 854 | /// Return the cost of spilling clearing out PhysReg and aliases so it is free |
| 855 | /// for allocation. Returns 0 when PhysReg is free or disabled with all aliases |
| 856 | /// disabled - it can be allocated directly. |
| 857 | /// \returns spillImpossible when PhysReg or an alias can't be spilled. |
| 858 | unsigned RegAllocFastImpl::calcSpillCost(MCPhysReg PhysReg) const { |
| 859 | for (MCRegUnit Unit : TRI->regunits(Reg: PhysReg)) { |
| 860 | switch (unsigned VirtReg = getRegUnitState(Unit)) { |
| 861 | case regFree: |
| 862 | break; |
| 863 | case regPreAssigned: |
| 864 | LLVM_DEBUG(dbgs() << "Cannot spill pre-assigned " |
| 865 | << printReg(PhysReg, TRI) << '\n'); |
| 866 | return spillImpossible; |
| 867 | default: { |
| 868 | bool SureSpill = StackSlotForVirtReg[VirtReg] != -1 || |
| 869 | findLiveVirtReg(VirtReg)->LiveOut; |
| 870 | return SureSpill ? spillClean : spillDirty; |
| 871 | } |
| 872 | } |
| 873 | } |
| 874 | return 0; |
| 875 | } |
| 876 | |
| 877 | void RegAllocFastImpl::assignDanglingDebugValues(MachineInstr &Definition, |
| 878 | Register VirtReg, |
| 879 | MCRegister Reg) { |
| 880 | auto UDBGValIter = DanglingDbgValues.find(Val: VirtReg); |
| 881 | if (UDBGValIter == DanglingDbgValues.end()) |
| 882 | return; |
| 883 | |
| 884 | SmallVectorImpl<MachineInstr *> &Dangling = UDBGValIter->second; |
| 885 | for (MachineInstr *DbgValue : Dangling) { |
| 886 | assert(DbgValue->isDebugValue()); |
| 887 | if (!DbgValue->hasDebugOperandForReg(Reg: VirtReg)) |
| 888 | continue; |
| 889 | |
| 890 | // Test whether the physreg survives from the definition to the DBG_VALUE. |
| 891 | // A tied use that took over its def's register is assigned at an |
| 892 | // instruction that overwrites it, so start the scan there. |
| 893 | MCRegister SetToReg = Reg; |
| 894 | unsigned Limit = 20; |
| 895 | MachineBasicBlock::iterator I = Definition.getIterator(); |
| 896 | if (!Definition.definesRegister(Reg, TRI)) |
| 897 | ++I; |
| 898 | for (MachineBasicBlock::iterator E = DbgValue->getIterator(); I != E; ++I) { |
| 899 | if (I->modifiesRegister(Reg, TRI) || --Limit == 0) { |
| 900 | LLVM_DEBUG(dbgs() << "Register did not survive for " << *DbgValue |
| 901 | << '\n'); |
| 902 | SetToReg = MCRegister(); |
| 903 | break; |
| 904 | } |
| 905 | } |
| 906 | for (MachineOperand &MO : DbgValue->getDebugOperandsForReg(Reg: VirtReg)) { |
| 907 | MO.setReg(SetToReg); |
| 908 | if (SetToReg) |
| 909 | MO.setIsRenamable(); |
| 910 | } |
| 911 | } |
| 912 | Dangling.clear(); |
| 913 | } |
| 914 | |
| 915 | /// This method updates local state so that we know that PhysReg is the |
| 916 | /// proper container for VirtReg now. The physical register must not be used |
| 917 | /// for anything else when this is called. |
| 918 | void RegAllocFastImpl::assignVirtToPhysReg(MachineInstr &AtMI, LiveReg &LR, |
| 919 | MCRegister PhysReg) { |
| 920 | Register VirtReg = LR.VirtReg; |
| 921 | LLVM_DEBUG(dbgs() << "Assigning " << printReg(VirtReg, TRI) << " to " |
| 922 | << printReg(PhysReg, TRI) << '\n'); |
| 923 | assert(!LR.PhysReg && "Already assigned a physreg" ); |
| 924 | assert(PhysReg && "Trying to assign no register" ); |
| 925 | LR.PhysReg = PhysReg; |
| 926 | setPhysRegState(PhysReg, NewState: VirtReg.id()); |
| 927 | |
| 928 | assignDanglingDebugValues(Definition&: AtMI, VirtReg, Reg: PhysReg); |
| 929 | } |
| 930 | |
| 931 | static bool isCoalescable(const MachineInstr &MI) { return MI.isFullCopy(); } |
| 932 | |
| 933 | /// The operand \p MO is tied to. |
| 934 | static const MachineOperand &getTiedOperand(const MachineInstr &MI, |
| 935 | const MachineOperand &MO) { |
| 936 | return MI.getOperand(i: MI.findTiedOperandIdx(OpIdx: MI.getOperandNo(I: &MO))); |
| 937 | } |
| 938 | |
| 939 | /// The register \p DefMO's tied use reads, when the two end up in the same |
| 940 | /// register: a subregister index on either side makes them differ. |
| 941 | static Register getTiedUseReg(const MachineInstr &MI, |
| 942 | const MachineOperand &DefMO) { |
| 943 | if (!DefMO.isTied() || DefMO.getSubReg()) |
| 944 | return Register(); |
| 945 | const MachineOperand &UseMO = getTiedOperand(MI, MO: DefMO); |
| 946 | return UseMO.getSubReg() ? Register() : UseMO.getReg(); |
| 947 | } |
| 948 | |
| 949 | Register RegAllocFastImpl::traceCopyChain(Register Reg) const { |
| 950 | static const unsigned ChainLengthLimit = 3; |
| 951 | for (unsigned C = 0; C <= ChainLengthLimit; ++C) { |
| 952 | if (Reg.isPhysical()) |
| 953 | return Reg; |
| 954 | assert(Reg.isVirtual()); |
| 955 | |
| 956 | const MachineOperand *DefMO = MRI->getOneDef(Reg); |
| 957 | if (!DefMO) |
| 958 | return Register(); |
| 959 | const MachineInstr *Def = DefMO->getParent(); |
| 960 | if (isCoalescable(MI: *Def)) { |
| 961 | Reg = Def->getOperand(i: 1).getReg(); |
| 962 | continue; |
| 963 | } |
| 964 | // A two-address instruction's def and tied use end up in the same |
| 965 | // register, so the tie continues the chain. |
| 966 | Reg = LowerTiedOps ? getTiedUseReg(MI: *Def, DefMO: *DefMO) : Register(); |
| 967 | if (!Reg) |
| 968 | return Register(); |
| 969 | } |
| 970 | return Register(); |
| 971 | } |
| 972 | |
| 973 | /// Check if any of \p VirtReg's definitions is a copy or a tied def. If it is |
| 974 | /// follow the chain of copies to check whether we reach a physical register we |
| 975 | /// can coalesce with. |
| 976 | Register RegAllocFastImpl::traceCopies(Register VirtReg) const { |
| 977 | static const unsigned DefLimit = 3; |
| 978 | unsigned C = 0; |
| 979 | for (const MachineOperand &DefMO : MRI->def_operands(Reg: VirtReg)) { |
| 980 | const MachineInstr &MI = *DefMO.getParent(); |
| 981 | Register Reg; |
| 982 | if (isCoalescable(MI)) |
| 983 | Reg = MI.getOperand(i: 1).getReg(); |
| 984 | else if (LowerTiedOps) |
| 985 | Reg = getTiedUseReg(MI, DefMO); |
| 986 | if (Reg) { |
| 987 | Reg = traceCopyChain(Reg); |
| 988 | if (Reg.isValid()) |
| 989 | return Reg; |
| 990 | } |
| 991 | |
| 992 | if (++C >= DefLimit) |
| 993 | break; |
| 994 | } |
| 995 | return Register(); |
| 996 | } |
| 997 | |
| 998 | /// Allocates a physical register for VirtReg. |
| 999 | void RegAllocFastImpl::allocVirtReg(MachineInstr &MI, LiveReg &LR, |
| 1000 | Register Hint0, bool LookAtPhysRegUses) { |
| 1001 | const Register VirtReg = LR.VirtReg; |
| 1002 | assert(!LR.PhysReg); |
| 1003 | |
| 1004 | const TargetRegisterClass &RC = *MRI->getRegClass(Reg: VirtReg); |
| 1005 | LLVM_DEBUG(dbgs() << "Search register for " << printReg(VirtReg) |
| 1006 | << " in class " << TRI->getRegClassName(&RC) |
| 1007 | << " with hint " << printReg(Hint0, TRI) << '\n'); |
| 1008 | |
| 1009 | // Take hint when possible. |
| 1010 | if (Hint0.isPhysical() && MRI->isAllocatable(PhysReg: Hint0) && RC.contains(Reg: Hint0) && |
| 1011 | !isRegUsedInInstr(PhysReg: Hint0, LookAtPhysRegUses)) { |
| 1012 | // Take hint if the register is currently free. |
| 1013 | if (isPhysRegFree(PhysReg: Hint0)) { |
| 1014 | LLVM_DEBUG(dbgs() << "\tPreferred Register 1: " << printReg(Hint0, TRI) |
| 1015 | << '\n'); |
| 1016 | assignVirtToPhysReg(AtMI&: MI, LR, PhysReg: Hint0); |
| 1017 | return; |
