| 1 | //===- llvm/CodeGen/TargetInstrInfo.h - Instruction Info --------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file describes the target machine instruction set to the code generator. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #ifndef LLVM_CODEGEN_TARGETINSTRINFO_H |
| 14 | #define LLVM_CODEGEN_TARGETINSTRINFO_H |
| 15 | |
| 16 | #include "llvm/ADT/ArrayRef.h" |
| 17 | #include "llvm/ADT/DenseMap.h" |
| 18 | #include "llvm/ADT/DenseMapInfo.h" |
| 19 | #include "llvm/ADT/Uniformity.h" |
| 20 | #include "llvm/CodeGen/MIRFormatter.h" |
| 21 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 22 | #include "llvm/CodeGen/MachineCombinerPattern.h" |
| 23 | #include "llvm/CodeGen/MachineCycleAnalysis.h" |
| 24 | #include "llvm/CodeGen/MachineFunction.h" |
| 25 | #include "llvm/CodeGen/MachineInstr.h" |
| 26 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 27 | #include "llvm/CodeGen/MachineOperand.h" |
| 28 | #include "llvm/CodeGen/MachineOutliner.h" |
| 29 | #include "llvm/CodeGen/RegisterClassInfo.h" |
| 30 | #include "llvm/CodeGen/VirtRegMap.h" |
| 31 | #include "llvm/MC/MCInstrInfo.h" |
| 32 | #include "llvm/Support/BranchProbability.h" |
| 33 | #include "llvm/Support/Compiler.h" |
| 34 | #include "llvm/Support/ErrorHandling.h" |
| 35 | #include "llvm/Support/TypeSize.h" |
| 36 | #include <array> |
| 37 | #include <cassert> |
| 38 | #include <cstddef> |
| 39 | #include <cstdint> |
| 40 | #include <utility> |
| 41 | #include <vector> |
| 42 | |
| 43 | namespace llvm { |
| 44 | |
| 45 | class DFAPacketizer; |
| 46 | class InstrItineraryData; |
| 47 | class LiveIntervals; |
| 48 | class LiveVariables; |
| 49 | class MachineLoop; |
| 50 | class MachineMemOperand; |
| 51 | class MachineModuleInfo; |
| 52 | class MachineRegisterInfo; |
| 53 | class MCAsmInfo; |
| 54 | class MCInst; |
| 55 | struct MCSchedModel; |
| 56 | class Module; |
| 57 | class ScheduleDAG; |
| 58 | class ScheduleDAGMI; |
| 59 | class ScheduleHazardRecognizer; |
| 60 | class SDNode; |
| 61 | class SelectionDAG; |
| 62 | class SMSchedule; |
| 63 | class SwingSchedulerDAG; |
| 64 | class RegScavenger; |
| 65 | class TargetRegisterClass; |
| 66 | class TargetRegisterInfo; |
| 67 | class TargetSchedModel; |
| 68 | class TargetSubtargetInfo; |
| 69 | enum class MachineTraceStrategy; |
| 70 | |
| 71 | template <class T> class SmallVectorImpl; |
| 72 | |
| 73 | using ParamLoadedValue = std::pair<MachineOperand, DIExpression*>; |
| 74 | |
| 75 | struct DestSourcePair { |
| 76 | const MachineOperand *Destination; |
| 77 | const MachineOperand *Source; |
| 78 | |
| 79 | DestSourcePair(const MachineOperand &Dest, const MachineOperand &Src) |
| 80 | : Destination(&Dest), Source(&Src) {} |
| 81 | }; |
| 82 | |
| 83 | /// Used to describe a register and immediate addition. |
| 84 | struct RegImmPair { |
| 85 | Register Reg; |
| 86 | int64_t Imm; |
| 87 | |
| 88 | RegImmPair(Register Reg, int64_t Imm) : Reg(Reg), Imm(Imm) {} |
| 89 | }; |
| 90 | |
| 91 | /// Used to describe addressing mode similar to ExtAddrMode in CodeGenPrepare. |
| 92 | /// It holds the register values, the scale value and the displacement. |
| 93 | /// It also holds a descriptor for the expression used to calculate the address |
| 94 | /// from the operands. |
| 95 | struct ExtAddrMode { |
| 96 | enum class Formula { |
| 97 | Basic = 0, // BaseReg + ScaledReg * Scale + Displacement |
| 98 | SExtScaledReg = 1, // BaseReg + sext(ScaledReg) * Scale + Displacement |
| 99 | ZExtScaledReg = 2 // BaseReg + zext(ScaledReg) * Scale + Displacement |
| 100 | }; |
| 101 | |
| 102 | Register BaseReg; |
| 103 | Register ScaledReg; |
| 104 | int64_t Scale = 0; |
