| 1 | //===- MachineSMEABIPass.cpp ----------------------------------------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This pass implements the SME ABI requirements for ZA state. This includes |
| 10 | // implementing the lazy (and agnostic) ZA state save schemes around calls. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | // |
| 14 | // This pass works by collecting instructions that require ZA to be in a |
| 15 | // specific state (e.g., "ACTIVE" or "SAVED") and inserting the necessary state |
| 16 | // transitions to ensure ZA is in the required state before instructions. State |
| 17 | // transitions represent actions such as setting up or restoring a lazy save. |
| 18 | // Certain points within a function may also have predefined states independent |
| 19 | // of any instructions, for example, a "shared_za" function is always entered |
| 20 | // and exited in the "ACTIVE" state. |
| 21 | // |
| 22 | // To handle ZA state across control flow, we make use of edge bundling. This |
| 23 | // assigns each block an "incoming" and "outgoing" edge bundle (representing |
| 24 | // incoming and outgoing edges). Initially, these are unique to each block; |
| 25 | // then, in the process of forming bundles, the outgoing bundle of a block is |
| 26 | // joined with the incoming bundle of all successors. The result is that each |
| 27 | // bundle can be assigned a single ZA state, which ensures the state required by |
| 28 | // all a blocks' successors is the same, and that each basic block will always |
| 29 | // be entered with the same ZA state. This eliminates the need for splitting |
| 30 | // edges to insert state transitions or "phi" nodes for ZA states. |
| 31 | // |
| 32 | // See below for a simple example of edge bundling. |
| 33 | // |
| 34 | // The following shows a conditionally executed basic block (BB1): |
| 35 | // |
| 36 | // if (cond) |
| 37 | // BB1 |
| 38 | // BB2 |
| 39 | // |
| 40 | // Initial Bundles Joined Bundles |
| 41 | // |
| 42 | // ┌──0──┐ ┌──0──┐ |
| 43 | // │ BB0 │ │ BB0 │ |
| 44 | // └──1──┘ └──1──┘ |
| 45 | // ├───────┐ ├───────┐ |
| 46 | // ▼ │ ▼ │ |
| 47 | // ┌──2──┐ │ ─────► ┌──1──┐ │ |
| 48 | // │ BB1 │ ▼ │ BB1 │ ▼ |
| 49 | // └──3──┘ ┌──4──┐ └──1──┘ ┌──1──┐ |
| 50 | // └───►4 BB2 │ └───►1 BB2 │ |
| 51 | // └──5──┘ └──2──┘ |
| 52 | // |
| 53 | // On the left are the initial per-block bundles, and on the right are the |
| 54 | // joined bundles (which are the result of the EdgeBundles analysis). |
| 55 | |
| 56 | #include "AArch64InstrInfo.h" |
| 57 | #include "AArch64MachineFunctionInfo.h" |
| 58 | #include "AArch64Subtarget.h" |
| 59 | #include "MCTargetDesc/AArch64AddressingModes.h" |
| 60 | #include "llvm/ADT/BitmaskEnum.h" |
| 61 | #include "llvm/ADT/SmallVector.h" |
| 62 | #include "llvm/CodeGen/EdgeBundles.h" |
| 63 | #include "llvm/CodeGen/LivePhysRegs.h" |
| 64 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 65 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 66 | #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h" |
| 67 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 68 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 69 | |
| 70 | using namespace llvm; |
| 71 | |
| 72 | #define DEBUG_TYPE "aarch64-machine-sme-abi" |
| 73 | |
| 74 | namespace { |
| 75 | |
| 76 | // Note: For agnostic ZA, we assume the function is always entered/exited in the |
| 77 | // "ACTIVE" state -- this _may_ not be the case (since OFF is also a |
| 78 | // possibility, but for the purpose of placing ZA saves/restores, that does not |
| 79 | // matter). |
| 80 | enum ZAState : uint8_t { |
| 81 | // Any/unknown state (not valid) |
| 82 | ANY = 0, |
| 83 | |
| 84 | // ZA is in use and active (i.e. within the accumulator) |
| 85 | ACTIVE, |
| 86 | |
| 87 | // ZA is active, but ZT0 has been saved. |
| 88 | // This handles the edge case of sharedZA && !sharesZT0. |
| 89 | ACTIVE_ZT0_SAVED, |
| 90 | |
| 91 | // A ZA save has been set up or committed (i.e. ZA is dormant or off) |
| 92 | // If the function uses ZT0 it must also be saved. |
| 93 | LOCAL_SAVED, |
| 94 | |
| 95 | // ZA has been committed to the lazy save buffer of the current function. |
| 96 | // If the function uses ZT0 it must also be saved. |
| 97 | // ZA is off. |
| 98 | LOCAL_COMMITTED, |
| 99 | |
| 100 | // The ZA/ZT0 state on entry to the function. |
| 101 | ENTRY, |
| 102 | |
| 103 | // ZA is off. |
| 104 | OFF, |
| 105 | |
| 106 | // The number of ZA states (not a valid state) |
| 107 | NUM_ZA_STATE |
| 108 | }; |
| 109 | |
| 110 | /// A bitmask enum to record live physical registers that the "emit*" routines |
| 111 | /// may need to preserve. Note: This only tracks registers we may clobber. |
| 112 | enum LiveRegs : uint8_t { |
| 113 | None = 0, |
| 114 | NZCV = 1 << 0, |
| 115 | W0 = 1 << 1, |
| 116 | W0_HI = 1 << 2, |
| 117 | X0 = W0 | W0_HI, |
| 118 | LLVM_MARK_AS_BITMASK_ENUM(/* LargestValue = */ W0_HI) |
| 119 | }; |
| 120 | |
| 121 | /// Holds the virtual registers live physical registers have been saved to. |
| 122 | struct PhysRegSave { |
| 123 | LiveRegs PhysLiveRegs; |
| 124 | Register StatusFlags = AArch64::NoRegister; |
| 125 | Register X0Save = AArch64::NoRegister; |
| 126 | }; |
| 127 | |
| 128 | /// Contains the needed ZA state (and live registers) at an instruction. That is |
| 129 | /// the state ZA must be in _before_ "InsertPt". |
| 130 | struct InstInfo { |
| 131 | ZAState NeededState{ZAState::ANY}; |
| 132 | MachineBasicBlock::iterator InsertPt; |
| 133 | LiveRegs PhysLiveRegs = LiveRegs::None; |
| 134 | }; |
| 135 | |
| 136 | /// Contains the needed ZA state for each instruction in a block. Instructions |
| 137 | /// that do not require a ZA state are not recorded. |
| 138 | struct BlockInfo { |
| 139 | SmallVector<InstInfo> Insts; |
| 140 | ZAState FixedEntryState{ZAState::ANY}; |
| 141 | ZAState DesiredIncomingState{ZAState::ANY}; |
| 142 | ZAState DesiredOutgoingState{ZAState::ANY}; |
| 143 | LiveRegs PhysLiveRegsAtEntry = LiveRegs::None; |
| 144 | LiveRegs PhysLiveRegsAtExit = LiveRegs::None; |
| 145 | }; |
| 146 | |
| 147 | /// Contains the needed ZA state information for all blocks within a function. |
| 148 | struct FunctionInfo { |
| 149 | SmallVector<BlockInfo> Blocks; |
| 150 | std::optional<MachineBasicBlock::iterator> AfterSMEProloguePt; |
| 151 | LiveRegs PhysLiveRegsAfterSMEPrologue = LiveRegs::None; |
| 152 | }; |
| 153 | |
| 154 | /// State/helpers that is only needed when emitting code to handle |
| 155 | /// saving/restoring ZA. |
| 156 | class EmitContext { |
| 157 | public: |
| 158 | EmitContext() = default; |
| 159 | |
| 160 | /// Get or create a TPIDR2 block in \p MF. |
| 161 | int getTPIDR2Block(MachineFunction &MF) { |
| 162 | if (TPIDR2BlockFI) |
| 163 | return *TPIDR2BlockFI; |
| 164 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 165 | TPIDR2BlockFI = MFI.CreateStackObject(Size: 16, Alignment: Align(16), isSpillSlot: false); |
| 166 | return *TPIDR2BlockFI; |
| 167 | } |
| 168 | |
| 169 | /// Get or create agnostic ZA buffer pointer in \p MF. |
| 170 | Register getAgnosticZABufferPtr(MachineFunction &MF) { |