| 1018 | } else { |
| 1019 | LLVM_DEBUG(dbgs() << "\tPreferred Register 0: " << printReg(Hint0, TRI) |
| 1020 | << " occupied\n" ); |
| 1021 | } |
| 1022 | } else { |
| 1023 | Hint0 = Register(); |
| 1024 | } |
| 1025 | |
| 1026 | // Try other hint. |
| 1027 | Register Hint1 = traceCopies(VirtReg); |
| 1028 | if (Hint1.isPhysical() && MRI->isAllocatable(PhysReg: Hint1) && RC.contains(Reg: Hint1) && |
| 1029 | !isRegUsedInInstr(PhysReg: Hint1, LookAtPhysRegUses)) { |
| 1030 | // Take hint if the register is currently free. |
| 1031 | if (isPhysRegFree(PhysReg: Hint1)) { |
| 1032 | LLVM_DEBUG(dbgs() << "\tPreferred Register 0: " << printReg(Hint1, TRI) |
| 1033 | << '\n'); |
| 1034 | assignVirtToPhysReg(AtMI&: MI, LR, PhysReg: Hint1); |
| 1035 | return; |
| 1036 | } else { |
| 1037 | LLVM_DEBUG(dbgs() << "\tPreferred Register 1: " << printReg(Hint1, TRI) |
| 1038 | << " occupied\n" ); |
| 1039 | } |
| 1040 | } else { |
| 1041 | Hint1 = Register(); |
| 1042 | } |
| 1043 | |
| 1044 | MCPhysReg BestReg = 0; |
| 1045 | unsigned BestCost = spillImpossible; |
| 1046 | ArrayRef<MCPhysReg> AllocationOrder = RegClassInfo.getOrder(RC: &RC); |
| 1047 | for (MCPhysReg PhysReg : AllocationOrder) { |
| 1048 | LLVM_DEBUG(dbgs() << "\tRegister: " << printReg(PhysReg, TRI) << ' '); |
| 1049 | if (isRegUsedInInstr(PhysReg, LookAtPhysRegUses)) { |
| 1050 | LLVM_DEBUG(dbgs() << "already used in instr.\n" ); |
| 1051 | continue; |
| 1052 | } |
| 1053 | |
| 1054 | unsigned Cost = calcSpillCost(PhysReg); |
| 1055 | LLVM_DEBUG(dbgs() << "Cost: " << Cost << " BestCost: " << BestCost << '\n'); |
| 1056 | // Immediate take a register with cost 0. |
| 1057 | if (Cost == 0) { |
| 1058 | assignVirtToPhysReg(AtMI&: MI, LR, PhysReg); |
| 1059 | return; |
| 1060 | } |
| 1061 | |
| 1062 | if (PhysReg == Hint0 || PhysReg == Hint1) |
| 1063 | Cost -= spillPrefBonus; |
| 1064 | |
| 1065 | if (Cost < BestCost) { |
| 1066 | BestReg = PhysReg; |
| 1067 | BestCost = Cost; |
| 1068 | } |
| 1069 | } |
| 1070 | |
| 1071 | if (!BestReg) { |
| 1072 | // Nothing we can do: Report an error and keep going with an invalid |
| 1073 | // allocation. |
| 1074 | LR.PhysReg = getErrorAssignment(LR, MI, RC); |
| 1075 | LR.Error = true; |
| 1076 | return; |
| 1077 | } |
| 1078 | |
| 1079 | displacePhysReg(MI, PhysReg: BestReg); |
| 1080 | assignVirtToPhysReg(AtMI&: MI, LR, PhysReg: BestReg); |
| 1081 | } |
| 1082 | |
| 1083 | void RegAllocFastImpl::allocVirtRegUndef(MachineOperand &MO) { |
| 1084 | assert(MO.isUndef() && "expected undef use" ); |
| 1085 | Register VirtReg = MO.getReg(); |
| 1086 | assert(VirtReg.isVirtual() && "Expected virtreg" ); |
| 1087 | if (!shouldAllocateRegister(Reg: VirtReg)) |
| 1088 | return; |
| 1089 | |
| 1090 | // If there are multiple undef uses, give them the same register. The def is |
| 1091 | // already freed, so take the register from the tie, not the lookup below. |
| 1092 | MachineInstr &MI = *MO.getParent(); |
| 1093 | for (const MachineOperand &Tied : MI.all_uses()) { |
| 1094 | if (!Tied.isTied() || Tied.getReg() != VirtReg) |
| 1095 | continue; |
| 1096 | MCRegister DefReg = getTiedOperand(MI, MO: Tied).getReg().asMCReg(); |
| 1097 | for (MachineOperand &O : MI.all_uses()) { |
| 1098 | if (O.getReg() != VirtReg) |
| 1099 | continue; |
| 1100 | // The def is already narrowed, so a tie takes its register whole. |
| 1101 | unsigned SubIdx = O.isTied() ? 0 : O.getSubReg(); |
| 1102 | O.setReg(SubIdx ? TRI->getSubReg(Reg: DefReg, Idx: SubIdx) : DefReg); |
| 1103 | O.setSubReg(0); |
| 1104 | O.setIsRenamable(!MRI->isReserved(PhysReg: O.getReg())); |
| 1105 | } |
| 1106 | return; |
| 1107 | } |
| 1108 | |
| 1109 | LiveRegMap::iterator LRI = findLiveVirtReg(VirtReg); |
| 1110 | MCRegister PhysReg; |
| 1111 | bool IsRenamable = true; |
| 1112 | if (LRI != LiveVirtRegs.end() && LRI->PhysReg) { |
| 1113 | PhysReg = LRI->PhysReg; |
| 1114 | } else { |
| 1115 | const TargetRegisterClass &RC = *MRI->getRegClass(Reg: VirtReg); |
| 1116 | ArrayRef<MCPhysReg> AllocationOrder = RegClassInfo.getOrder(RC: &RC); |
| 1117 | if (AllocationOrder.empty()) { |
| 1118 | // All registers in the class were reserved. |
| 1119 | // |
| 1120 | // It might be OK to take any entry from the class as this is an undef |
| 1121 | // use, but accepting this would give different behavior than greedy and |
| 1122 | // basic. |
| 1123 | PhysReg = getErrorAssignment(LR: *LRI, MI&: *MO.getParent(), RC); |
| 1124 | LRI->Error = true; |
| 1125 | IsRenamable = false; |
| 1126 | } else |
| 1127 | PhysReg = AllocationOrder.front(); |
| 1128 | } |
| 1129 | |
| 1130 | unsigned SubRegIdx = MO.getSubReg(); |
| 1131 | if (SubRegIdx != 0) { |
| 1132 | PhysReg = TRI->getSubReg(Reg: PhysReg, Idx: SubRegIdx); |
| 1133 | MO.setSubReg(0); |
| 1134 | } |
| 1135 | MO.setReg(PhysReg); |
| 1136 | MO.setIsRenamable(IsRenamable); |
| 1137 | } |
| 1138 | |
| 1139 | /// Variation of defineVirtReg() with special handling for livethrough regs |
| 1140 | /// (tied or earlyclobber) that may interfere with preassigned uses. |
| 1141 | /// \return true if MI's MachineOperands were re-arranged/invalidated. |
| 1142 | bool RegAllocFastImpl::defineLiveThroughVirtReg(MachineInstr &MI, |
| 1143 | unsigned OpNum, |
| 1144 | Register VirtReg) { |
| 1145 | if (!shouldAllocateRegister(Reg: VirtReg)) |
| 1146 | return false; |
| 1147 | LiveRegMap::iterator LRI = findLiveVirtReg(VirtReg); |
| 1148 | if (LRI != LiveVirtRegs.end()) { |
| 1149 | MCRegister PrevReg = LRI->PhysReg; |
| 1150 | if (PrevReg && isRegUsedInInstr(PhysReg: PrevReg, LookAtPhysRegUses: true)) { |
| 1151 | LLVM_DEBUG(dbgs() << "Need new assignment for " << printReg(PrevReg, TRI) |
| 1152 | << " (tied/earlyclobber resolution)\n" ); |
| 1153 | freePhysReg(PhysReg: PrevReg); |
| 1154 | LRI->PhysReg = MCRegister(); |
| 1155 | allocVirtReg(MI, LR&: *LRI, Hint0: Register(), LookAtPhysRegUses: true); |
| 1156 | MachineBasicBlock::iterator InsertBefore = |
| 1157 | std::next(x: (MachineBasicBlock::iterator)MI.getIterator()); |
| 1158 | LLVM_DEBUG(dbgs() << "Copy " << printReg(LRI->PhysReg, TRI) << " to " |
| 1159 | << printReg(PrevReg, TRI) << '\n'); |
| 1160 | BuildMI(BB&: *MBB, I: InsertBefore, MIMD: MI.getDebugLoc(), |
| 1161 | MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: PrevReg) |
| 1162 | .addReg(RegNo: LRI->PhysReg, Flags: llvm::RegState::Kill); |
| 1163 | } |
| 1164 | MachineOperand &MO = MI.getOperand(i: OpNum); |
| 1165 | if (MO.getSubReg() && !MO.isUndef()) { |
| 1166 | LRI->LastUse = &MI; |
| 1167 | } |
| 1168 | } |
| 1169 | return defineVirtReg(MI, OpNum, VirtReg, LookAtPhysRegUses: true); |
| 1170 | } |
| 1171 | |
| 1172 | /// Allocates a register for VirtReg definition. Typically the register is |
| 1173 | /// already assigned from a use of the virtreg, however we still need to |
| 1174 | /// perform an allocation if: |
| 1175 | /// - It is a dead definition without any uses. |
| 1176 | /// - The value is live out and all uses are in different basic blocks. |
| 1177 | /// |
| 1178 | /// \return true if MI's MachineOperands were re-arranged/invalidated. |
| 1179 | bool RegAllocFastImpl::defineVirtReg(MachineInstr &MI, unsigned OpNum, |