| 105 | int64_t Displacement = 0; |
| 106 | Formula Form = Formula::Basic; |
| 107 | ExtAddrMode() = default; |
| 108 | }; |
| 109 | |
| 110 | //--------------------------------------------------------------------------- |
| 111 | /// |
| 112 | /// TargetInstrInfo - Interface to description of machine instruction set |
| 113 | /// |
| 114 | class LLVM_ABI TargetInstrInfo : public MCInstrInfo { |
| 115 | public: |
| 116 | TargetInstrInfo(unsigned CFSetupOpcode = ~0u, unsigned CFDestroyOpcode = ~0u, |
| 117 | unsigned CatchRetOpcode = ~0u, unsigned ReturnOpcode = ~0u) |
| 118 | : CallFrameSetupOpcode(CFSetupOpcode), |
| 119 | CallFrameDestroyOpcode(CFDestroyOpcode), CatchRetOpcode(CatchRetOpcode), |
| 120 | ReturnOpcode(ReturnOpcode) {} |
| 121 | TargetInstrInfo(const TargetInstrInfo &) = delete; |
| 122 | TargetInstrInfo &operator=(const TargetInstrInfo &) = delete; |
| 123 | virtual ~TargetInstrInfo(); |
| 124 | |
| 125 | static bool isGenericOpcode(unsigned Opc) { |
| 126 | return Opc <= TargetOpcode::GENERIC_OP_END; |
| 127 | } |
| 128 | |
| 129 | static bool isGenericAtomicRMWOpcode(unsigned Opc) { |
| 130 | return Opc >= TargetOpcode::GENERIC_ATOMICRMW_OP_START && |
| 131 | Opc <= TargetOpcode::GENERIC_ATOMICRMW_OP_END; |
| 132 | } |
| 133 | |
| 134 | /// Given a machine instruction descriptor, returns the register |
| 135 | /// class constraint for OpNum, or NULL. |
| 136 | virtual |
| 137 | const TargetRegisterClass *getRegClass(const MCInstrDesc &MCID, unsigned OpNum, |
| 138 | const TargetRegisterInfo *TRI, |
| 139 | const MachineFunction &MF) const; |
| 140 | |
| 141 | /// Returns true if MI is an instruction we are unable to reason about |
| 142 | /// (like a call or something with unmodeled side effects). |
| 143 | virtual bool isGlobalMemoryObject(const MachineInstr *MI) const; |
| 144 | |
| 145 | /// Return true if the instruction is trivially rematerializable, meaning it |
| 146 | /// has no side effects and requires no operands that aren't always available. |
| 147 | /// This means the only allowed uses are constants and unallocatable physical |
| 148 | /// registers so that the instructions result is independent of the place |
| 149 | /// in the function. |
| 150 | bool isTriviallyReMaterializable(const MachineInstr &MI) const { |
| 151 | return (MI.getOpcode() == TargetOpcode::IMPLICIT_DEF && |
| 152 | MI.getNumOperands() == 1) || |
| 153 | (MI.getDesc().isRematerializable() && |
| 154 | isReallyTriviallyReMaterializable(MI)); |
| 155 | } |
| 156 | |
| 157 | /// Given \p MO is a PhysReg use return if it can be ignored for the purpose |
| 158 | /// of instruction rematerialization or sinking. |
| 159 | virtual bool isIgnorableUse(const MachineOperand &MO) const { |
| 160 | return false; |
| 161 | } |
| 162 | |
| 163 | virtual bool isSafeToSink(MachineInstr &MI, MachineBasicBlock *SuccToSinkTo, |
| 164 | MachineCycleInfo *CI) const { |
| 165 | return true; |
| 166 | } |
| 167 | |
| 168 | /// For a "cheap" instruction which doesn't enable additional sinking, |
| 169 | /// should MachineSink break a critical edge to sink it anyways? |
| 170 | virtual bool shouldBreakCriticalEdgeToSink(MachineInstr &MI) const { |
| 171 | return false; |
| 172 | } |
| 173 | |
| 174 | protected: |
| 175 | /// For instructions with opcodes for which the M_REMATERIALIZABLE flag is |
| 176 | /// set, this hook lets the target specify whether the instruction is actually |
| 177 | /// trivially rematerializable, taking into consideration its operands. This |
| 178 | /// predicate must return false if the instruction has any side effects other |