| 171 | if (AgnosticZABufferPtr != AArch64::NoRegister) |
| 172 | return AgnosticZABufferPtr; |
| 173 | Register BufferPtr = |
| 174 | MF.getInfo<AArch64FunctionInfo>()->getEarlyAllocSMESaveBuffer(); |
| 175 | AgnosticZABufferPtr = |
| 176 | BufferPtr != AArch64::NoRegister |
| 177 | ? BufferPtr |
| 178 | : MF.getRegInfo().createVirtualRegister(RegClass: &AArch64::GPR64RegClass); |
| 179 | return AgnosticZABufferPtr; |
| 180 | } |
| 181 | |
| 182 | int getZT0SaveSlot(MachineFunction &MF) { |
| 183 | if (ZT0SaveFI) |
| 184 | return *ZT0SaveFI; |
| 185 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 186 | ZT0SaveFI = MFI.CreateSpillStackObject(Size: 64, Alignment: Align(16)); |
| 187 | return *ZT0SaveFI; |
| 188 | } |
| 189 | |
| 190 | /// Returns true if the function must allocate a ZA save buffer on entry. This |
| 191 | /// will be the case if, at any point in the function, a ZA save was emitted. |
| 192 | bool needsSaveBuffer() const { |
| 193 | assert(!(TPIDR2BlockFI && AgnosticZABufferPtr) && |
| 194 | "Cannot have both a TPIDR2 block and agnostic ZA buffer" ); |
| 195 | return TPIDR2BlockFI || AgnosticZABufferPtr != AArch64::NoRegister; |
| 196 | } |
| 197 | |
| 198 | private: |
| 199 | std::optional<int> ZT0SaveFI; |
| 200 | std::optional<int> TPIDR2BlockFI; |
| 201 | Register AgnosticZABufferPtr = AArch64::NoRegister; |
| 202 | }; |
| 203 | |
| 204 | StringRef getZAStateString(ZAState State) { |
| 205 | #define MAKE_CASE(V) \ |
| 206 | case V: \ |
| 207 | return #V; |
| 208 | switch (State) { |
| 209 | MAKE_CASE(ZAState::ANY) |
| 210 | MAKE_CASE(ZAState::ACTIVE) |
| 211 | MAKE_CASE(ZAState::ACTIVE_ZT0_SAVED) |
| 212 | MAKE_CASE(ZAState::LOCAL_SAVED) |
| 213 | MAKE_CASE(ZAState::LOCAL_COMMITTED) |
| 214 | MAKE_CASE(ZAState::ENTRY) |
| 215 | MAKE_CASE(ZAState::OFF) |
| 216 | default: |
| 217 | llvm_unreachable("Unexpected ZAState" ); |
| 218 | } |
| 219 | #undef MAKE_CASE |
| 220 | } |
| 221 | |
| 222 | static bool isZAorZTRegOp(const TargetRegisterInfo &TRI, |
| 223 | const MachineOperand &MO) { |
| 224 | if (!MO.isReg() || !MO.getReg().isPhysical()) |
| 225 | return false; |
| 226 | return any_of(Range: TRI.subregs_inclusive(Reg: MO.getReg()), P: [](const MCPhysReg &SR) { |
| 227 | return AArch64::MPR128RegClass.contains(Reg: SR) || |
| 228 | AArch64::ZTRRegClass.contains(Reg: SR); |
| 229 | }); |
| 230 | } |
| 231 | |
| 232 | /// Returns the required ZA state needed before \p MI and an iterator pointing |
| 233 | /// to where any code required to change the ZA state should be inserted. |
| 234 | static std::pair<ZAState, MachineBasicBlock::iterator> |
| 235 | getInstNeededZAState(const TargetRegisterInfo &TRI, MachineInstr &MI, |
| 236 | SMEAttrs SMEFnAttrs) { |
| 237 | MachineBasicBlock::iterator InsertPt(MI); |
| 238 | |
| 239 | // Note: InOutZAUsePseudo, RequiresZASavePseudo, and RequiresZT0SavePseudo are |
| 240 | // intended to mark the position immediately before a call. Due to |
| 241 | // SelectionDAG constraints, these markers occur after the ADJCALLSTACKDOWN, |
| 242 | // so we use std::prev(InsertPt) to get the position before the call. |
| 243 | |
| 244 | if (MI.getOpcode() == AArch64::InOutZAUsePseudo) |
| 245 | return {ZAState::ACTIVE, std::prev(x: InsertPt)}; |
| 246 | |
| 247 | // Note: If we need to save both ZA and ZT0 we use RequiresZASavePseudo. |
| 248 | if (MI.getOpcode() == AArch64::RequiresZASavePseudo) |
| 249 | return {ZAState::LOCAL_SAVED, std::prev(x: InsertPt)}; |
| 250 | |
| 251 | // If we only need to save ZT0 there's two cases to consider: |
| 252 | // 1. The function has ZA state (that we don't need to save). |
| 253 | // - In this case we switch to the "ACTIVE_ZT0_SAVED" state. |
| 254 | // This only saves ZT0. |
| 255 | // 2. The function does not have ZA state |
| 256 | // - In this case we switch to "LOCAL_COMMITTED" state. |
| 257 | // This saves ZT0 and turns ZA off. |
| 258 | if (MI.getOpcode() == AArch64::RequiresZT0SavePseudo) { |
| 259 | return {SMEFnAttrs.hasZAState() ? ZAState::ACTIVE_ZT0_SAVED |
| 260 | : ZAState::LOCAL_COMMITTED, |
| 261 | std::prev(x: InsertPt)}; |
| 262 | } |
| 263 | |
| 264 | if (MI.isReturn()) { |
| 265 | bool ZAOffAtReturn = SMEFnAttrs.hasPrivateZAInterface(); |
| 266 | return {ZAOffAtReturn ? ZAState::OFF : ZAState::ACTIVE, InsertPt}; |
| 267 | } |
| 268 | |
| 269 | for (auto &MO : MI.operands()) { |
| 270 | if (isZAorZTRegOp(TRI, MO)) |
| 271 | return {ZAState::ACTIVE, InsertPt}; |
| 272 | } |
| 273 | |
| 274 | return {ZAState::ANY, InsertPt}; |
| 275 | } |
| 276 | |
| 277 | struct MachineSMEABI : public MachineFunctionPass { |
| 278 | inline static char ID = 0; |
| 279 | |
| 280 | MachineSMEABI(CodeGenOptLevel OptLevel = CodeGenOptLevel::Default) |
| 281 | : MachineFunctionPass(ID), OptLevel(OptLevel) {} |
| 282 | |
| 283 | bool runOnMachineFunction(MachineFunction &MF) override; |
| 284 | |
| 285 | StringRef getPassName() const override { return "Machine SME ABI pass" ; } |
| 286 | |
| 287 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 288 | AU.setPreservesCFG(); |
| 289 | AU.addRequired<EdgeBundlesWrapperLegacy>(); |
| 290 | AU.addRequired<MachineOptimizationRemarkEmitterPass>(); |
| 291 | AU.addRequired<LibcallLoweringInfoWrapper>(); |
| 292 | AU.addPreservedID(ID&: MachineLoopInfoID); |
| 293 | AU.addPreservedID(ID&: MachineDominatorsID); |
| 294 | MachineFunctionPass::getAnalysisUsage(AU); |
| 295 | } |
| 296 | |
| 297 | /// Collects the needed ZA state (and live registers) before each instruction |
| 298 | /// within the machine function. |
| 299 | FunctionInfo collectNeededZAStates(SMEAttrs SMEFnAttrs); |
| 300 | |
| 301 | /// Assigns each edge bundle a ZA state based on the needed states of blocks |
| 302 | /// that have incoming or outgoing edges in that bundle. |
| 303 | SmallVector<ZAState> assignBundleZAStates(const EdgeBundles &Bundles, |
| 304 | const FunctionInfo &FnInfo); |
| 305 | |
| 306 | /// Inserts code to handle changes between ZA states within the function. |
| 307 | /// E.g., ACTIVE -> LOCAL_SAVED will insert code required to save ZA. |
| 308 | void insertStateChanges(EmitContext &, const FunctionInfo &FnInfo, |
| 309 | const EdgeBundles &Bundles, |
| 310 | ArrayRef<ZAState> BundleStates); |
| 311 | |
| 312 | void addSMELibCall(MachineInstrBuilder &MIB, RTLIB::Libcall LC, |
| 313 | CallingConv::ID ExpectedCC); |
| 314 | |
| 315 | void emitZT0SaveRestore(EmitContext &, MachineBasicBlock &MBB, |
| 316 | MachineBasicBlock::iterator MBBI, bool IsSave); |
| 317 | |
| 318 | // Emission routines for private and shared ZA functions (using lazy saves). |
| 319 | void emitSMEPrologue(MachineBasicBlock &MBB, |
| 320 | MachineBasicBlock::iterator MBBI); |
| 321 | void emitRestoreLazySave(EmitContext &, MachineBasicBlock &MBB, |
| 322 | MachineBasicBlock::iterator MBBI, |
| 323 | LiveRegs PhysLiveRegs); |
| 324 | void emitSetupLazySave(EmitContext &, MachineBasicBlock &MBB, |
| 325 | MachineBasicBlock::iterator MBBI); |
| 326 | void emitAllocateLazySaveBuffer(EmitContext &, MachineBasicBlock &MBB, |
| 327 | MachineBasicBlock::iterator MBBI); |
| 328 | void emitZAMode(MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, |