| 1180 | Register VirtReg, bool LookAtPhysRegUses) { |
| 1181 | assert(VirtReg.isVirtual() && "Not a virtual register" ); |
| 1182 | if (!shouldAllocateRegister(Reg: VirtReg)) |
| 1183 | return false; |
| 1184 | MachineOperand &MO = MI.getOperand(i: OpNum); |
| 1185 | LiveRegMap::iterator LRI; |
| 1186 | bool New; |
| 1187 | std::tie(args&: LRI, args&: New) = LiveVirtRegs.insert(Val: LiveReg(VirtReg)); |
| 1188 | if (New) { |
| 1189 | if (!MO.isDead()) { |
| 1190 | if (mayLiveOut(VirtReg)) { |
| 1191 | LRI->LiveOut = true; |
| 1192 | } else { |
| 1193 | // It is a dead def without the dead flag; add the flag now. |
| 1194 | MO.setIsDead(true); |
| 1195 | } |
| 1196 | } |
| 1197 | } |
| 1198 | if (!LRI->PhysReg) { |
| 1199 | allocVirtReg(MI, LR&: *LRI, Hint0: Register(), LookAtPhysRegUses); |
| 1200 | } else { |
| 1201 | assert((!isRegUsedInInstr(LRI->PhysReg, LookAtPhysRegUses) || LRI->Error) && |
| 1202 | "TODO: preassign mismatch" ); |
| 1203 | LLVM_DEBUG(dbgs() << "In def of " << printReg(VirtReg, TRI) |
| 1204 | << " use existing assignment to " |
| 1205 | << printReg(LRI->PhysReg, TRI) << '\n'); |
| 1206 | } |
| 1207 | |
| 1208 | MCRegister PhysReg = LRI->PhysReg; |
| 1209 | // Either flag means a reader below depends on the slot. |
| 1210 | if (LRI->Reloaded || LRI->LiveOut) { |
| 1211 | if (!MI.isImplicitDef()) { |
| 1212 | MachineBasicBlock::iterator SpillBefore = |
| 1213 | std::next(x: (MachineBasicBlock::iterator)MI.getIterator()); |
| 1214 | LLVM_DEBUG(dbgs() << "Spill Reason: LO: " << LRI->LiveOut |
| 1215 | << " RL: " << LRI->Reloaded << '\n'); |
| 1216 | bool Kill = LRI->LastUse == nullptr; |
| 1217 | spill(Before: SpillBefore, VirtReg, AssignedReg: PhysReg, Kill, LiveOut: LRI->LiveOut); |
| 1218 | |
| 1219 | // We need to place additional spills for each indirect destination of an |
| 1220 | // INLINEASM_BR. |
| 1221 | if (MI.getOpcode() == TargetOpcode::INLINEASM_BR) { |
| 1222 | int FI = StackSlotForVirtReg[VirtReg]; |
| 1223 | const TargetRegisterClass &RC = *MRI->getRegClass(Reg: VirtReg); |
| 1224 | for (MachineOperand &MO : MI.operands()) { |
| 1225 | if (MO.isMBB()) { |
| 1226 | MachineBasicBlock *Succ = MO.getMBB(); |
| 1227 | TII->storeRegToStackSlot(MBB&: *Succ, MI: Succ->begin(), SrcReg: PhysReg, isKill: Kill, FrameIndex: FI, |
| 1228 | RC: &RC, VReg: VirtReg); |
| 1229 | ++NumStores; |
| 1230 | Succ->addLiveIn(PhysReg); |
| 1231 | } |
| 1232 | } |
| 1233 | } |
| 1234 | |
| 1235 | LRI->LastUse = nullptr; |
| 1236 | } else if (!LRI->LastUse) { |
| 1237 | // No spill was inserted, so nothing below reads this def. |
| 1238 | MO.setIsDead(true); |
| 1239 | } |
| 1240 | // A def above spills only if a displacement above reloads again. |
| 1241 | LRI->LiveOut = false; |
| 1242 | LRI->Reloaded = false; |
| 1243 | } |
| 1244 | if (MI.getOpcode() == TargetOpcode::BUNDLE) { |
| 1245 | BundleVirtRegsMap[VirtReg] = *LRI; |
| 1246 | } |
| 1247 | markRegUsedInInstr(PhysReg); |
| 1248 | return setPhysReg(MI, MO, Assignment: *LRI); |
| 1249 | } |
| 1250 | |
| 1251 | /// Place MO's value in its tied def's register, by taking the register over or |
| 1252 | /// copying into it. Return false if useVirtReg() should finish MO. |
| 1253 | bool RegAllocFastImpl::lowerTiedUse(MachineInstr &MI, MachineOperand &MO, |
| 1254 | LiveReg &LR) { |
| 1255 | const MachineOperand &DefMO = getTiedOperand(MI, MO); |
| 1256 | assert(DefMO.getReg().isPhysical() && "tied def allocated before its use" ); |
| 1257 | MCRegister DefReg = DefMO.getReg().asMCReg(); |
| 1258 | unsigned SubReg = MO.getSubReg(); |
| 1259 | if (!LR.PhysReg) { |
| 1260 | // No register holds the value below MI, so it can live in DefReg, unless |
| 1261 | // MO reads a subregister, DefReg cannot hold the value, or an early-clobber |
| 1262 | // def would overwrite DefReg before another operand reads the value. |
| 1263 | bool MustCopy = SubReg || !MRI->isAllocatable(PhysReg: DefReg) || |
| 1264 | !MRI->getRegClass(Reg: LR.VirtReg)->contains(Reg: DefReg) || |
| 1265 | (DefMO.isEarlyClobber() && |
| 1266 | any_of(Range: MI.all_uses(), P: [&](const MachineOperand &O) { |
| 1267 | return &O != &MO && O.getReg() == LR.VirtReg; |
| 1268 | })); |
| 1269 | if (!MustCopy) { |
| 1270 | // The def is not live above MI, so the value can occupy DefReg there. |
| 1271 | freePhysReg(PhysReg: DefReg); |
| 1272 | assignVirtToPhysReg(AtMI&: MI, LR, PhysReg: DefReg); |
| 1273 | return false; |
| 1274 | } |
| 1275 | allocVirtReg(MI, LR, Hint0: Register(), LookAtPhysRegUses: false); |
| 1276 | // The def phase marked DefReg used in MI, so allocVirtReg skips it. |
| 1277 | assert((LR.Error || !TRI->regsOverlap(LR.PhysReg, DefReg)) && |
| 1278 | "copy source overlaps the tied def" ); |
| 1279 | } |
| 1280 | |
| 1281 | MCRegister SrcReg = SubReg ? TRI->getSubReg(Reg: LR.PhysReg, Idx: SubReg) : LR.PhysReg; |
| 1282 | // Only an already rewritten tie (%x = OP %x) finds the value in DefReg. |
| 1283 | if (SrcReg == DefReg) |
| 1284 | return false; |
| 1285 | |
| 1286 | // The copy reads SrcReg above MI, so no other operand may take it. |
| 1287 | BuildMI(BB&: *MBB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: DefReg) |
| 1288 | .addReg(RegNo: SrcReg); |
| 1289 | LR.LastUse = &MI; |
| 1290 | markRegUsedInInstr(PhysReg: LR.PhysReg); |
| 1291 | |
| 1292 | bool Renamable = !MRI->isReserved(PhysReg: DefReg); |
| 1293 | auto ReadDefReg = [&](MachineOperand &O) { |
| 1294 | O.setReg(DefReg); |
| 1295 | O.setSubReg(0); |
| 1296 | O.setIsRenamable(Renamable); |
| 1297 | }; |
| 1298 | ReadDefReg(MO); |
| 1299 | // The other reads of the value follow it into DefReg, so SrcReg dies at the |
| 1300 | // copy. They cannot when an early-clobber def overwrites DefReg first, and a |
| 1301 | // read tied to another def owes that def's register. |
| 1302 | if (!DefMO.isEarlyClobber()) { |
| 1303 | for (MachineOperand &O : MI.all_uses()) { |
| 1304 | if (O.isTied() || O.getReg() != LR.VirtReg || O.getSubReg() != SubReg) |
| 1305 | continue; |
| 1306 | ReadDefReg(O); |
| 1307 | O.setIsKill(false); |
| 1308 | } |
| 1309 | } |
| 1310 | |
| 1311 | // The free-defs step skips tied defs, so DefReg still holds the def. |
| 1312 | freePhysReg(PhysReg: DefReg); |
| 1313 | return true; |
| 1314 | } |
| 1315 | |
| 1316 | /// Allocates a register for a VirtReg use. |
| 1317 | /// \return true if MI's MachineOperands were re-arranged/invalidated. |
| 1318 | bool RegAllocFastImpl::useVirtReg(MachineInstr &MI, MachineOperand &MO, |
| 1319 | Register VirtReg) { |
| 1320 | assert(VirtReg.isVirtual() && "Not a virtual register" ); |
| 1321 | if (!shouldAllocateRegister(Reg: VirtReg)) |
| 1322 | return false; |
| 1323 | LiveRegMap::iterator LRI; |
| 1324 | bool New; |
| 1325 | std::tie(args&: LRI, args&: New) = LiveVirtRegs.insert(Val: LiveReg(VirtReg)); |
| 1326 | if (New) { |
| 1327 | if (!MO.isKill()) { |
| 1328 | if (mayLiveOut(VirtReg)) { |
| 1329 | LRI->LiveOut = true; |
| 1330 | } else { |
| 1331 | // It is a last (killing) use without the kill flag; add the flag now. |
| 1332 | MO.setIsKill(true); |
| 1333 | } |
| 1334 | } |
| 1335 | } else { |