| 179 | /// than producing a value, or if it requres any address registers that are |
| 180 | /// not always available. |
| 181 | virtual bool isReallyTriviallyReMaterializable(const MachineInstr &MI) const; |
| 182 | |
| 183 | /// This method commutes the operands of the given machine instruction MI. |
| 184 | /// The operands to be commuted are specified by their indices OpIdx1 and |
| 185 | /// OpIdx2. |
| 186 | /// |
| 187 | /// If a target has any instructions that are commutable but require |
| 188 | /// converting to different instructions or making non-trivial changes |
| 189 | /// to commute them, this method can be overloaded to do that. |
| 190 | /// The default implementation simply swaps the commutable operands. |
| 191 | /// |
| 192 | /// If NewMI is false, MI is modified in place and returned; otherwise, a |
| 193 | /// new machine instruction is created and returned. |
| 194 | /// |
| 195 | /// Do not call this method for a non-commutable instruction. |
| 196 | /// Even though the instruction is commutable, the method may still |
| 197 | /// fail to commute the operands, null pointer is returned in such cases. |
| 198 | virtual MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI, |
| 199 | unsigned OpIdx1, |
| 200 | unsigned OpIdx2) const; |
| 201 | |
| 202 | /// Assigns the (CommutableOpIdx1, CommutableOpIdx2) pair of commutable |
| 203 | /// operand indices to (ResultIdx1, ResultIdx2). |
| 204 | /// One or both input values of the pair: (ResultIdx1, ResultIdx2) may be |
| 205 | /// predefined to some indices or be undefined (designated by the special |
| 206 | /// value 'CommuteAnyOperandIndex'). |
| 207 | /// The predefined result indices cannot be re-defined. |
| 208 | /// The function returns true iff after the result pair redefinition |
| 209 | /// the fixed result pair is equal to or equivalent to the source pair of |
| 210 | /// indices: (CommutableOpIdx1, CommutableOpIdx2). It is assumed here that |
| 211 | /// the pairs (x,y) and (y,x) are equivalent. |
| 212 | static bool fixCommutedOpIndices(unsigned &ResultIdx1, unsigned &ResultIdx2, |
| 213 | unsigned CommutableOpIdx1, |
| 214 | unsigned CommutableOpIdx2); |
| 215 | |
| 216 | public: |
| 217 | /// These methods return the opcode of the frame setup/destroy instructions |
| 218 | /// if they exist (-1 otherwise). Some targets use pseudo instructions in |
| 219 | /// order to abstract away the difference between operating with a frame |
| 220 | /// pointer and operating without, through the use of these two instructions. |
| 221 | /// A FrameSetup MI in MF implies MFI::AdjustsStack. |
| 222 | /// |
| 223 | unsigned getCallFrameSetupOpcode() const { return CallFrameSetupOpcode; } |
| 224 | unsigned getCallFrameDestroyOpcode() const { return CallFrameDestroyOpcode; } |
| 225 | |
| 226 | /// Returns true if the argument is a frame pseudo instruction. |
| 227 | bool isFrameInstr(const MachineInstr &I) const { |
| 228 | return I.getOpcode() == getCallFrameSetupOpcode() || |
| 229 | I.getOpcode() == getCallFrameDestroyOpcode(); |
| 230 | } |
| 231 | |
| 232 | /// Returns true if the argument is a frame setup pseudo instruction. |
| 233 | bool isFrameSetup(const MachineInstr &I) const { |
| 234 | return I.getOpcode() == getCallFrameSetupOpcode(); |
| 235 | } |
| 236 | |
| 237 | /// Returns size of the frame associated with the given frame instruction. |
| 238 | /// For frame setup instruction this is frame that is set up space set up |
| 239 | /// after the instruction. For frame destroy instruction this is the frame |
| 240 | /// freed by the caller. |
| 241 | /// Note, in some cases a call frame (or a part of it) may be prepared prior |