| 329 | bool ClearTPIDR2, bool On); |
| 330 | |
| 331 | // Emission routines for agnostic ZA functions. |
| 332 | void emitSetupFullZASave(MachineBasicBlock &MBB, |
| 333 | MachineBasicBlock::iterator MBBI, |
| 334 | LiveRegs PhysLiveRegs); |
| 335 | // Emit a "full" ZA save or restore. It is "full" in the sense that this |
| 336 | // function will emit a call to __arm_sme_save or __arm_sme_restore, which |
| 337 | // handles saving and restoring both ZA and ZT0. |
| 338 | void emitFullZASaveRestore(EmitContext &, MachineBasicBlock &MBB, |
| 339 | MachineBasicBlock::iterator MBBI, |
| 340 | LiveRegs PhysLiveRegs, bool IsSave); |
| 341 | void emitAllocateFullZASaveBuffer(EmitContext &, MachineBasicBlock &MBB, |
| 342 | MachineBasicBlock::iterator MBBI, |
| 343 | LiveRegs PhysLiveRegs); |
| 344 | |
| 345 | /// Attempts to find an insertion point before \p Inst where the status flags |
| 346 | /// are not live. If \p Inst is `Block.Insts.end()` a point before the end of |
| 347 | /// the block is found. |
| 348 | std::pair<MachineBasicBlock::iterator, LiveRegs> |
| 349 | findStateChangeInsertionPoint(MachineBasicBlock &MBB, const BlockInfo &Block, |
| 350 | SmallVectorImpl<InstInfo>::const_iterator Inst); |
| 351 | void emitStateChange(EmitContext &, MachineBasicBlock &MBB, |
| 352 | MachineBasicBlock::iterator MBBI, ZAState From, |
| 353 | ZAState To, LiveRegs PhysLiveRegs); |
| 354 | |
| 355 | // Helpers for switching between lazy/full ZA save/restore routines. |
| 356 | void emitZASave(EmitContext &Context, MachineBasicBlock &MBB, |
| 357 | MachineBasicBlock::iterator MBBI, LiveRegs PhysLiveRegs) { |
| 358 | if (AFI->getSMEFnAttrs().hasAgnosticZAInterface()) |
| 359 | return emitFullZASaveRestore(Context, MBB, MBBI, PhysLiveRegs, |
| 360 | /*IsSave=*/true); |
| 361 | return emitSetupLazySave(Context, MBB, MBBI); |
| 362 | } |
| 363 | void emitZARestore(EmitContext &Context, MachineBasicBlock &MBB, |
| 364 | MachineBasicBlock::iterator MBBI, LiveRegs PhysLiveRegs) { |
| 365 | if (AFI->getSMEFnAttrs().hasAgnosticZAInterface()) |
| 366 | return emitFullZASaveRestore(Context, MBB, MBBI, PhysLiveRegs, |
| 367 | /*IsSave=*/false); |
| 368 | return emitRestoreLazySave(Context, MBB, MBBI, PhysLiveRegs); |
| 369 | } |
| 370 | void emitAllocateZASaveBuffer(EmitContext &Context, MachineBasicBlock &MBB, |
| 371 | MachineBasicBlock::iterator MBBI, |
| 372 | LiveRegs PhysLiveRegs) { |
| 373 | if (AFI->getSMEFnAttrs().hasAgnosticZAInterface()) |
| 374 | return emitAllocateFullZASaveBuffer(Context, MBB, MBBI, PhysLiveRegs); |
| 375 | return emitAllocateLazySaveBuffer(Context, MBB, MBBI); |
| 376 | } |
| 377 | |
| 378 | /// Collects the reachable calls from \p MBBI marked with \p Marker. This is |
| 379 | /// intended to be used to emit lazy save remarks. Note: This stops at the |
| 380 | /// first marked call along any path. |
| 381 | void collectReachableMarkedCalls(const MachineBasicBlock &MBB, |
| 382 | MachineBasicBlock::const_iterator MBBI, |
| 383 | SmallVectorImpl<const MachineInstr *> &Calls, |
| 384 | unsigned Marker) const; |
| 385 | |
| 386 | void emitCallSaveRemarks(const MachineBasicBlock &MBB, |
| 387 | MachineBasicBlock::const_iterator MBBI, DebugLoc DL, |
| 388 | unsigned Marker, StringRef , |
| 389 | StringRef SaveName) const; |
| 390 | |
| 391 | void emitError(const Twine &Message) { |
| 392 | LLVMContext &Context = MF->getFunction().getContext(); |
| 393 | Context.emitError(ErrorStr: MF->getName() + ": " + Message); |
| 394 | } |
| 395 | |
| 396 | /// Save live physical registers to virtual registers. |
| 397 | PhysRegSave createPhysRegSave(LiveRegs PhysLiveRegs, MachineBasicBlock &MBB, |
| 398 | MachineBasicBlock::iterator MBBI, DebugLoc DL); |
| 399 | /// Restore physical registers from a save of their previous values. |
| 400 | void restorePhyRegSave(const PhysRegSave &RegSave, MachineBasicBlock &MBB, |
| 401 | MachineBasicBlock::iterator MBBI, DebugLoc DL); |
| 402 | |
| 403 | private: |
| 404 | CodeGenOptLevel OptLevel = CodeGenOptLevel::Default; |
| 405 | |
| 406 | MachineFunction *MF = nullptr; |
| 407 | const AArch64Subtarget *Subtarget = nullptr; |
| 408 | const AArch64RegisterInfo *TRI = nullptr; |
| 409 | const AArch64FunctionInfo *AFI = nullptr; |
| 410 | const AArch64InstrInfo *TII = nullptr; |
| 411 | const LibcallLoweringInfo *LLI = nullptr; |
| 412 | |
| 413 | MachineOptimizationRemarkEmitter *ORE = nullptr; |
| 414 | MachineRegisterInfo *MRI = nullptr; |
| 415 | MachineLoopInfo *MLI = nullptr; |
| 416 | }; |
| 417 | |
| 418 | static LiveRegs getPhysLiveRegs(LiveRegUnits const &LiveUnits) { |
| 419 | LiveRegs PhysLiveRegs = LiveRegs::None; |
| 420 | if (!LiveUnits.available(Reg: AArch64::NZCV)) |
| 421 | PhysLiveRegs |= LiveRegs::NZCV; |
| 422 | // We have to track W0 and X0 separately as otherwise things can get |
| 423 | // confused if we attempt to preserve X0 but only W0 was defined. |
| 424 | if (!LiveUnits.available(Reg: AArch64::W0)) |
| 425 | PhysLiveRegs |= LiveRegs::W0; |
| 426 | if (!LiveUnits.available(Reg: AArch64::W0_HI)) |
| 427 | PhysLiveRegs |= LiveRegs::W0_HI; |
| 428 | return PhysLiveRegs; |
| 429 | } |
| 430 | |
| 431 | static void setPhysLiveRegs(LiveRegUnits &LiveUnits, LiveRegs PhysLiveRegs) { |
| 432 | if (PhysLiveRegs & LiveRegs::NZCV) |
| 433 | LiveUnits.addReg(Reg: AArch64::NZCV); |
| 434 | if (PhysLiveRegs & LiveRegs::W0) |
| 435 | LiveUnits.addReg(Reg: AArch64::W0); |
| 436 | if (PhysLiveRegs & LiveRegs::W0_HI) |
| 437 | LiveUnits.addReg(Reg: AArch64::W0_HI); |
| 438 | } |
| 439 | |
| 440 | [[maybe_unused]] bool isCallStartOpcode(unsigned Opc) { |
| 441 | switch (Opc) { |
| 442 | case AArch64::TLSDESC_CALLSEQ: |
| 443 | case AArch64::TLSDESC_AUTH_CALLSEQ: |
| 444 | case AArch64::ADJCALLSTACKDOWN: |
| 445 | return true; |
| 446 | default: |
| 447 | return false; |
| 448 | } |
| 449 | } |
| 450 | |
| 451 | FunctionInfo MachineSMEABI::collectNeededZAStates(SMEAttrs SMEFnAttrs) { |
| 452 | assert((SMEFnAttrs.hasAgnosticZAInterface() || SMEFnAttrs.hasZT0State() || |
| 453 | SMEFnAttrs.hasZAState()) && |
| 454 | "Expected function to have ZA/ZT0 state!" ); |
| 455 | |
| 456 | SmallVector<BlockInfo> Blocks(MF->getNumBlockIDs()); |
| 457 | LiveRegs PhysLiveRegsAfterSMEPrologue = LiveRegs::None; |
| 458 | std::optional<MachineBasicBlock::iterator> AfterSMEProloguePt; |
| 459 | |
| 460 | for (MachineBasicBlock &MBB : *MF) { |
| 461 | BlockInfo &Block = Blocks[MBB.getNumber()]; |
| 462 | |
| 463 | if (MBB.isEntryBlock()) { |
| 464 | // Entry block: |
| 465 | Block.FixedEntryState = ZAState::ENTRY; |
| 466 | } else if (MBB.isEHPad()) { |
| 467 | // EH entry block: |
| 468 | Block.FixedEntryState = ZAState::LOCAL_COMMITTED; |
| 469 | } |
| 470 | |
| 471 | LiveRegUnits LiveUnits(*TRI); |
| 472 | LiveUnits.addLiveOuts(MBB); |
| 473 | |
| 474 | Block.PhysLiveRegsAtExit = getPhysLiveRegs(LiveUnits); |
| 475 | auto FirstTerminatorInsertPt = MBB.getFirstTerminator(); |
| 476 | auto FirstNonPhiInsertPt = MBB.getFirstNonPHI(); |
| 477 | for (MachineInstr &MI : reverse(C&: MBB)) { |
| 478 | MachineBasicBlock::iterator MBBI(MI); |
| 479 | LiveUnits.stepBackward(MI); |