| 1336 | assert((!MO.isKill() || LRI->LastUse == &MI) && "Invalid kill flag" ); |
| 1337 | } |
| 1338 | |
| 1339 | if (LowerTiedOps && MO.isTied() && lowerTiedUse(MI, MO, LR&: *LRI)) |
| 1340 | return false; |
| 1341 | |
| 1342 | // If necessary allocate a register. |
| 1343 | if (!LRI->PhysReg) { |
| 1344 | assert(!MO.isTied() && "tied op should be allocated" ); |
| 1345 | Register Hint; |
| 1346 | if (MI.isCopy() && MI.getOperand(i: 1).getSubReg() == 0) { |
| 1347 | Hint = MI.getOperand(i: 0).getReg(); |
| 1348 | if (Hint.isVirtual()) { |
| 1349 | assert(!shouldAllocateRegister(Hint)); |
| 1350 | Hint = Register(); |
| 1351 | } else { |
| 1352 | assert(Hint.isPhysical() && |
| 1353 | "Copy destination should already be assigned" ); |
| 1354 | } |
| 1355 | } |
| 1356 | allocVirtReg(MI, LR&: *LRI, Hint0: Hint, LookAtPhysRegUses: false); |
| 1357 | } |
| 1358 | |
| 1359 | LRI->LastUse = &MI; |
| 1360 | |
| 1361 | if (MI.getOpcode() == TargetOpcode::BUNDLE) { |
| 1362 | BundleVirtRegsMap[VirtReg] = *LRI; |
| 1363 | } |
| 1364 | markRegUsedInInstr(PhysReg: LRI->PhysReg); |
| 1365 | return setPhysReg(MI, MO, Assignment: *LRI); |
| 1366 | } |
| 1367 | |
| 1368 | /// Query a physical register to use as a filler in contexts where the |
| 1369 | /// allocation has failed. This will raise an error, but not abort the |
| 1370 | /// compilation. |
| 1371 | MCPhysReg RegAllocFastImpl::getErrorAssignment(const LiveReg &LR, |
| 1372 | MachineInstr &MI, |
| 1373 | const TargetRegisterClass &RC) { |
| 1374 | MachineFunction &MF = *MI.getMF(); |
| 1375 | |
| 1376 | // Avoid repeating the error every time a register is used. |
| 1377 | bool EmitError = !MF.getProperties().hasFailedRegAlloc(); |
| 1378 | if (EmitError) |
| 1379 | MF.getProperties().setFailedRegAlloc(); |
| 1380 | |
| 1381 | // If the allocation order was empty, all registers in the class were |
| 1382 | // probably reserved. Fall back to taking the first register in the class, |
| 1383 | // even if it's reserved. |
| 1384 | ArrayRef<MCPhysReg> AllocationOrder = RegClassInfo.getOrder(RC: &RC); |
| 1385 | if (AllocationOrder.empty()) { |
| 1386 | const Function &Fn = MF.getFunction(); |
| 1387 | if (EmitError) { |
| 1388 | Fn.getContext().diagnose(DI: DiagnosticInfoRegAllocFailure( |
| 1389 | "no registers from class available to allocate" , Fn, |
| 1390 | MI.getDebugLoc())); |
| 1391 | } |
| 1392 | |
| 1393 | ArrayRef<MCPhysReg> RawRegs = RC.getRegisters(); |
| 1394 | assert(!RawRegs.empty() && "register classes cannot have no registers" ); |
| 1395 | return RawRegs.front(); |
| 1396 | } |
| 1397 | |
| 1398 | if (!LR.Error && EmitError) { |
| 1399 | // Nothing we can do: Report an error and keep going with an invalid |
| 1400 | // allocation. |
| 1401 | if (MI.isInlineAsm()) { |
| 1402 | MI.emitInlineAsmError( |
| 1403 | ErrMsg: "inline assembly requires more registers than available" ); |
| 1404 | } else { |
| 1405 | const Function &Fn = MBB->getParent()->getFunction(); |
| 1406 | Fn.getContext().diagnose(DI: DiagnosticInfoRegAllocFailure( |
| 1407 | "ran out of registers during register allocation" , Fn, |
| 1408 | MI.getDebugLoc())); |
| 1409 | } |
| 1410 | } |
| 1411 | |
| 1412 | return AllocationOrder.front(); |
| 1413 | } |
| 1414 | |
| 1415 | /// Changes operand OpNum in MI the refer the PhysReg, considering subregs. |
| 1416 | /// \return true if MI's MachineOperands were re-arranged/invalidated. |
| 1417 | bool RegAllocFastImpl::setPhysReg(MachineInstr &MI, MachineOperand &MO, |
| 1418 | const LiveReg &Assignment) { |
| 1419 | MCRegister PhysReg = Assignment.PhysReg; |
| 1420 | assert(PhysReg && "assignments should always be to a valid physreg" ); |
| 1421 | |
| 1422 | if (LLVM_UNLIKELY(Assignment.Error)) { |
| 1423 | // Make sure we don't set renamable in error scenarios, as we may have |
| 1424 | // assigned to a reserved register. |
| 1425 | if (MO.isUse()) |
| 1426 | MO.setIsUndef(true); |
| 1427 | } |
| 1428 | |
| 1429 | if (!MO.getSubReg()) { |
| 1430 | MO.setReg(PhysReg); |
| 1431 | MO.setIsRenamable(!Assignment.Error); |
| 1432 | return false; |
| 1433 | } |
| 1434 | |
| 1435 | // Handle subregister index. |
| 1436 | MO.setReg(TRI->getSubReg(Reg: PhysReg, Idx: MO.getSubReg())); |
| 1437 | MO.setIsRenamable(!Assignment.Error); |
| 1438 | |
| 1439 | // Note: We leave the subreg number around a little longer in case of defs. |
| 1440 | // This is so that the register freeing logic in allocateInstruction can still |
| 1441 | // recognize this as subregister defs. The code there will clear the number. |
| 1442 | if (!MO.isDef()) |
| 1443 | MO.setSubReg(0); |
| 1444 | |
| 1445 | // A kill flag implies killing the full register. Add corresponding super |
| 1446 | // register kill. |
| 1447 | if (MO.isKill()) { |
| 1448 | MI.addRegisterKilled(IncomingReg: PhysReg, RegInfo: TRI, AddIfNotFound: true); |
| 1449 | // Conservatively assume implicit MOs were re-arranged |
| 1450 | return true; |
| 1451 | } |
| 1452 | |
| 1453 | // A <def,read-undef> of a sub-register requires an implicit def of the full |
| 1454 | // register. |
| 1455 | if (MO.isDef() && MO.isUndef()) { |
| 1456 | if (MO.isDead()) |
| 1457 | MI.addRegisterDead(Reg: PhysReg, RegInfo: TRI, AddIfNotFound: true); |
| 1458 | else |
| 1459 | MI.addRegisterDefined(Reg: PhysReg, RegInfo: TRI); |
| 1460 | // Conservatively assume implicit MOs were re-arranged |
| 1461 | return true; |
| 1462 | } |
| 1463 | return false; |
| 1464 | } |
| 1465 | |
| 1466 | #ifndef NDEBUG |
| 1467 | |
| 1468 | void RegAllocFastImpl::dumpState() const { |
| 1469 | for (MCRegUnit Unit : TRI->regunits()) { |
| 1470 | switch (unsigned VirtReg = getRegUnitState(Unit)) { |
| 1471 | case regFree: |
| 1472 | break; |
| 1473 | case regPreAssigned: |
| 1474 | dbgs() << " " << printRegUnit(Unit, TRI) << "[P]" ; |
| 1475 | break; |
| 1476 | case regLiveIn: |
| 1477 | llvm_unreachable("Should not have regLiveIn in map" ); |
| 1478 | default: { |
| 1479 | dbgs() << ' ' << printRegUnit(Unit, TRI) << '=' << printReg(VirtReg); |
| 1480 | LiveRegMap::const_iterator I = findLiveVirtReg(VirtReg); |
| 1481 | assert(I != LiveVirtRegs.end() && "have LiveVirtRegs entry" ); |
| 1482 | if (I->LiveOut || I->Reloaded) { |
| 1483 | dbgs() << '['; |
| 1484 | if (I->LiveOut) |
| 1485 | dbgs() << 'O'; |
| 1486 | if (I->Reloaded) |
| 1487 | dbgs() << 'R'; |
| 1488 | dbgs() << ']'; |
| 1489 | } |
| 1490 | assert(TRI->hasRegUnit(I->PhysReg, Unit) && "inverse mapping present" ); |
| 1491 | break; |
| 1492 | } |
| 1493 | } |
| 1494 | } |
| 1495 | dbgs() << '\n'; |
| 1496 | // Check that LiveVirtRegs is the inverse. |
| 1497 | for (const LiveReg &LR : LiveVirtRegs) { |
| 1498 | Register VirtReg = LR.VirtReg; |
| 1499 | assert(VirtReg.isVirtual() && "Bad map key" ); |
| 1500 | MCRegister PhysReg = LR.PhysReg; |
| 1501 | if (PhysReg) { |
| 1502 | assert(PhysReg.isPhysical() && "mapped to physreg" ); |
| 1503 | for (MCRegUnit Unit : TRI->regunits(PhysReg)) { |
| 1504 | assert(getRegUnitState(Unit) == VirtReg && "inverse map valid" ); |
| 1505 | } |
| 1506 | } |
| 1507 | } |
| 1508 | } |
| 1509 | #endif |
| 1510 | |
| 1511 | /// Count number of defs consumed from each register class by \p Reg |