| 242 | /// to the frame setup instruction. It occurs in the calls that involve |
| 243 | /// inalloca arguments. This function reports only the size of the frame part |
| 244 | /// that is set up between the frame setup and destroy pseudo instructions. |
| 245 | int64_t getFrameSize(const MachineInstr &I) const { |
| 246 | assert(isFrameInstr(I) && "Not a frame instruction" ); |
| 247 | assert(I.getOperand(0).getImm() >= 0); |
| 248 | return I.getOperand(i: 0).getImm(); |
| 249 | } |
| 250 | |
| 251 | /// Returns the total frame size, which is made up of the space set up inside |
| 252 | /// the pair of frame start-stop instructions and the space that is set up |
| 253 | /// prior to the pair. |
| 254 | int64_t getFrameTotalSize(const MachineInstr &I) const { |
| 255 | if (isFrameSetup(I)) { |
| 256 | assert(I.getOperand(1).getImm() >= 0 && |
| 257 | "Frame size must not be negative" ); |
| 258 | return getFrameSize(I) + I.getOperand(i: 1).getImm(); |
| 259 | } |
| 260 | return getFrameSize(I); |
| 261 | } |
| 262 | |
| 263 | unsigned getCatchReturnOpcode() const { return CatchRetOpcode; } |
| 264 | unsigned getReturnOpcode() const { return ReturnOpcode; } |
| 265 | |
| 266 | /// Returns the actual stack pointer adjustment made by an instruction |
| 267 | /// as part of a call sequence. By default, only call frame setup/destroy |
| 268 | /// instructions adjust the stack, but targets may want to override this |
| 269 | /// to enable more fine-grained adjustment, or adjust by a different value. |
| 270 | virtual int getSPAdjust(const MachineInstr &MI) const; |
| 271 | |
| 272 | /// Return true if the instruction is a "coalescable" extension instruction. |
| 273 | /// That is, it's like a copy where it's legal for the source to overlap the |
| 274 | /// destination. e.g. X86::MOVSX64rr32. If this returns true, then it's |
| 275 | /// expected the pre-extension value is available as a subreg of the result |
| 276 | /// register. This also returns the sub-register index in SubIdx. |
| 277 | virtual bool isCoalescableExtInstr(const MachineInstr &MI, Register &SrcReg, |
| 278 | Register &DstReg, unsigned &SubIdx) const { |
| 279 | return false; |
| 280 | } |
| 281 | |
| 282 | /// If the specified machine instruction is a direct |
| 283 | /// load from a stack slot, return the virtual or physical register number of |
| 284 | /// the destination along with the FrameIndex of the loaded stack slot. If |
| 285 | /// not, return 0. This predicate must return 0 if the instruction has |
| 286 | /// any side effects other than loading from the stack slot. |
| 287 | virtual Register isLoadFromStackSlot(const MachineInstr &MI, |
| 288 | int &FrameIndex) const { |
| 289 | return 0; |
| 290 | } |
| 291 | |
| 292 | /// Optional extension of isLoadFromStackSlot that returns the number of |
| 293 | /// bytes loaded from the stack. This must be implemented if a backend |
| 294 | /// supports partial stack slot spills/loads to further disambiguate |
| 295 | /// what the load does. |
| 296 | virtual Register isLoadFromStackSlot(const MachineInstr &MI, |
| 297 | int &FrameIndex, |
| 298 | TypeSize &MemBytes) const { |
| 299 | MemBytes = TypeSize::getZero(); |
| 300 | return isLoadFromStackSlot(MI, FrameIndex); |
| 301 | } |
| 302 | |
| 303 | /// Check for post-frame ptr elimination stack locations as well. |
| 304 | /// This uses a heuristic so it isn't reliable for correctness. |
| 305 | virtual Register isLoadFromStackSlotPostFE(const MachineInstr &MI, |
| 306 | int &FrameIndex) const { |
| 307 | return 0; |
| 308 | } |
| 309 | |
| 310 | /// If the specified machine instruction has a load from a stack slot, |