| 480 | LiveRegs PhysLiveRegs = getPhysLiveRegs(LiveUnits); |
| 481 | // The SMEStateAllocPseudo marker is added to a function if the save |
| 482 | // buffer was allocated in SelectionDAG. It marks the end of the |
| 483 | // allocation -- which is a safe point for this pass to insert any TPIDR2 |
| 484 | // block setup. |
| 485 | if (MI.getOpcode() == AArch64::SMEStateAllocPseudo) { |
| 486 | AfterSMEProloguePt = MBBI; |
| 487 | PhysLiveRegsAfterSMEPrologue = PhysLiveRegs; |
| 488 | } |
| 489 | // Note: We treat Agnostic ZA as inout_za with an alternate save/restore. |
| 490 | auto [NeededState, InsertPt] = getInstNeededZAState(TRI: *TRI, MI, SMEFnAttrs); |
| 491 | assert((InsertPt == MBBI || isCallStartOpcode(InsertPt->getOpcode())) && |
| 492 | "Unexpected state change insertion point!" ); |
| 493 | // TODO: Do something to avoid state changes where NZCV is live. |
| 494 | if (MBBI == FirstTerminatorInsertPt) |
| 495 | Block.PhysLiveRegsAtExit = PhysLiveRegs; |
| 496 | if (MBBI == FirstNonPhiInsertPt) |
| 497 | Block.PhysLiveRegsAtEntry = PhysLiveRegs; |
| 498 | if (NeededState != ZAState::ANY) |
| 499 | Block.Insts.push_back(Elt: {.NeededState: NeededState, .InsertPt: InsertPt, .PhysLiveRegs: PhysLiveRegs}); |
| 500 | } |
| 501 | |
| 502 | // Reverse vector (as we had to iterate backwards for liveness). |
| 503 | std::reverse(first: Block.Insts.begin(), last: Block.Insts.end()); |
| 504 | |
| 505 | // Record the desired states on entry/exit of this block. These are the |
| 506 | // states that would not incur a state transition. |
| 507 | if (!Block.Insts.empty()) { |
| 508 | Block.DesiredIncomingState = Block.Insts.front().NeededState; |
| 509 | Block.DesiredOutgoingState = Block.Insts.back().NeededState; |
| 510 | } |
| 511 | } |
| 512 | |
| 513 | return FunctionInfo{.Blocks: std::move(Blocks), .AfterSMEProloguePt: AfterSMEProloguePt, |
| 514 | .PhysLiveRegsAfterSMEPrologue: PhysLiveRegsAfterSMEPrologue}; |
| 515 | } |
| 516 | |
| 517 | /// Assigns each edge bundle a ZA state based on the needed states of blocks |
| 518 | /// that have incoming or outgoing blocks in that bundle. |
| 519 | SmallVector<ZAState> |
| 520 | MachineSMEABI::assignBundleZAStates(const EdgeBundles &Bundles, |
| 521 | const FunctionInfo &FnInfo) { |
| 522 | SmallVector<ZAState> BundleStates(Bundles.getNumBundles()); |
| 523 | for (unsigned I = 0, E = Bundles.getNumBundles(); I != E; ++I) { |
| 524 | std::optional<ZAState> BundleState; |
| 525 | for (unsigned BlockID : Bundles.getBlocks(Bundle: I)) { |
| 526 | const BlockInfo &Block = FnInfo.Blocks[BlockID]; |
| 527 | // Check if the block is an incoming block in the bundle. Note: We skip |
| 528 | // Block.FixedEntryState != ANY to ignore EH pads (which are only |
| 529 | // reachable via exceptions). |
| 530 | if (Block.FixedEntryState != ZAState::ANY || |
| 531 | Bundles.getBundle(N: BlockID, /*Out=*/false) != I) |
| 532 | continue; |
| 533 | |
| 534 | // Pick a state that matches all incoming blocks. Fallback to "ACTIVE" if |
| 535 | // any blocks doesn't match. This will hoist the state from incoming |
| 536 | // blocks to outgoing blocks. |
| 537 | if (!BundleState) |
| 538 | BundleState = Block.DesiredIncomingState; |
| 539 | else if (BundleState != Block.DesiredIncomingState) |
| 540 | BundleState = ZAState::ACTIVE; |
| 541 | } |
| 542 | |
| 543 | if (!BundleState || BundleState == ZAState::ANY) |
| 544 | BundleState = ZAState::ACTIVE; |
| 545 | |
| 546 | BundleStates[I] = *BundleState; |
| 547 | } |
| 548 | |
| 549 | return BundleStates; |
| 550 | } |
| 551 | |
| 552 | std::pair<MachineBasicBlock::iterator, LiveRegs> |
| 553 | MachineSMEABI::findStateChangeInsertionPoint( |
| 554 | MachineBasicBlock &MBB, const BlockInfo &Block, |
| 555 | SmallVectorImpl<InstInfo>::const_iterator Inst) { |
| 556 | LiveRegs PhysLiveRegs; |
| 557 | MachineBasicBlock::iterator InsertPt; |
| 558 | if (Inst != Block.Insts.end()) { |
| 559 | InsertPt = Inst->InsertPt; |
| 560 | PhysLiveRegs = Inst->PhysLiveRegs; |
| 561 | } else { |
| 562 | InsertPt = MBB.getFirstTerminator(); |
| 563 | PhysLiveRegs = Block.PhysLiveRegsAtExit; |
| 564 | } |
| 565 | |
| 566 | if (PhysLiveRegs == LiveRegs::None) |
| 567 | return {InsertPt, PhysLiveRegs}; // Nothing to do (no live regs). |
| 568 | |
| 569 | // Find the previous state change. We can not move before this point. |
| 570 | MachineBasicBlock::iterator PrevStateChangeI; |
| 571 | if (Inst == Block.Insts.begin()) { |
| 572 | PrevStateChangeI = MBB.begin(); |
| 573 | } else { |
| 574 | // Note: `std::prev(Inst)` is the previous InstInfo. We only create an |
| 575 | // InstInfo object for instructions that require a specific ZA state, so the |
| 576 | // InstInfo is the site of the previous state change in the block (which can |
| 577 | // be several MIs earlier). |
| 578 | PrevStateChangeI = std::prev(x: Inst)->InsertPt; |
| 579 | } |
| 580 | |
| 581 | // Note: LiveUnits will only accurately track X0 and NZCV. |
| 582 | LiveRegUnits LiveUnits(*TRI); |
| 583 | setPhysLiveRegs(LiveUnits, PhysLiveRegs); |
| 584 | auto BestCandidate = std::make_pair(x&: InsertPt, y&: PhysLiveRegs); |
| 585 | for (MachineBasicBlock::iterator I = InsertPt; I != PrevStateChangeI; --I) { |
| 586 | // Don't move before/into a call (which may have a state change before it). |
| 587 | if (I->getOpcode() == TII->getCallFrameDestroyOpcode() || I->isCall()) |
| 588 | break; |
| 589 | LiveUnits.stepBackward(MI: *I); |
| 590 | LiveRegs CurrentPhysLiveRegs = getPhysLiveRegs(LiveUnits); |
| 591 | // Find places where NZCV is available, but keep looking for locations where |
| 592 | // both NZCV and X0 are available, which can avoid some copies. |
| 593 | if (!(CurrentPhysLiveRegs & LiveRegs::NZCV)) |
| 594 | BestCandidate = {I, CurrentPhysLiveRegs}; |
| 595 | if (CurrentPhysLiveRegs == LiveRegs::None) |
| 596 | break; |
| 597 | } |
| 598 | return BestCandidate; |
| 599 | } |
| 600 | |
| 601 | void MachineSMEABI::insertStateChanges(EmitContext &Context, |
| 602 | const FunctionInfo &FnInfo, |
| 603 | const EdgeBundles &Bundles, |
| 604 | ArrayRef<ZAState> BundleStates) { |
| 605 | for (MachineBasicBlock &MBB : *MF) { |
| 606 | const BlockInfo &Block = FnInfo.Blocks[MBB.getNumber()]; |
| 607 | ZAState InState = BundleStates[Bundles.getBundle(N: MBB.getNumber(), |
| 608 | /*Out=*/false)]; |
| 609 | |
| 610 | ZAState CurrentState = Block.FixedEntryState; |
| 611 | if (CurrentState == ZAState::ANY) |
| 612 | CurrentState = InState; |
| 613 | |
| 614 | for (auto &Inst : Block.Insts) { |
| 615 | if (CurrentState != Inst.NeededState) { |
| 616 | auto [InsertPt, PhysLiveRegs] = |
| 617 | findStateChangeInsertionPoint(MBB, Block, Inst: &Inst); |
| 618 | emitStateChange(Context, MBB, MBBI: InsertPt, From: CurrentState, To: Inst.NeededState, |
| 619 | PhysLiveRegs); |
| 620 | CurrentState = Inst.NeededState; |
| 621 | } |
| 622 | } |
| 623 | |
| 624 | if (MBB.succ_empty()) |
| 625 | continue; |
| 626 | |
| 627 | ZAState OutState = |
| 628 | BundleStates[Bundles.getBundle(N: MBB.getNumber(), /*Out=*/true)]; |
| 629 | if (CurrentState != OutState) { |
| 630 | auto [InsertPt, PhysLiveRegs] = |