| 1512 | void RegAllocFastImpl::addRegClassDefCounts( |
| 1513 | MutableArrayRef<unsigned> RegClassDefCounts, Register Reg) const { |
| 1514 | assert(RegClassDefCounts.size() == TRI->getNumRegClasses()); |
| 1515 | |
| 1516 | if (Reg.isVirtual()) { |
| 1517 | if (!shouldAllocateRegister(Reg)) |
| 1518 | return; |
| 1519 | const TargetRegisterClass *OpRC = MRI->getRegClass(Reg); |
| 1520 | for (unsigned RCIdx = 0, RCIdxEnd = TRI->getNumRegClasses(); |
| 1521 | RCIdx != RCIdxEnd; ++RCIdx) { |
| 1522 | const TargetRegisterClass *IdxRC = TRI->getRegClass(i: RCIdx); |
| 1523 | // FIXME: Consider aliasing sub/super registers. |
| 1524 | if (OpRC->hasSubClassEq(RC: IdxRC)) |
| 1525 | ++RegClassDefCounts[RCIdx]; |
| 1526 | } |
| 1527 | |
| 1528 | return; |
| 1529 | } |
| 1530 | |
| 1531 | for (unsigned RCIdx = 0, RCIdxEnd = TRI->getNumRegClasses(); |
| 1532 | RCIdx != RCIdxEnd; ++RCIdx) { |
| 1533 | const TargetRegisterClass *IdxRC = TRI->getRegClass(i: RCIdx); |
| 1534 | for (MCRegAliasIterator Alias(Reg, TRI, true); Alias.isValid(); ++Alias) { |
| 1535 | if (IdxRC->contains(Reg: *Alias)) { |
| 1536 | ++RegClassDefCounts[RCIdx]; |
| 1537 | break; |
| 1538 | } |
| 1539 | } |
| 1540 | } |
| 1541 | } |
| 1542 | |
| 1543 | /// Early clobber, partial def, or tied to a use that carries a value: the |
| 1544 | /// register is occupied while the uses are read. |
| 1545 | static bool isLiveThroughDef(const MachineInstr &MI, const MachineOperand &MO) { |
| 1546 | assert(MO.isDef() && "expected def operand" ); |
| 1547 | if (MO.isEarlyClobber() || MO.readsReg()) |
| 1548 | return true; |
| 1549 | return MO.isTied() && |
| 1550 | !MI.getOperand(i: MI.findTiedOperandIdx(OpIdx: MI.getOperandNo(I: &MO))).isUndef(); |
| 1551 | } |
| 1552 | |
| 1553 | /// Compute \ref DefOperandIndexes so it contains the indices of "def" operands |
| 1554 | /// that are to be allocated. Those are ordered in a way that small classes, |
| 1555 | /// early clobbers and livethroughs are allocated first. |
| 1556 | void RegAllocFastImpl::findAndSortDefOperandIndexes(const MachineInstr &MI) { |
| 1557 | DefOperandIndexes.clear(); |
| 1558 | |
| 1559 | LLVM_DEBUG(dbgs() << "Need to assign livethroughs\n" ); |
| 1560 | for (unsigned I = 0, E = MI.getNumOperands(); I < E; ++I) { |
| 1561 | const MachineOperand &MO = MI.getOperand(i: I); |
| 1562 | if (!MO.isReg()) |
| 1563 | continue; |
| 1564 | Register Reg = MO.getReg(); |
| 1565 | if (MO.readsReg()) { |
| 1566 | if (Reg.isPhysical()) { |
| 1567 | LLVM_DEBUG(dbgs() << "mark extra used: " << printReg(Reg, TRI) << '\n'); |
| 1568 | markPhysRegUsedInInstr(PhysReg: Reg); |
| 1569 | } |
| 1570 | } |
| 1571 | |
| 1572 | if (MO.isDef() && Reg.isVirtual() && shouldAllocateRegister(Reg)) |
| 1573 | DefOperandIndexes.push_back(Elt: I); |
| 1574 | } |
| 1575 | |
| 1576 | // Most instructions only have one virtual def, so there's no point in |
| 1577 | // computing the possible number of defs for every register class. |
| 1578 | if (DefOperandIndexes.size() <= 1) |
| 1579 | return; |
| 1580 | |
| 1581 | // Track number of defs which may consume a register from the class. This is |
| 1582 | // used to assign registers for possibly-too-small classes first. Example: |
| 1583 | // defs are eax, 3 * gr32_abcd, 2 * gr32 => we want to assign the gr32_abcd |
| 1584 | // registers first so that the gr32 don't use the gr32_abcd registers before |
| 1585 | // we assign these. |
| 1586 | SmallVector<unsigned> RegClassDefCounts(TRI->getNumRegClasses(), 0); |
| 1587 | |
| 1588 | for (const MachineOperand &MO : MI.all_defs()) |
| 1589 | addRegClassDefCounts(RegClassDefCounts, Reg: MO.getReg()); |
| 1590 | |
| 1591 | llvm::sort(C&: DefOperandIndexes, Comp: [&](unsigned I0, unsigned I1) { |
| 1592 | const MachineOperand &MO0 = MI.getOperand(i: I0); |
| 1593 | const MachineOperand &MO1 = MI.getOperand(i: I1); |
| 1594 | Register Reg0 = MO0.getReg(); |
| 1595 | Register Reg1 = MO1.getReg(); |
| 1596 | const TargetRegisterClass &RC0 = *MRI->getRegClass(Reg: Reg0); |
| 1597 | const TargetRegisterClass &RC1 = *MRI->getRegClass(Reg: Reg1); |
| 1598 | |
| 1599 | // Identify regclass that are easy to use up completely just in this |
| 1600 | // instruction. |
| 1601 | unsigned ClassSize0 = RegClassInfo.getOrder(RC: &RC0).size(); |
| 1602 | unsigned ClassSize1 = RegClassInfo.getOrder(RC: &RC1).size(); |
| 1603 | |
| 1604 | bool SmallClass0 = ClassSize0 < RegClassDefCounts[RC0.getID()]; |
| 1605 | bool SmallClass1 = ClassSize1 < RegClassDefCounts[RC1.getID()]; |
| 1606 | if (SmallClass0 > SmallClass1) |
| 1607 | return true; |
| 1608 | if (SmallClass0 < SmallClass1) |
| 1609 | return false; |
| 1610 | |
| 1611 | // Allocate early clobbers and livethrough operands first. |
| 1612 | bool Livethrough0 = isLiveThroughDef(MI, MO: MO0); |
| 1613 | bool Livethrough1 = isLiveThroughDef(MI, MO: MO1); |
| 1614 | if (Livethrough0 > Livethrough1) |
| 1615 | return true; |
| 1616 | if (Livethrough0 < Livethrough1) |
| 1617 | return false; |
| 1618 | |
| 1619 | // Tie-break rule: operand index. |
| 1620 | return I0 < I1; |
| 1621 | }); |
| 1622 | } |
| 1623 | |
| 1624 | void RegAllocFastImpl::allocateInstruction(MachineInstr &MI) { |
| 1625 | // Backwards, a def frees a register and a use occupies it. The phases: |
| 1626 | // * pre-assigned physreg defs |
| 1627 | // * virtual register defs |
| 1628 | // * free the def operands' registers |
| 1629 | // * displace registers clobbered by regmasks |
| 1630 | // * pre-assigned physreg uses |
| 1631 | // * virtual register uses, inserting reloads and tied-operand copies |
| 1632 | // * undef uses |
| 1633 | // * free early-clobber defs |
| 1634 | // |
| 1635 | // Freeing follows the def allocation so a def is not handed a register this |
| 1636 | // instruction also writes, and precedes the uses so a use may take one. It |
| 1637 | // skips tied defs, whose register the tied use reads, and early-clobber defs, |
| 1638 | // freed last so that no use lands on them. |
| 1639 | |
| 1640 | InstrGen += 2; |
| 1641 | // In the event we ever get more than 2**31 instructions... |
| 1642 | if (LLVM_UNLIKELY(InstrGen == 0)) { |
| 1643 | UsedInInstr.assign(NumElts: UsedInInstr.size(), Elt: 0); |
| 1644 | LiveDefUnits.assign(NumElts: LiveDefUnits.size(), Elt: 0); |
| 1645 | InstrGen = 2; |
| 1646 | } |
| 1647 | RegMasks.clear(); |
| 1648 | BundleVirtRegsMap.clear(); |
| 1649 | |
| 1650 | // Scan for special cases; Apply pre-assigned register defs to state. |
| 1651 | bool HasPhysRegUse = false; |
| 1652 | bool HasRegMask = false; |
| 1653 | bool HasVRegDef = false; |
| 1654 | bool HasDef = false; |
| 1655 | bool HasEarlyClobber = false; |
| 1656 | bool HasTiedDef = false; |
| 1657 | bool NeedToAssignLiveThroughs = false; |
| 1658 | for (MachineOperand &MO : MI.operands()) { |
| 1659 | if (MO.isReg()) { |
| 1660 | Register Reg = MO.getReg(); |
| 1661 | if (Reg.isVirtual()) { |
| 1662 | if (!shouldAllocateRegister(Reg)) |
| 1663 | continue; |
| 1664 | if (MO.isDef()) { |
| 1665 | HasDef = true; |
| 1666 | HasVRegDef = true; |
| 1667 | if (MO.isEarlyClobber()) |
| 1668 | HasEarlyClobber = true; |