| 311 | /// return true along with the FrameIndices of the loaded stack slot and the |
| 312 | /// machine mem operands containing the reference. |
| 313 | /// If not, return false. Unlike isLoadFromStackSlot, this returns true for |
| 314 | /// any instructions that loads from the stack. This is just a hint, as some |
| 315 | /// cases may be missed. |
| 316 | virtual bool hasLoadFromStackSlot( |
| 317 | const MachineInstr &MI, |
| 318 | SmallVectorImpl<const MachineMemOperand *> &Accesses) const; |
| 319 | |
| 320 | /// If the specified machine instruction is a direct |
| 321 | /// store to a stack slot, return the virtual or physical register number of |
| 322 | /// the source reg along with the FrameIndex of the loaded stack slot. If |
| 323 | /// not, return 0. This predicate must return 0 if the instruction has |
| 324 | /// any side effects other than storing to the stack slot. |
| 325 | virtual Register isStoreToStackSlot(const MachineInstr &MI, |
| 326 | int &FrameIndex) const { |
| 327 | return 0; |
| 328 | } |
| 329 | |
| 330 | /// Optional extension of isStoreToStackSlot that returns the number of |
| 331 | /// bytes stored to the stack. This must be implemented if a backend |
| 332 | /// supports partial stack slot spills/loads to further disambiguate |
| 333 | /// what the store does. |
| 334 | virtual Register isStoreToStackSlot(const MachineInstr &MI, |
| 335 | int &FrameIndex, |
| 336 | TypeSize &MemBytes) const { |
| 337 | MemBytes = TypeSize::getZero(); |
| 338 | return isStoreToStackSlot(MI, FrameIndex); |
| 339 | } |
| 340 | |
| 341 | /// Check for post-frame ptr elimination stack locations as well. |
| 342 | /// This uses a heuristic, so it isn't reliable for correctness. |
| 343 | virtual Register isStoreToStackSlotPostFE(const MachineInstr &MI, |
| 344 | int &FrameIndex) const { |
| 345 | return 0; |
| 346 | } |
| 347 | |
| 348 | /// If the specified machine instruction has a store to a stack slot, |
| 349 | /// return true along with the FrameIndices of the loaded stack slot and the |
| 350 | /// machine mem operands containing the reference. |
| 351 | /// If not, return false. Unlike isStoreToStackSlot, |
| 352 | /// this returns true for any instructions that stores to the |
| 353 | /// stack. This is just a hint, as some cases may be missed. |
| 354 | virtual bool hasStoreToStackSlot( |
| 355 | const MachineInstr &MI, |
| 356 | SmallVectorImpl<const MachineMemOperand *> &Accesses) const; |
| 357 | |
| 358 | /// Return true if the specified machine instruction |
| 359 | /// is a copy of one stack slot to another and has no other effect. |
| 360 | /// Provide the identity of the two frame indices. |
| 361 | virtual bool isStackSlotCopy(const MachineInstr &MI, int &DestFrameIndex, |
| 362 | int &SrcFrameIndex) const { |
| 363 | return false; |
| 364 | } |
| 365 | |
| 366 | /// Compute the size in bytes and offset within a stack slot of a spilled |
| 367 | /// register or subregister. |
| 368 | /// |
| 369 | /// \param [out] Size in bytes of the spilled value. |
| 370 | /// \param [out] Offset in bytes within the stack slot. |
| 371 | /// \returns true if both Size and Offset are successfully computed. |
| 372 | /// |
| 373 | /// Not all subregisters have computable spill slots. For example, |
| 374 | /// subregisters registers may not be byte-sized, and a pair of discontiguous |
| 375 | /// subregisters has no single offset. |
| 376 | /// |
| 377 | /// Targets with nontrivial bigendian implementations may need to override |
| 378 | /// this, particularly to support spilled vector registers. |
| 379 | virtual bool getStackSlotRange(const TargetRegisterClass *RC, unsigned SubIdx, |