| 631 | findStateChangeInsertionPoint(MBB, Block, Inst: Block.Insts.end()); |
| 632 | emitStateChange(Context, MBB, MBBI: InsertPt, From: CurrentState, To: OutState, |
| 633 | PhysLiveRegs); |
| 634 | } |
| 635 | } |
| 636 | } |
| 637 | |
| 638 | static DebugLoc getDebugLoc(MachineBasicBlock &MBB, |
| 639 | MachineBasicBlock::iterator MBBI) { |
| 640 | if (MBB.empty()) |
| 641 | return DebugLoc(); |
| 642 | return MBBI != MBB.end() ? MBBI->getDebugLoc() : MBB.back().getDebugLoc(); |
| 643 | } |
| 644 | |
| 645 | /// Finds the first call (as determined by MachineInstr::isCall()) starting from |
| 646 | /// \p MBBI in \p MBB marked with \p Marker (which is a marker opcode such as |
| 647 | /// RequiresZASavePseudo). If a marked call is found, it is pushed to \p Calls |
| 648 | /// and the function returns true. |
| 649 | static bool findMarkedCall(const MachineBasicBlock &MBB, |
| 650 | MachineBasicBlock::const_iterator MBBI, |
| 651 | SmallVectorImpl<const MachineInstr *> &Calls, |
| 652 | unsigned Marker, unsigned CallDestroyOpcode) { |
| 653 | auto IsMarker = [&](auto &MI) { return MI.getOpcode() == Marker; }; |
| 654 | auto MarkerInst = std::find_if(first: MBBI, last: MBB.end(), pred: IsMarker); |
| 655 | if (MarkerInst == MBB.end()) |
| 656 | return false; |
| 657 | MachineBasicBlock::const_iterator I = MarkerInst; |
| 658 | while (++I != MBB.end()) { |
| 659 | if (I->isCall() || I->getOpcode() == CallDestroyOpcode) |
| 660 | break; |
| 661 | } |
| 662 | if (I != MBB.end() && I->isCall()) |
| 663 | Calls.push_back(Elt: &*I); |
| 664 | // Note: This function always returns true if a "Marker" was found. |
| 665 | return true; |
| 666 | } |
| 667 | |
| 668 | void MachineSMEABI::collectReachableMarkedCalls( |
| 669 | const MachineBasicBlock &StartMBB, |
| 670 | MachineBasicBlock::const_iterator StartInst, |
| 671 | SmallVectorImpl<const MachineInstr *> &Calls, unsigned Marker) const { |
| 672 | assert(Marker == AArch64::InOutZAUsePseudo || |
| 673 | Marker == AArch64::RequiresZASavePseudo || |
| 674 | Marker == AArch64::RequiresZT0SavePseudo); |
| 675 | unsigned CallDestroyOpcode = TII->getCallFrameDestroyOpcode(); |
| 676 | if (findMarkedCall(MBB: StartMBB, MBBI: StartInst, Calls, Marker, CallDestroyOpcode)) |
| 677 | return; |
| 678 | |
| 679 | SmallPtrSet<const MachineBasicBlock *, 4> Visited; |
| 680 | SmallVector<const MachineBasicBlock *> Worklist(StartMBB.succ_rbegin(), |
| 681 | StartMBB.succ_rend()); |
| 682 | while (!Worklist.empty()) { |
| 683 | const MachineBasicBlock *MBB = Worklist.pop_back_val(); |
| 684 | auto [_, Inserted] = Visited.insert(Ptr: MBB); |
| 685 | if (!Inserted) |
| 686 | continue; |
| 687 | |
| 688 | if (!findMarkedCall(MBB: *MBB, MBBI: MBB->begin(), Calls, Marker, CallDestroyOpcode)) |
| 689 | Worklist.append(in_start: MBB->succ_rbegin(), in_end: MBB->succ_rend()); |
| 690 | } |
| 691 | } |
| 692 | |
| 693 | static StringRef getCalleeName(const MachineInstr &CallInst) { |
| 694 | assert(CallInst.isCall() && "expected a call" ); |
| 695 | for (const MachineOperand &MO : CallInst.operands()) { |
| 696 | if (MO.isSymbol()) |
| 697 | return MO.getSymbolName(); |
| 698 | if (MO.isGlobal()) |
| 699 | return MO.getGlobal()->getName(); |
| 700 | } |
| 701 | return {}; |
| 702 | } |
| 703 | |
| 704 | void MachineSMEABI::(const MachineBasicBlock &MBB, |
| 705 | MachineBasicBlock::const_iterator MBBI, |
| 706 | DebugLoc DL, unsigned Marker, |
| 707 | StringRef , |
| 708 | StringRef SaveName) const { |
| 709 | auto = [&](DebugLoc DL, const MachineBasicBlock &MBB) { |
| 710 | return MachineOptimizationRemarkAnalysis("sme" , RemarkName, DL, &MBB); |
| 711 | }; |
| 712 | StringRef StateName = Marker == AArch64::RequiresZT0SavePseudo ? "ZT0" : "ZA" ; |
| 713 | ORE->emit(RemarkBuilder: [&] { |
| 714 | return SaveRemark(DL, MBB) << SaveName << " of " << StateName |
| 715 | << " emitted in '" << MF->getName() << "'" ; |
| 716 | }); |
| 717 | if (!ORE->allowExtraAnalysis(PassName: "sme" )) |
| 718 | return; |
| 719 | SmallVector<const MachineInstr *> CallsRequiringSaves; |
| 720 | collectReachableMarkedCalls(StartMBB: MBB, StartInst: MBBI, Calls&: CallsRequiringSaves, Marker); |
| 721 | for (const MachineInstr *CallInst : CallsRequiringSaves) { |
| 722 | auto R = SaveRemark(CallInst->getDebugLoc(), *CallInst->getParent()); |
| 723 | R << "call" ; |
| 724 | if (StringRef CalleeName = getCalleeName(CallInst: *CallInst); !CalleeName.empty()) |
| 725 | R << " to '" << CalleeName << "'" ; |
| 726 | R << " requires " << StateName << " save" ; |
| 727 | ORE->emit(OptDiag&: R); |
| 728 | } |
| 729 | } |
| 730 | |
| 731 | void MachineSMEABI::emitSetupLazySave(EmitContext &Context, |
| 732 | MachineBasicBlock &MBB, |
| 733 | MachineBasicBlock::iterator MBBI) { |
| 734 | DebugLoc DL = getDebugLoc(MBB, MBBI); |
| 735 | |
| 736 | emitCallSaveRemarks(MBB, MBBI, DL, Marker: AArch64::RequiresZASavePseudo, |
| 737 | RemarkName: "SMELazySaveZA" , SaveName: "lazy save" ); |
| 738 | |
| 739 | // Get pointer to TPIDR2 block. |
| 740 | Register TPIDR2 = MRI->createVirtualRegister(RegClass: &AArch64::GPR64spRegClass); |
| 741 | Register TPIDR2Ptr = MRI->createVirtualRegister(RegClass: &AArch64::GPR64RegClass); |
| 742 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::ADDXri), DestReg: TPIDR2) |
| 743 | .addFrameIndex(Idx: Context.getTPIDR2Block(MF&: *MF)) |
| 744 | .addImm(Val: 0) |
| 745 | .addImm(Val: 0); |
| 746 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: TPIDR2Ptr) |
| 747 | .addReg(RegNo: TPIDR2); |
| 748 | // Set TPIDR2_EL0 to point to TPIDR2 block. |
| 749 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::MSR)) |
| 750 | .addImm(Val: AArch64SysReg::TPIDR2_EL0) |
| 751 | .addReg(RegNo: TPIDR2Ptr); |
| 752 | } |
| 753 | |
| 754 | PhysRegSave MachineSMEABI::createPhysRegSave(LiveRegs PhysLiveRegs, |
| 755 | MachineBasicBlock &MBB, |
| 756 | MachineBasicBlock::iterator MBBI, |
| 757 | DebugLoc DL) { |
| 758 | PhysRegSave RegSave{.PhysLiveRegs: PhysLiveRegs}; |
| 759 | if (PhysLiveRegs & LiveRegs::NZCV) { |
| 760 | RegSave.StatusFlags = MRI->createVirtualRegister(RegClass: &AArch64::GPR64RegClass); |
| 761 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::MRS), DestReg: RegSave.StatusFlags) |
| 762 | .addImm(Val: AArch64SysReg::NZCV) |
| 763 | .addReg(RegNo: AArch64::NZCV, Flags: RegState::Implicit); |
| 764 | } |
| 765 | // Note: Preserving X0 is "free" as this is before register allocation, so |
| 766 | // the register allocator is still able to optimize these copies. |
| 767 | if (PhysLiveRegs & LiveRegs::W0) { |
| 768 | RegSave.X0Save = MRI->createVirtualRegister(RegClass: PhysLiveRegs & LiveRegs::W0_HI |
| 769 | ? &AArch64::GPR64RegClass |
| 770 | : &AArch64::GPR32RegClass); |
| 771 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: RegSave.X0Save) |
| 772 | .addReg(RegNo: PhysLiveRegs & LiveRegs::W0_HI ? AArch64::X0 : AArch64::W0); |
| 773 | } |
| 774 | return RegSave; |
| 775 | } |
| 776 | |
| 777 | void MachineSMEABI::restorePhyRegSave(const PhysRegSave &RegSave, |
| 778 | MachineBasicBlock &MBB, |
| 779 | MachineBasicBlock::iterator MBBI, |