| 1669 | if (LowerTiedOps && MO.isTied()) |
| 1670 | HasTiedDef = true; |
| 1671 | if (isLiveThroughDef(MI, MO)) |
| 1672 | NeedToAssignLiveThroughs = true; |
| 1673 | } |
| 1674 | } else if (Reg.isPhysical()) { |
| 1675 | if (!MRI->isReserved(PhysReg: Reg)) { |
| 1676 | if (MO.isDef()) { |
| 1677 | HasDef = true; |
| 1678 | bool displacedAny = definePhysReg(MI, Reg); |
| 1679 | if (MO.isEarlyClobber()) |
| 1680 | HasEarlyClobber = true; |
| 1681 | if (!displacedAny) |
| 1682 | MO.setIsDead(true); |
| 1683 | if (!MO.isDead()) |
| 1684 | markLiveDefUnits(PhysReg: Reg.asMCReg()); |
| 1685 | } |
| 1686 | if (MO.readsReg()) |
| 1687 | HasPhysRegUse = true; |
| 1688 | } |
| 1689 | } |
| 1690 | } else if (MO.isRegMask()) { |
| 1691 | HasRegMask = true; |
| 1692 | RegMasks.push_back(Elt: MO.getRegMask()); |
| 1693 | } |
| 1694 | } |
| 1695 | |
| 1696 | // Allocate virtreg defs. |
| 1697 | if (HasDef) { |
| 1698 | if (HasVRegDef) { |
| 1699 | // Note that Implicit MOs can get re-arranged by defineVirtReg(), so loop |
| 1700 | // multiple times to ensure no operand is missed. |
| 1701 | bool ReArrangedImplicitOps = true; |
| 1702 | |
| 1703 | // Special handling for early clobbers, tied operands or subregister defs: |
| 1704 | // Compared to "normal" defs these: |
| 1705 | // - Must not use a register that is pre-assigned for a use operand. |
| 1706 | // - In order to solve tricky inline assembly constraints we change the |
| 1707 | // heuristic to figure out a good operand order before doing |
| 1708 | // assignments. |
| 1709 | if (NeedToAssignLiveThroughs) { |
| 1710 | while (ReArrangedImplicitOps) { |
| 1711 | ReArrangedImplicitOps = false; |
| 1712 | findAndSortDefOperandIndexes(MI); |
| 1713 | for (unsigned OpIdx : DefOperandIndexes) { |
| 1714 | MachineOperand &MO = MI.getOperand(i: OpIdx); |
| 1715 | LLVM_DEBUG(dbgs() << "Allocating " << MO << '\n'); |
| 1716 | Register Reg = MO.getReg(); |
| 1717 | if (isLiveThroughDef(MI, MO)) { |
| 1718 | ReArrangedImplicitOps = defineLiveThroughVirtReg(MI, OpNum: OpIdx, VirtReg: Reg); |
| 1719 | } else { |
| 1720 | ReArrangedImplicitOps = defineVirtReg(MI, OpNum: OpIdx, VirtReg: Reg); |
| 1721 | } |
| 1722 | // Implicit operands of MI were re-arranged, |
| 1723 | // re-compute DefOperandIndexes. |
| 1724 | if (ReArrangedImplicitOps) |
| 1725 | break; |
| 1726 | } |
| 1727 | } |
| 1728 | } else { |
| 1729 | // Assign virtual register defs. |
| 1730 | while (ReArrangedImplicitOps) { |
| 1731 | ReArrangedImplicitOps = false; |
| 1732 | for (MachineOperand &MO : MI.all_defs()) { |
| 1733 | Register Reg = MO.getReg(); |
| 1734 | if (Reg.isVirtual()) { |
| 1735 | ReArrangedImplicitOps = |
| 1736 | defineVirtReg(MI, OpNum: MI.getOperandNo(I: &MO), VirtReg: Reg); |
| 1737 | if (ReArrangedImplicitOps) |
| 1738 | break; |
| 1739 | } |
| 1740 | } |
| 1741 | } |
| 1742 | } |
| 1743 | } |
| 1744 | |
| 1745 | // Free registers occupied by defs. |
| 1746 | // Iterate operands in reverse order, so we see the implicit super register |
| 1747 | // defs first (we added them earlier in case of <def,read-undef>). |
| 1748 | for (MachineOperand &MO : reverse(C: MI.all_defs())) { |
| 1749 | Register Reg = MO.getReg(); |
| 1750 | |
| 1751 | // A dead def whose register units are all covered by non-dead aliasing |
| 1752 | // defs is kept alive by them, so clear the inconsistent dead flag. |
| 1753 | if (Reg.isPhysical() && MO.isDead() && hasLiveDefUnits(PhysReg: Reg.asMCReg())) |
| 1754 | MO.setIsDead(false); |
| 1755 | |
| 1756 | // subreg defs don't free the full register. We left the subreg number |
| 1757 | // around as a marker in setPhysReg() to recognize this case here. |
| 1758 | if (Reg.isPhysical() && MO.getSubReg() != 0) { |
| 1759 | MO.setSubReg(0); |
| 1760 | continue; |
| 1761 | } |
| 1762 | |
| 1763 | assert((!MO.isTied() || !isClobberedByRegMasks(MO.getReg())) && |
| 1764 | "tied def assigned to clobbered register" ); |
| 1765 | |
| 1766 | // Do not free live-through defs. |
| 1767 | if (isLiveThroughDef(MI, MO)) |
| 1768 | continue; |
| 1769 | if (!Reg) |
| 1770 | continue; |
| 1771 | if (Reg.isVirtual()) { |
| 1772 | assert(!shouldAllocateRegister(Reg)); |
| 1773 | continue; |
| 1774 | } |
| 1775 | assert(Reg.isPhysical()); |
| 1776 | if (MRI->isReserved(PhysReg: Reg)) |
| 1777 | continue; |
| 1778 | freePhysReg(PhysReg: Reg); |
| 1779 | unmarkRegUsedInInstr(PhysReg: Reg); |
| 1780 | } |
| 1781 | } |
| 1782 | |
| 1783 | // A regmask is a def of every clobbered register: reload what lives in one |
| 1784 | // below MI. Nothing is reserved, so the uses may still take those registers. |
| 1785 | if (HasRegMask) { |
| 1786 | assert(!RegMasks.empty() && "expected RegMask" ); |
| 1787 | // MRI bookkeeping. |
| 1788 | for (const auto *RM : RegMasks) |
| 1789 | MRI->addPhysRegsUsedFromRegMask(RegMask: RM); |
| 1790 | |
| 1791 | for (const LiveReg &LR : LiveVirtRegs) { |
| 1792 | MCRegister PhysReg = LR.PhysReg; |
| 1793 | if (PhysReg && isClobberedByRegMasks(PhysReg)) |
| 1794 | displacePhysReg(MI, PhysReg); |
| 1795 | } |
| 1796 | } |
| 1797 | |
| 1798 | // Apply pre-assigned register uses to state. |
| 1799 | if (HasPhysRegUse) { |
| 1800 | for (MachineOperand &MO : MI.operands()) { |
| 1801 | if (!MO.isReg() || !MO.readsReg()) |
| 1802 | continue; |
| 1803 | Register Reg = MO.getReg(); |
| 1804 | if (!Reg.isPhysical()) |
| 1805 | continue; |
| 1806 | if (MRI->isReserved(PhysReg: Reg)) |
| 1807 | continue; |
| 1808 | if (!usePhysReg(MI, Reg)) |
| 1809 | MO.setIsKill(true); |
| 1810 | } |
| 1811 | } |
| 1812 | |
| 1813 | // Allocate virtreg uses and insert reloads as necessary. |
| 1814 | // Implicit MOs can get moved/removed by useVirtReg(), so loop multiple |
| 1815 | // times to ensure no operand is missed. |
| 1816 | bool HasUndefUse = false; |
| 1817 | bool TiedOnly = HasTiedDef; |
| 1818 | bool ReArrangedImplicitMOs = true; |
| 1819 | while (ReArrangedImplicitMOs) { |
| 1820 | ReArrangedImplicitMOs = false; |
| 1821 | for (MachineOperand &MO : MI.operands()) { |
| 1822 | if (!MO.isReg() || !MO.isUse()) |
| 1823 | continue; |
| 1824 | Register Reg = MO.getReg(); |
| 1825 | if (!Reg.isVirtual() || !shouldAllocateRegister(Reg) || |
| 1826 | (TiedOnly && !MO.isTied())) |
| 1827 | continue; |
| 1828 | |
| 1829 | if (MO.isUndef()) { |
| 1830 | HasUndefUse = true; |
| 1831 | continue; |
| 1832 | } |
| 1833 | |
| 1834 | // Populate MayLiveAcrossBlocks now: these uses are about to be rewritten |
| 1835 | // to physregs, so a def block allocated later can no longer see them. |
| 1836 | mayLiveIn(VirtReg: Reg); |
| 1837 | |
| 1838 | assert(!MO.isInternalRead() && "Bundles not supported" ); |
| 1839 | assert(MO.readsReg() && "reading use" ); |
| 1840 | ReArrangedImplicitMOs = useVirtReg(MI, MO, VirtReg: Reg); |
| 1841 | if (ReArrangedImplicitMOs) |
| 1842 | break; |
| 1843 | } |
| 1844 | // Given %1 = OP %0, %0(tied-def 0), allocate tied %0 first, so that %0 |
| 1845 | // takes %1's register. In operand order, the untied %0 would take another |
| 1846 | // register and the tied use would need a copy. |
| 1847 | if (TiedOnly && !ReArrangedImplicitMOs) { |
| 1848 | TiedOnly = false; |