| 780 | DebugLoc DL) { |
| 781 | if (RegSave.StatusFlags != AArch64::NoRegister) |
| 782 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::MSR)) |
| 783 | .addImm(Val: AArch64SysReg::NZCV) |
| 784 | .addReg(RegNo: RegSave.StatusFlags) |
| 785 | .addReg(RegNo: AArch64::NZCV, Flags: RegState::ImplicitDefine); |
| 786 | |
| 787 | if (RegSave.X0Save != AArch64::NoRegister) |
| 788 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), |
| 789 | DestReg: RegSave.PhysLiveRegs & LiveRegs::W0_HI ? AArch64::X0 : AArch64::W0) |
| 790 | .addReg(RegNo: RegSave.X0Save); |
| 791 | } |
| 792 | |
| 793 | void MachineSMEABI::addSMELibCall(MachineInstrBuilder &MIB, RTLIB::Libcall LC, |
| 794 | CallingConv::ID ExpectedCC) { |
| 795 | RTLIB::LibcallImpl LCImpl = LLI->getLibcallImpl(Call: LC); |
| 796 | if (LCImpl == RTLIB::Unsupported) |
| 797 | emitError(Message: "cannot lower SME ABI (SME routines unsupported)" ); |
| 798 | CallingConv::ID CC = LLI->getLibcallImplCallingConv(Call: LCImpl); |
| 799 | StringRef ImplName = RTLIB::RuntimeLibcallsInfo::getLibcallImplName(CallImpl: LCImpl); |
| 800 | if (CC != ExpectedCC) |
| 801 | emitError(Message: "invalid calling convention for SME routine: '" + ImplName + "'" ); |
| 802 | // FIXME: This assumes the ImplName StringRef is null-terminated. |
| 803 | MIB.addExternalSymbol(FnName: ImplName.data()); |
| 804 | MIB.addRegMask(Mask: TRI->getCallPreservedMask(MF: *MF, CC)); |
| 805 | } |
| 806 | |
| 807 | void MachineSMEABI::emitRestoreLazySave(EmitContext &Context, |
| 808 | MachineBasicBlock &MBB, |
| 809 | MachineBasicBlock::iterator MBBI, |
| 810 | LiveRegs PhysLiveRegs) { |
| 811 | DebugLoc DL = getDebugLoc(MBB, MBBI); |
| 812 | Register TPIDR2EL0 = MRI->createVirtualRegister(RegClass: &AArch64::GPR64RegClass); |
| 813 | Register TPIDR2 = AArch64::X0; |
| 814 | |
| 815 | // TODO: Emit these within the restore MBB to prevent unnecessary saves. |
| 816 | PhysRegSave RegSave = createPhysRegSave(PhysLiveRegs, MBB, MBBI, DL); |
| 817 | |
| 818 | // Enable ZA. |
| 819 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::MSRpstatesvcrImm1)) |
| 820 | .addImm(Val: AArch64SVCR::SVCRZA) |
| 821 | .addImm(Val: 1); |
| 822 | // Get current TPIDR2_EL0. |
| 823 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::MRS), DestReg: TPIDR2EL0) |
| 824 | .addImm(Val: AArch64SysReg::TPIDR2_EL0); |
| 825 | // Get pointer to TPIDR2 block. |
| 826 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::ADDXri), DestReg: TPIDR2) |
| 827 | .addFrameIndex(Idx: Context.getTPIDR2Block(MF&: *MF)) |
| 828 | .addImm(Val: 0) |
| 829 | .addImm(Val: 0); |
| 830 | // (Conditionally) restore ZA state. |
| 831 | auto RestoreZA = BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::RestoreZAPseudo)) |
| 832 | .addReg(RegNo: TPIDR2EL0) |
| 833 | .addReg(RegNo: TPIDR2); |
| 834 | addSMELibCall( |
| 835 | MIB&: RestoreZA, LC: RTLIB::SMEABI_TPIDR2_RESTORE, |
| 836 | ExpectedCC: CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0); |
| 837 | // Zero TPIDR2_EL0. |
| 838 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::MSR)) |
| 839 | .addImm(Val: AArch64SysReg::TPIDR2_EL0) |
| 840 | .addReg(RegNo: AArch64::XZR); |
| 841 | |
| 842 | restorePhyRegSave(RegSave, MBB, MBBI, DL); |
| 843 | } |
| 844 | |
| 845 | void MachineSMEABI::emitZAMode(MachineBasicBlock &MBB, |
| 846 | MachineBasicBlock::iterator MBBI, |
| 847 | bool ClearTPIDR2, bool On) { |
| 848 | DebugLoc DL = getDebugLoc(MBB, MBBI); |
| 849 | |
| 850 | if (ClearTPIDR2) |
| 851 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::MSR)) |
| 852 | .addImm(Val: AArch64SysReg::TPIDR2_EL0) |
| 853 | .addReg(RegNo: AArch64::XZR); |
| 854 | |
| 855 | // Disable ZA. |
| 856 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::MSRpstatesvcrImm1)) |
| 857 | .addImm(Val: AArch64SVCR::SVCRZA) |
| 858 | .addImm(Val: On ? 1 : 0); |
| 859 | } |
| 860 | |
| 861 | void MachineSMEABI::emitAllocateLazySaveBuffer( |
| 862 | EmitContext &Context, MachineBasicBlock &MBB, |
| 863 | MachineBasicBlock::iterator MBBI) { |
| 864 | MachineFrameInfo &MFI = MF->getFrameInfo(); |
| 865 | DebugLoc DL = getDebugLoc(MBB, MBBI); |
| 866 | Register SP = MRI->createVirtualRegister(RegClass: &AArch64::GPR64RegClass); |
| 867 | Register SVL = MRI->createVirtualRegister(RegClass: &AArch64::GPR64RegClass); |
| 868 | Register Buffer = AFI->getEarlyAllocSMESaveBuffer(); |
| 869 | |
| 870 | // Calculate SVL. |
| 871 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::RDSVLI_XI), DestReg: SVL).addImm(Val: 1); |
| 872 | |
| 873 | // 1. Allocate the lazy save buffer. |
| 874 | if (Buffer == AArch64::NoRegister) { |
| 875 | // TODO: On Windows, we allocate the lazy save buffer in SelectionDAG (so |
| 876 | // Buffer != AArch64::NoRegister). This is done to reuse the existing |
| 877 | // expansions (which can insert stack checks). This works, but it means we |
| 878 | // will always allocate the lazy save buffer (even if the function contains |
| 879 | // no lazy saves). If we want to handle Windows here, we'll need to |
| 880 | // implement something similar to LowerWindowsDYNAMIC_STACKALLOC. |
| 881 | assert(!Subtarget->isTargetWindows() && |
| 882 | "Lazy ZA save is not yet supported on Windows" ); |
| 883 | Buffer = MRI->createVirtualRegister(RegClass: &AArch64::GPR64RegClass); |
| 884 | // Get original stack pointer. |
| 885 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: SP) |
| 886 | .addReg(RegNo: AArch64::SP); |
| 887 | // Allocate a lazy-save buffer object of the size given, normally SVL * SVL |
| 888 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::MSUBXrrr), DestReg: Buffer) |
| 889 | .addReg(RegNo: SVL) |
| 890 | .addReg(RegNo: SVL) |
| 891 | .addReg(RegNo: SP); |
| 892 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: AArch64::SP) |
| 893 | .addReg(RegNo: Buffer); |
| 894 | // We have just allocated a variable sized object, tell this to PEI. |
| 895 | MFI.CreateVariableSizedObject(Alignment: Align(16), Alloca: nullptr); |
| 896 | } |
| 897 | |
| 898 | // 2. Setup the TPIDR2 block. |
| 899 | { |
| 900 | // Note: This case just needs to do `SVL << 48`. It is not implemented as we |
| 901 | // generally don't support big-endian SVE/SME. |
| 902 | if (!Subtarget->isLittleEndian()) |
| 903 | reportFatalInternalError( |
| 904 | reason: "TPIDR2 block initialization is not supported on big-endian targets" ); |
| 905 | |
| 906 | // Store buffer pointer and num_za_save_slices. |
| 907 | // Bytes 10-15 are implicitly zeroed. |
| 908 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::STPXi)) |
| 909 | .addReg(RegNo: Buffer) |
| 910 | .addReg(RegNo: SVL) |
| 911 | .addFrameIndex(Idx: Context.getTPIDR2Block(MF&: *MF)) |
| 912 | .addImm(Val: 0); |
| 913 | } |
| 914 | } |
| 915 | |
| 916 | static constexpr unsigned ZERO_ALL_ZA_MASK = 0b11111111; |
| 917 | |
| 918 | void MachineSMEABI::emitSMEPrologue(MachineBasicBlock &MBB, |
| 919 | MachineBasicBlock::iterator MBBI) { |
| 920 | DebugLoc DL = getDebugLoc(MBB, MBBI); |
| 921 | |
| 922 | bool ZeroZA = AFI->getSMEFnAttrs().isNewZA(); |
| 923 | bool ZeroZT0 = AFI->getSMEFnAttrs().isNewZT0(); |
| 924 | if (AFI->getSMEFnAttrs().hasPrivateZAInterface()) { |