| 1849 | ReArrangedImplicitMOs = true; |
| 1850 | } |
| 1851 | } |
| 1852 | |
| 1853 | // Allocate undef operands. This is a separate step because in a situation |
| 1854 | // like ` = OP undef %X, %X` both operands need the same register assign |
| 1855 | // so we should perform the normal assignment first. |
| 1856 | if (HasUndefUse) { |
| 1857 | for (MachineOperand &MO : MI.all_uses()) { |
| 1858 | Register Reg = MO.getReg(); |
| 1859 | if (!Reg.isVirtual() || !shouldAllocateRegister(Reg)) |
| 1860 | continue; |
| 1861 | |
| 1862 | assert(MO.isUndef() && "Should only have undef virtreg uses left" ); |
| 1863 | allocVirtRegUndef(MO); |
| 1864 | } |
| 1865 | } |
| 1866 | |
| 1867 | // Free early clobbers. Last, because they must not share a register with any |
| 1868 | // use. |
| 1869 | if (HasEarlyClobber) { |
| 1870 | for (MachineOperand &MO : reverse(C: MI.all_defs())) { |
| 1871 | if (!MO.isEarlyClobber()) |
| 1872 | continue; |
| 1873 | assert(!MO.getSubReg() && "should be already handled in def processing" ); |
| 1874 | |
| 1875 | Register Reg = MO.getReg(); |
| 1876 | if (!Reg) |
| 1877 | continue; |
| 1878 | if (Reg.isVirtual()) { |
| 1879 | assert(!shouldAllocateRegister(Reg)); |
| 1880 | continue; |
| 1881 | } |
| 1882 | assert(Reg.isPhysical() && "should have register assigned" ); |
| 1883 | |
| 1884 | // We sometimes get odd situations like: |
| 1885 | // early-clobber %x0 = INSTRUCTION %x0 |
| 1886 | // which is semantically questionable as the early-clobber should |
| 1887 | // apply before the use. But in practice we consider the use to |
| 1888 | // happen before the early clobber now. Don't free the early clobber |
| 1889 | // register in this case. |
| 1890 | if (MI.readsRegister(Reg, TRI)) |
| 1891 | continue; |
| 1892 | |
| 1893 | freePhysReg(PhysReg: Reg); |
| 1894 | } |
| 1895 | } |
| 1896 | |
| 1897 | LLVM_DEBUG(dbgs() << "<< " << MI); |
| 1898 | if (MI.isCopy() && |
| 1899 | (MI.getOperand(i: 0).getReg() == MI.getOperand(i: 1).getReg() || |
| 1900 | MI.getOperand(i: 0).isDead()) && |
| 1901 | MI.getNumOperands() == 2) { |
| 1902 | LLVM_DEBUG(dbgs() << "Mark unnecessary copy for removal: " << MI); |
| 1903 | Coalesced.push_back(Elt: &MI); |
| 1904 | } |
| 1905 | } |
| 1906 | |
| 1907 | void RegAllocFastImpl::handleDebugValue(MachineInstr &MI) { |
| 1908 | // Ignore DBG_VALUEs that aren't based on virtual registers. These are |
| 1909 | // mostly constants and frame indices. |
| 1910 | assert(MI.isDebugValue() && "not a DBG_VALUE*" ); |
| 1911 | for (const auto &MO : MI.debug_operands()) { |
| 1912 | if (!MO.isReg()) |
| 1913 | continue; |
| 1914 | Register Reg = MO.getReg(); |
| 1915 | if (!Reg.isVirtual()) |
| 1916 | continue; |
| 1917 | if (!shouldAllocateRegister(Reg)) |
| 1918 | continue; |
| 1919 | |
| 1920 | // Already spilled to a stackslot? |
| 1921 | int SS = StackSlotForVirtReg[Reg]; |
| 1922 | if (SS != -1) { |
| 1923 | // Modify DBG_VALUE now that the value is in a spill slot. |
| 1924 | updateDbgValueForSpill(Orig&: MI, FrameIndex: SS, Reg); |
| 1925 | LLVM_DEBUG(dbgs() << "Rewrite DBG_VALUE for spilled memory: " << MI); |
| 1926 | continue; |
| 1927 | } |
| 1928 | |
| 1929 | // See if this virtual register has already been allocated to a physical |
| 1930 | // register or spilled to a stack slot. |
| 1931 | LiveRegMap::iterator LRI = findLiveVirtReg(VirtReg: Reg); |
| 1932 | SmallVector<MachineOperand *> DbgOps( |
| 1933 | llvm::make_pointer_range(Range: MI.getDebugOperandsForReg(Reg))); |
| 1934 | |
| 1935 | if (LRI != LiveVirtRegs.end() && LRI->PhysReg) { |
| 1936 | // Update every use of Reg within MI. |
| 1937 | for (auto &RegMO : DbgOps) |
| 1938 | setPhysReg(MI, MO&: *RegMO, Assignment: *LRI); |
| 1939 | } else { |
| 1940 | DanglingDbgValues[Reg].push_back(Elt: &MI); |
| 1941 | } |
| 1942 | |
| 1943 | // If Reg hasn't been spilled, put this DBG_VALUE in LiveDbgValueMap so |
| 1944 | // that future spills of Reg will have DBG_VALUEs. |
| 1945 | LiveDbgValueMap[Reg].append(in_start: DbgOps.begin(), in_end: DbgOps.end()); |
| 1946 | } |
| 1947 | } |
| 1948 | |
| 1949 | void RegAllocFastImpl::handleBundle(MachineInstr &MI) { |
| 1950 | MachineBasicBlock::instr_iterator BundledMI = MI.getIterator(); |
| 1951 | ++BundledMI; |
| 1952 | while (BundledMI->isBundledWithPred()) { |
| 1953 | for (MachineOperand &MO : BundledMI->operands()) { |
| 1954 | if (!MO.isReg()) |
| 1955 | continue; |
| 1956 | |
| 1957 | Register Reg = MO.getReg(); |
| 1958 | if (!Reg.isVirtual() || !shouldAllocateRegister(Reg)) |
| 1959 | continue; |
| 1960 | |
| 1961 | auto DI = BundleVirtRegsMap.find(Val: Reg); |
| 1962 | assert(DI != BundleVirtRegsMap.end() && "Unassigned virtual register" ); |
| 1963 | |
| 1964 | setPhysReg(MI, MO, Assignment: DI->second); |
| 1965 | } |
| 1966 | |
| 1967 | ++BundledMI; |
| 1968 | } |
| 1969 | } |
| 1970 | |
| 1971 | void RegAllocFastImpl::allocateBasicBlock(MachineBasicBlock &MBB) { |
| 1972 | this->MBB = &MBB; |
| 1973 | LLVM_DEBUG(dbgs() << "\nAllocating " << MBB); |
| 1974 | |
| 1975 | PosIndexes.unsetInitialized(); |
| 1976 | RegUnitStates.assign(n: TRI->getNumRegUnits(), val: regFree); |
| 1977 | assert(LiveVirtRegs.empty() && "Mapping not cleared from last block?" ); |
| 1978 | |
| 1979 | for (const auto &LiveReg : MBB.liveouts()) |
| 1980 | setPhysRegState(PhysReg: LiveReg.PhysReg, NewState: regPreAssigned); |
| 1981 | |
| 1982 | Coalesced.clear(); |
| 1983 | |
| 1984 | // Lowering a tied operand inserts a copy ahead of MI. Its registers are |
| 1985 | // already assigned, so visiting it would evict what still lives in the |
| 1986 | // source. |
| 1987 | for (MachineInstr &MI : make_early_inc_range(Range: reverse(C&: MBB))) { |
| 1988 | LLVM_DEBUG(dbgs() << "\n>> " << MI << "Regs:" ; dumpState()); |
| 1989 | |
| 1990 | // Special handling for debug values. Note that they are not allowed to |
| 1991 | // affect codegen of the other instructions in any way. |
| 1992 | if (MI.isDebugValue()) { |
| 1993 | handleDebugValue(MI); |
| 1994 | continue; |
| 1995 | } |
| 1996 | |
| 1997 | allocateInstruction(MI); |
| 1998 | |
| 1999 | // Once BUNDLE header is assigned registers, same assignments need to be |
| 2000 | // done for bundled MIs. |
| 2001 | if (MI.getOpcode() == TargetOpcode::BUNDLE) { |
| 2002 | handleBundle(MI); |
| 2003 | } |
| 2004 | } |
| 2005 | |
| 2006 | LLVM_DEBUG(dbgs() << "Begin Regs:" ; dumpState()); |
| 2007 | |
| 2008 | // Spill all physical registers holding virtual registers now. |
| 2009 | LLVM_DEBUG(dbgs() << "Loading live registers at begin of block.\n" ); |
| 2010 | reloadAtBegin(MBB); |
| 2011 | |
| 2012 | // Erase all the coalesced copies. We are delaying it until now because |
| 2013 | // LiveVirtRegs might refer to the instrs. |
| 2014 | for (MachineInstr *MI : Coalesced) |
| 2015 | MBB.erase(I: MI); |
| 2016 | NumCoalesced += Coalesced.size(); |
| 2017 | |
| 2018 | for (auto &UDBGPair : DanglingDbgValues) { |
| 2019 | for (MachineInstr *DbgValue : UDBGPair.second) { |
| 2020 | assert(DbgValue->isDebugValue() && "expected DBG_VALUE" ); |
| 2021 | // Nothing to do if the vreg was spilled in the meantime. |