| 925 | // Get current TPIDR2_EL0. |
| 926 | Register TPIDR2EL0 = MRI->createVirtualRegister(RegClass: &AArch64::GPR64RegClass); |
| 927 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::MRS)) |
| 928 | .addReg(RegNo: TPIDR2EL0, Flags: RegState::Define) |
| 929 | .addImm(Val: AArch64SysReg::TPIDR2_EL0); |
| 930 | // If TPIDR2_EL0 is non-zero, commit the lazy save. |
| 931 | // NOTE: Functions that only use ZT0 don't need to zero ZA. |
| 932 | auto CommitZASave = |
| 933 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::CommitZASavePseudo)) |
| 934 | .addReg(RegNo: TPIDR2EL0) |
| 935 | .addImm(Val: ZeroZA) |
| 936 | .addImm(Val: ZeroZT0); |
| 937 | addSMELibCall( |
| 938 | MIB&: CommitZASave, LC: RTLIB::SMEABI_TPIDR2_SAVE, |
| 939 | ExpectedCC: CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0); |
| 940 | if (ZeroZA) |
| 941 | CommitZASave.addDef(RegNo: AArch64::ZAB0, Flags: RegState::ImplicitDefine); |
| 942 | if (ZeroZT0) |
| 943 | CommitZASave.addDef(RegNo: AArch64::ZT0, Flags: RegState::ImplicitDefine); |
| 944 | // Enable ZA (as ZA could have previously been in the OFF state). |
| 945 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::MSRpstatesvcrImm1)) |
| 946 | .addImm(Val: AArch64SVCR::SVCRZA) |
| 947 | .addImm(Val: 1); |
| 948 | } else if (AFI->getSMEFnAttrs().hasSharedZAInterface()) { |
| 949 | if (ZeroZA) |
| 950 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::ZERO_M)) |
| 951 | .addImm(Val: ZERO_ALL_ZA_MASK) |
| 952 | .addDef(RegNo: AArch64::ZAB0, Flags: RegState::ImplicitDefine); |
| 953 | if (ZeroZT0) |
| 954 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::ZERO_T)).addDef(RegNo: AArch64::ZT0); |
| 955 | } |
| 956 | } |
| 957 | |
| 958 | void MachineSMEABI::emitFullZASaveRestore(EmitContext &Context, |
| 959 | MachineBasicBlock &MBB, |
| 960 | MachineBasicBlock::iterator MBBI, |
| 961 | LiveRegs PhysLiveRegs, bool IsSave) { |
| 962 | DebugLoc DL = getDebugLoc(MBB, MBBI); |
| 963 | |
| 964 | if (IsSave) |
| 965 | emitCallSaveRemarks(MBB, MBBI, DL, Marker: AArch64::RequiresZASavePseudo, |
| 966 | RemarkName: "SMEFullZASave" , SaveName: "full save" ); |
| 967 | |
| 968 | PhysRegSave RegSave = createPhysRegSave(PhysLiveRegs, MBB, MBBI, DL); |
| 969 | |
| 970 | // Copy the buffer pointer into X0. |
| 971 | Register BufferPtr = AArch64::X0; |
| 972 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: BufferPtr) |
| 973 | .addReg(RegNo: Context.getAgnosticZABufferPtr(MF&: *MF)); |
| 974 | |
| 975 | // Call __arm_sme_save/__arm_sme_restore. |
| 976 | auto SaveRestoreZA = BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::BL)) |
| 977 | .addReg(RegNo: BufferPtr, Flags: RegState::Implicit); |
| 978 | addSMELibCall( |
| 979 | MIB&: SaveRestoreZA, |
| 980 | LC: IsSave ? RTLIB::SMEABI_SME_SAVE : RTLIB::SMEABI_SME_RESTORE, |
| 981 | ExpectedCC: CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1); |
| 982 | |
| 983 | restorePhyRegSave(RegSave, MBB, MBBI, DL); |
| 984 | } |
| 985 | |
| 986 | void MachineSMEABI::emitZT0SaveRestore(EmitContext &Context, |
| 987 | MachineBasicBlock &MBB, |
| 988 | MachineBasicBlock::iterator MBBI, |
| 989 | bool IsSave) { |
| 990 | DebugLoc DL = getDebugLoc(MBB, MBBI); |
| 991 | |
| 992 | // Note: This will report calls that _only_ need ZT0 saved. Call that save |
| 993 | // both ZA and ZT0 will be under the SMELazySaveZA remark. This prevents |
| 994 | // reporting the same calls twice. |
| 995 | if (IsSave) |
| 996 | emitCallSaveRemarks(MBB, MBBI, DL, Marker: AArch64::RequiresZT0SavePseudo, |
| 997 | RemarkName: "SMEZT0Save" , SaveName: "spill" ); |
| 998 | |
| 999 | Register ZT0Save = MRI->createVirtualRegister(RegClass: &AArch64::GPR64spRegClass); |
| 1000 | |
| 1001 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::ADDXri), DestReg: ZT0Save) |
| 1002 | .addFrameIndex(Idx: Context.getZT0SaveSlot(MF&: *MF)) |
| 1003 | .addImm(Val: 0) |
| 1004 | .addImm(Val: 0); |
| 1005 | |
| 1006 | if (IsSave) { |
| 1007 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::STR_TX)) |
| 1008 | .addReg(RegNo: AArch64::ZT0) |
| 1009 | .addReg(RegNo: ZT0Save); |
| 1010 | } else { |
| 1011 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::LDR_TX), DestReg: AArch64::ZT0) |
| 1012 | .addReg(RegNo: ZT0Save); |
| 1013 | } |
| 1014 | } |
| 1015 | |
| 1016 | void MachineSMEABI::emitAllocateFullZASaveBuffer( |
| 1017 | EmitContext &Context, MachineBasicBlock &MBB, |
| 1018 | MachineBasicBlock::iterator MBBI, LiveRegs PhysLiveRegs) { |
| 1019 | // Buffer already allocated in SelectionDAG. |
| 1020 | if (AFI->getEarlyAllocSMESaveBuffer()) |
| 1021 | return; |
| 1022 | |
| 1023 | DebugLoc DL = getDebugLoc(MBB, MBBI); |
| 1024 | Register BufferPtr = Context.getAgnosticZABufferPtr(MF&: *MF); |
| 1025 | Register BufferSize = MRI->createVirtualRegister(RegClass: &AArch64::GPR64RegClass); |
| 1026 | |
| 1027 | PhysRegSave RegSave = createPhysRegSave(PhysLiveRegs, MBB, MBBI, DL); |
| 1028 | |
| 1029 | // Calculate the SME state size. |
| 1030 | { |
| 1031 | auto SMEStateSize = BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::BL)) |
| 1032 | .addReg(RegNo: AArch64::X0, Flags: RegState::ImplicitDefine); |
| 1033 | addSMELibCall( |
| 1034 | MIB&: SMEStateSize, LC: RTLIB::SMEABI_SME_STATE_SIZE, |
| 1035 | ExpectedCC: CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1); |
| 1036 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: BufferSize) |
| 1037 | .addReg(RegNo: AArch64::X0); |
| 1038 | } |
| 1039 | |
| 1040 | // Allocate a buffer object of the size given __arm_sme_state_size. |
| 1041 | { |
| 1042 | MachineFrameInfo &MFI = MF->getFrameInfo(); |
| 1043 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::SUBXrx64), DestReg: AArch64::SP) |
| 1044 | .addReg(RegNo: AArch64::SP) |
| 1045 | .addReg(RegNo: BufferSize) |
| 1046 | .addImm(Val: AArch64_AM::getArithExtendImm(ET: AArch64_AM::UXTX, Imm: 0)); |
| 1047 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: BufferPtr) |
| 1048 | .addReg(RegNo: AArch64::SP); |
| 1049 | |
| 1050 | // We have just allocated a variable sized object, tell this to PEI. |
| 1051 | MFI.CreateVariableSizedObject(Alignment: Align(16), Alloca: nullptr); |
| 1052 | } |
| 1053 | |
| 1054 | restorePhyRegSave(RegSave, MBB, MBBI, DL); |
| 1055 | } |
| 1056 | |
| 1057 | struct FromState { |
| 1058 | ZAState From; |
| 1059 | |
| 1060 | constexpr uint8_t to(ZAState To) const { |
| 1061 | static_assert(NUM_ZA_STATE < 16, "expected ZAState to fit in 4-bits" ); |
| 1062 | return uint8_t(From) << 4 | uint8_t(To); |
| 1063 | } |
| 1064 | }; |
| 1065 | |
| 1066 | constexpr FromState transitionFrom(ZAState From) { return FromState{.From: From}; } |
| 1067 | |
| 1068 | void MachineSMEABI::emitStateChange(EmitContext &Context, |
| 1069 | MachineBasicBlock &MBB, |
| 1070 | MachineBasicBlock::iterator InsertPt, |
| 1071 | ZAState From, ZAState To, |
| 1072 | LiveRegs PhysLiveRegs) { |
| 1073 | // ZA not used. |
| 1074 | if (From == ZAState::ANY || To == ZAState::ANY) |
| 1075 | return; |
| 1076 | |
| 1077 | // If we're exiting from the ENTRY state that means that the function has not |