| 2022 | if (!DbgValue->hasDebugOperandForReg(Reg: UDBGPair.first)) |
| 2023 | continue; |
| 2024 | LLVM_DEBUG(dbgs() << "Register did not survive for " << *DbgValue |
| 2025 | << '\n'); |
| 2026 | DbgValue->setDebugValueUndef(); |
| 2027 | } |
| 2028 | } |
| 2029 | DanglingDbgValues.clear(); |
| 2030 | |
| 2031 | LLVM_DEBUG(MBB.dump()); |
| 2032 | } |
| 2033 | |
| 2034 | /// Expand REG_SEQUENCE and INSERT_SUBREG into subregister COPYs: the lowering |
| 2035 | /// TwoAddressInstructionPass performs when it runs before allocation, plus the |
| 2036 | /// base-value copy its tie processing provides. |
| 2037 | void RegAllocFastImpl::expandSubregPseudo(MachineInstr &MI) { |
| 2038 | MachineBasicBlock &MBB = *MI.getParent(); |
| 2039 | const DebugLoc &DL = MI.getDebugLoc(); |
| 2040 | if (MI.isInsertSubreg()) { |
| 2041 | // %d = INSERT_SUBREG %base, %sub, idx -> %d = COPY %base |
| 2042 | // %d.idx = COPY %sub |
| 2043 | const MachineOperand &BaseMO = MI.getOperand(i: 1); |
| 2044 | if (!BaseMO.isUndef()) |
| 2045 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), |
| 2046 | DestReg: MI.getOperand(i: 0).getReg()) |
| 2047 | .addReg(RegNo: BaseMO.getReg(), Flags: RegState::NoFlags, SubReg: BaseMO.getSubReg()); |
| 2048 | unsigned SubIdx = MI.getOperand(i: 3).getImm(); |
| 2049 | MI.removeOperand(OpNo: 3); |
| 2050 | assert(MI.getOperand(0).getSubReg() == 0 && "Unexpected subreg idx" ); |
| 2051 | MI.getOperand(i: 0).setSubReg(SubIdx); |
| 2052 | MI.getOperand(i: 0).setIsUndef(MI.getOperand(i: 1).isUndef()); |
| 2053 | MI.removeOperand(OpNo: 1); |
| 2054 | MI.setDesc(TII->get(Opcode: TargetOpcode::COPY)); |
| 2055 | return; |
| 2056 | } |
| 2057 | |
| 2058 | // %d = REG_SEQUENCE %s1, idx1, ... -> undef %d.idx1 = COPY %s1 |
| 2059 | // %d.idx2 = COPY %s2 ... |
| 2060 | assert(MI.isRegSequence()); |
| 2061 | Register Dst = MI.getOperand(i: 0).getReg(); |
| 2062 | // An undef source needs no copy: the read-undef flag on the first copy |
| 2063 | // defines the whole register. One is still needed where a use reads that |
| 2064 | // lane on its own, which would otherwise read an undefined subregister. |
| 2065 | LaneBitmask ReadLanes = LaneBitmask::getNone(); |
| 2066 | for (const MachineOperand &Use : MRI->use_nodbg_operands(Reg: Dst)) |
| 2067 | if (unsigned UseSubIdx = Use.getSubReg()) |
| 2068 | ReadLanes |= TRI->getSubRegIndexLaneMask(SubIdx: UseSubIdx); |
| 2069 | |
| 2070 | bool DefEmitted = false; |
| 2071 | for (unsigned I = 1, E = MI.getNumOperands(); I + 1 < E; I += 2) { |
| 2072 | const MachineOperand &SrcMO = MI.getOperand(i: I); |
| 2073 | unsigned SubIdx = MI.getOperand(i: I + 1).getImm(); |
| 2074 | if (SrcMO.isUndef() && |
| 2075 | (ReadLanes & TRI->getSubRegIndexLaneMask(SubIdx)).none()) |
| 2076 | continue; |
| 2077 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY)) |
| 2078 | .addReg(RegNo: Dst, Flags: RegState::Define | getUndefRegState(B: !DefEmitted), SubReg: SubIdx) |
| 2079 | .addReg(RegNo: SrcMO.getReg(), Flags: getUndefRegState(B: SrcMO.isUndef()), |
| 2080 | SubReg: SrcMO.getSubReg()); |
| 2081 | DefEmitted = true; |
| 2082 | } |
| 2083 | // Every source was undef: uses of Dst still need a definition. |
| 2084 | if (!DefEmitted) { |
| 2085 | MI.setDesc(TII->get(Opcode: TargetOpcode::IMPLICIT_DEF)); |
| 2086 | while (MI.getNumOperands() > 1) |
| 2087 | MI.removeOperand(OpNo: MI.getNumOperands() - 1); |
| 2088 | return; |
| 2089 | } |
| 2090 | MI.eraseFromParent(); |
| 2091 | } |
| 2092 | |
| 2093 | bool RegAllocFastImpl::runOnMachineFunction(MachineFunction &MF) { |
| 2094 | LLVM_DEBUG(dbgs() << "********** FAST REGISTER ALLOCATION **********\n" |
| 2095 | << "********** Function: " << MF.getName() << '\n'); |
| 2096 | MRI = &MF.getRegInfo(); |
| 2097 | const TargetSubtargetInfo &STI = MF.getSubtarget(); |
| 2098 | TRI = STI.getRegisterInfo(); |
| 2099 | TII = STI.getInstrInfo(); |
| 2100 | MFI = &MF.getFrameInfo(); |
| 2101 | MRI->freezeReservedRegs(); |
| 2102 | RegClassInfo.runOnMachineFunction(MF); |
| 2103 | unsigned NumRegUnits = TRI->getNumRegUnits(); |
| 2104 | InstrGen = 0; |
| 2105 | UsedInInstr.assign(NumElts: NumRegUnits, Elt: 0); |
| 2106 | LiveDefUnits.assign(NumElts: NumRegUnits, Elt: 0); |
| 2107 | |
| 2108 | // MIR that already went through TwoAddressInstructionPass carries |
| 2109 | // TiedOpsRewritten, so partial pipelines (-run-pass, -start-before) follow |
| 2110 | // the input they are given. |
| 2111 | LowerTiedOps = !MF.getProperties().hasTiedOpsRewritten(); |
| 2112 | if (LowerTiedOps) { |
| 2113 | for (MachineBasicBlock &MBB : MF) |
| 2114 | for (MachineInstr &MI : make_early_inc_range(Range&: MBB)) |
| 2115 | if (MI.isRegSequence() || MI.isInsertSubreg()) |
| 2116 | expandSubregPseudo(MI); |
| 2117 | } |
| 2118 | |
| 2119 | // initialize the virtual->physical register map to have a 'null' |
| 2120 | // mapping for all virtual registers |
| 2121 | unsigned NumVirtRegs = MRI->getNumVirtRegs(); |
| 2122 | StackSlotForVirtReg.resize(S: NumVirtRegs); |
| 2123 | LiveVirtRegs.setUniverse(NumVirtRegs); |
| 2124 | MayLiveAcrossBlocks.clear(); |
| 2125 | MayLiveAcrossBlocks.resize(N: NumVirtRegs); |
| 2126 | |
| 2127 | // Loop over all of the basic blocks, eliminating virtual register references |
| 2128 | for (MachineBasicBlock &MBB : MF) |
| 2129 | allocateBasicBlock(MBB); |
| 2130 | |
| 2131 | if (ClearVirtRegs) { |
| 2132 | // All machine operands and other references to virtual registers have been |
| 2133 | // replaced. Remove the virtual registers. |
| 2134 | MRI->clearVirtRegs(); |
| 2135 | } |
| 2136 | |
| 2137 | StackSlotForVirtReg.clear(); |
| 2138 | LiveDbgValueMap.clear(); |
| 2139 | return true; |
| 2140 | } |
| 2141 | |
| 2142 | PreservedAnalyses RegAllocFastPass::run(MachineFunction &MF, |
| 2143 | MachineFunctionAnalysisManager &) { |
| 2144 | MFPropsModifier _(*this, MF); |
| 2145 | RegAllocFastImpl Impl(Opts.Filter, Opts.ClearVRegs); |
| 2146 | bool Changed = Impl.runOnMachineFunction(MF); |
| 2147 | if (!Changed) |
| 2148 | return PreservedAnalyses::all(); |
| 2149 | auto PA = getMachineFunctionPassPreservedAnalyses(); |
| 2150 | PA.preserveSet<CFGAnalyses>(); |
| 2151 | return PA; |
| 2152 | } |
| 2153 | |
| 2154 | void RegAllocFastPass::printPipeline( |
| 2155 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
| 2156 | bool PrintFilterName = Opts.FilterName != "all" ; |
| 2157 | bool PrintNoClearVRegs = !Opts.ClearVRegs; |
| 2158 | bool PrintSemicolon = PrintFilterName && PrintNoClearVRegs; |
| 2159 | |
| 2160 | OS << "regallocfast" ; |
| 2161 | if (PrintFilterName || PrintNoClearVRegs) { |
| 2162 | OS << '<'; |
| 2163 | if (PrintFilterName) |
| 2164 | OS << "filter=" << Opts.FilterName; |
| 2165 | if (PrintSemicolon) |
| 2166 | OS << ';'; |
| 2167 | if (PrintNoClearVRegs) |
| 2168 | OS << "no-clear-vregs" ; |
| 2169 | OS << '>'; |
| 2170 | } |
| 2171 | } |
| 2172 | |
| 2173 | FunctionPass *llvm::createFastRegisterAllocator() { return new RegAllocFast(); } |
| 2174 | |
| 2175 | FunctionPass *llvm::createFastRegisterAllocator(RegAllocFilterFunc Ftor, |
| 2176 | bool ClearVirtRegs) { |
| 2177 | return new RegAllocFast(Ftor, ClearVirtRegs); |
| 2178 | } |
| 2179 | |