| 1078 | // used ZA, so in the case of private ZA/ZT0 functions we can omit any set up. |
| 1079 | if (From == ZAState::ENTRY && To == ZAState::OFF) |
| 1080 | return; |
| 1081 | |
| 1082 | // TODO: Avoid setting up the save buffer if there's no transition to |
| 1083 | // LOCAL_SAVED. |
| 1084 | if (From == ZAState::ENTRY) { |
| 1085 | assert(&MBB == &MBB.getParent()->front() && |
| 1086 | "ENTRY state only valid in entry block" ); |
| 1087 | emitSMEPrologue(MBB, MBBI: MBB.getFirstNonPHI()); |
| 1088 | if (To == ZAState::ACTIVE) |
| 1089 | return; // Nothing more to do (ZA is active after the prologue). |
| 1090 | |
| 1091 | // Note: "emitNewZAPrologue" zeros ZA, so we may need to setup a lazy save |
| 1092 | // if "To" is "ZAState::LOCAL_SAVED". It may be possible to improve this |
| 1093 | // case by changing the placement of the zero instruction. |
| 1094 | From = ZAState::ACTIVE; |
| 1095 | } |
| 1096 | |
| 1097 | SMEAttrs SMEFnAttrs = AFI->getSMEFnAttrs(); |
| 1098 | bool IsAgnosticZA = SMEFnAttrs.hasAgnosticZAInterface(); |
| 1099 | bool HasZT0State = SMEFnAttrs.hasZT0State(); |
| 1100 | bool HasZAState = IsAgnosticZA || SMEFnAttrs.hasZAState(); |
| 1101 | |
| 1102 | switch (transitionFrom(From).to(To)) { |
| 1103 | // This section handles: ACTIVE <-> ACTIVE_ZT0_SAVED |
| 1104 | case transitionFrom(From: ZAState::ACTIVE).to(To: ZAState::ACTIVE_ZT0_SAVED): |
| 1105 | emitZT0SaveRestore(Context, MBB, MBBI: InsertPt, /*IsSave=*/true); |
| 1106 | break; |
| 1107 | case transitionFrom(From: ZAState::ACTIVE_ZT0_SAVED).to(To: ZAState::ACTIVE): |
| 1108 | emitZT0SaveRestore(Context, MBB, MBBI: InsertPt, /*IsSave=*/false); |
| 1109 | break; |
| 1110 | |
| 1111 | // This section handles: ACTIVE[_ZT0_SAVED] -> LOCAL_SAVED |
| 1112 | case transitionFrom(From: ZAState::ACTIVE).to(To: ZAState::LOCAL_SAVED): |
| 1113 | case transitionFrom(From: ZAState::ACTIVE_ZT0_SAVED).to(To: ZAState::LOCAL_SAVED): |
| 1114 | if (HasZT0State && From == ZAState::ACTIVE) |
| 1115 | emitZT0SaveRestore(Context, MBB, MBBI: InsertPt, /*IsSave=*/true); |
| 1116 | if (HasZAState) |
| 1117 | emitZASave(Context, MBB, MBBI: InsertPt, PhysLiveRegs); |
| 1118 | break; |
| 1119 | |
| 1120 | // This section handles: ACTIVE -> LOCAL_COMMITTED |
| 1121 | case transitionFrom(From: ZAState::ACTIVE).to(To: ZAState::LOCAL_COMMITTED): |
| 1122 | // TODO: We could support ZA state here, but this transition is currently |
| 1123 | // only possible when we _don't_ have ZA state. |
| 1124 | assert(HasZT0State && !HasZAState && "Expect to only have ZT0 state." ); |
| 1125 | emitZT0SaveRestore(Context, MBB, MBBI: InsertPt, /*IsSave=*/true); |
| 1126 | emitZAMode(MBB, MBBI: InsertPt, /*ClearTPIDR2=*/false, /*On=*/false); |
| 1127 | break; |
| 1128 | |
| 1129 | // This section handles: LOCAL_COMMITTED -> (OFF|LOCAL_SAVED) |
| 1130 | case transitionFrom(From: ZAState::LOCAL_COMMITTED).to(To: ZAState::OFF): |
| 1131 | case transitionFrom(From: ZAState::LOCAL_COMMITTED).to(To: ZAState::LOCAL_SAVED): |
| 1132 | // These transistions are a no-op. |
| 1133 | break; |
| 1134 | |
| 1135 | // This section handles: LOCAL_(SAVED|COMMITTED) -> ACTIVE[_ZT0_SAVED] |
| 1136 | case transitionFrom(From: ZAState::LOCAL_COMMITTED).to(To: ZAState::ACTIVE): |
| 1137 | case transitionFrom(From: ZAState::LOCAL_COMMITTED).to(To: ZAState::ACTIVE_ZT0_SAVED): |
| 1138 | case transitionFrom(From: ZAState::LOCAL_SAVED).to(To: ZAState::ACTIVE): |
| 1139 | if (HasZAState) |
| 1140 | emitZARestore(Context, MBB, MBBI: InsertPt, PhysLiveRegs); |
| 1141 | else |
| 1142 | emitZAMode(MBB, MBBI: InsertPt, /*ClearTPIDR2=*/false, /*On=*/true); |
| 1143 | if (HasZT0State && To == ZAState::ACTIVE) |
| 1144 | emitZT0SaveRestore(Context, MBB, MBBI: InsertPt, /*IsSave=*/false); |
| 1145 | break; |
| 1146 | |
| 1147 | // This section handles transistions to OFF (not previously covered) |
| 1148 | case transitionFrom(From: ZAState::ACTIVE).to(To: ZAState::OFF): |
| 1149 | case transitionFrom(From: ZAState::ACTIVE_ZT0_SAVED).to(To: ZAState::OFF): |
| 1150 | case transitionFrom(From: ZAState::LOCAL_SAVED).to(To: ZAState::OFF): |
| 1151 | assert(SMEFnAttrs.hasPrivateZAInterface() && |
| 1152 | "Did not expect to turn ZA off in shared/agnostic ZA function" ); |
| 1153 | emitZAMode(MBB, MBBI: InsertPt, /*ClearTPIDR2=*/From == ZAState::LOCAL_SAVED, |
| 1154 | /*On=*/false); |
| 1155 | break; |
| 1156 | |
| 1157 | default: |
| 1158 | dbgs() << "Error: Transition from " << getZAStateString(State: From) << " to " |
| 1159 | << getZAStateString(State: To) << '\n'; |
| 1160 | llvm_unreachable("Unimplemented state transition" ); |
| 1161 | } |
| 1162 | } |
| 1163 | |
| 1164 | } // end anonymous namespace |
| 1165 | |
| 1166 | INITIALIZE_PASS(MachineSMEABI, "aarch64-machine-sme-abi" , "Machine SME ABI" , |
| 1167 | false, false) |
| 1168 | |
| 1169 | bool MachineSMEABI::runOnMachineFunction(MachineFunction &MF) { |
| 1170 | Subtarget = &MF.getSubtarget<AArch64Subtarget>(); |
| 1171 | if (!Subtarget->hasSME()) |
| 1172 | return false; |
| 1173 | |
| 1174 | AFI = MF.getInfo<AArch64FunctionInfo>(); |
| 1175 | SMEAttrs SMEFnAttrs = AFI->getSMEFnAttrs(); |
| 1176 | if (!SMEFnAttrs.hasZAState() && !SMEFnAttrs.hasZT0State() && |
| 1177 | !SMEFnAttrs.hasAgnosticZAInterface()) |
| 1178 | return false; |
| 1179 | |
| 1180 | assert(MF.getRegInfo().isSSA() && "Expected to be run on SSA form!" ); |
| 1181 | |
| 1182 | this->MF = &MF; |
| 1183 | ORE = &getAnalysis<MachineOptimizationRemarkEmitterPass>().getORE(); |
| 1184 | LLI = &getAnalysis<LibcallLoweringInfoWrapper>().getLibcallLowering( |
| 1185 | M: *MF.getFunction().getParent(), Subtarget: *Subtarget); |
| 1186 | TII = Subtarget->getInstrInfo(); |
| 1187 | TRI = Subtarget->getRegisterInfo(); |
| 1188 | MRI = &MF.getRegInfo(); |
| 1189 | |
| 1190 | const EdgeBundles &Bundles = |
| 1191 | getAnalysis<EdgeBundlesWrapperLegacy>().getEdgeBundles(); |
| 1192 | |
| 1193 | FunctionInfo FnInfo = collectNeededZAStates(SMEFnAttrs); |
| 1194 | |
| 1195 | SmallVector<ZAState> BundleStates = assignBundleZAStates(Bundles, FnInfo); |
| 1196 | |
| 1197 | EmitContext Context; |
| 1198 | insertStateChanges(Context, FnInfo, Bundles, BundleStates); |
| 1199 | |
| 1200 | if (Context.needsSaveBuffer()) { |
| 1201 | if (FnInfo.AfterSMEProloguePt) { |
| 1202 | // Note: With inline stack probes the AfterSMEProloguePt may not be in the |
| 1203 | // entry block (due to the probing loop). |
| 1204 | MachineBasicBlock::iterator MBBI = *FnInfo.AfterSMEProloguePt; |
| 1205 | emitAllocateZASaveBuffer(Context, MBB&: *MBBI->getParent(), MBBI, |
| 1206 | PhysLiveRegs: FnInfo.PhysLiveRegsAfterSMEPrologue); |
| 1207 | } else { |
| 1208 | MachineBasicBlock &EntryBlock = MF.front(); |
| 1209 | emitAllocateZASaveBuffer( |
| 1210 | Context, MBB&: EntryBlock, MBBI: EntryBlock.getFirstNonPHI(), |
| 1211 | PhysLiveRegs: FnInfo.Blocks[EntryBlock.getNumber()].PhysLiveRegsAtEntry); |
| 1212 | } |
| 1213 | } |
| 1214 | |
| 1215 | return true; |
| 1216 | } |
| 1217 | |
| 1218 | FunctionPass *llvm::createMachineSMEABIPass(CodeGenOptLevel OptLevel) { |
| 1219 | return new MachineSMEABI(OptLevel); |
| 1220 | } |
| 1221 | |