| 380 | unsigned &Size, unsigned &Offset, |
| 381 | const MachineFunction &MF) const; |
| 382 | |
| 383 | /// Return true if the given instruction is terminator that is unspillable, |
| 384 | /// according to isUnspillableTerminatorImpl. |
| 385 | bool isUnspillableTerminator(const MachineInstr *MI) const { |
| 386 | return MI->isTerminator() && isUnspillableTerminatorImpl(MI); |
| 387 | } |
| 388 | |
| 389 | /// Returns the size in bytes of the specified MachineInstr, or ~0U |
| 390 | /// when this function is not implemented by a target. |
| 391 | virtual unsigned getInstSizeInBytes(const MachineInstr &MI) const { |
| 392 | return ~0U; |
| 393 | } |
| 394 | |
| 395 | /// Return true if the instruction is as cheap as a move instruction. |
| 396 | /// |
| 397 | /// Targets for different archs need to override this, and different |
| 398 | /// micro-architectures can also be finely tuned inside. |
| 399 | virtual bool isAsCheapAsAMove(const MachineInstr &MI) const { |
| 400 | return MI.isAsCheapAsAMove(); |
| 401 | } |
| 402 | |
| 403 | /// Return true if the instruction should be sunk by MachineSink. |
| 404 | /// |
| 405 | /// MachineSink determines on its own whether the instruction is safe to sink; |
| 406 | /// this gives the target a hook to override the default behavior with regards |
| 407 | /// to which instructions should be sunk. |
| 408 | virtual bool shouldSink(const MachineInstr &MI) const { return true; } |
| 409 | |
| 410 | /// Return false if the instruction should not be hoisted by MachineLICM. |
| 411 | /// |
| 412 | /// MachineLICM determines on its own whether the instruction is safe to |
| 413 | /// hoist; this gives the target a hook to extend this assessment and prevent |
| 414 | /// an instruction being hoisted from a given loop for target specific |
| 415 | /// reasons. |
| 416 | virtual bool shouldHoist(const MachineInstr &MI, |
| 417 | const MachineLoop *FromLoop) const { |
| 418 | return true; |
| 419 | } |
| 420 | |
| 421 | /// Re-issue the specified 'original' instruction at the |
| 422 | /// specific location targeting a new destination register. |
| 423 | /// The register in Orig->getOperand(0).getReg() will be substituted by |
| 424 | /// DestReg:SubIdx. Any existing subreg index is preserved or composed with |
| 425 | /// SubIdx. |
| 426 | virtual void reMaterialize(MachineBasicBlock &MBB, |
| 427 | MachineBasicBlock::iterator MI, Register DestReg, |
| 428 | unsigned SubIdx, const MachineInstr &Orig, |
| 429 | const TargetRegisterInfo &TRI) const; |
| 430 | |
| 431 | /// Clones instruction or the whole instruction bundle \p Orig and |
| 432 | /// insert into \p MBB before \p InsertBefore. The target may update operands |
| 433 | /// that are required to be unique. |
| 434 | /// |
| 435 | /// \p Orig must not return true for MachineInstr::isNotDuplicable(). |
| 436 | virtual MachineInstr &duplicate(MachineBasicBlock &MBB, |
| 437 | MachineBasicBlock::iterator InsertBefore, |
| 438 | const MachineInstr &Orig) const; |
| 439 | |
| 440 | /// This method must be implemented by targets that |
| 441 | /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target |
| 442 | /// may be able to convert a two-address instruction into one or more true |
| 443 | /// three-address instructions on demand. This allows the X86 target (for |
| 444 | /// example) to convert ADD and SHL instructions into LEA instructions if they |
| 445 | /// would require register copies due to two-addressness. |
| 446 | /// |
| 447 | /// This method returns a null pointer if the transformation cannot be |
| 448 | /// performed, otherwise it returns the last new instruction. |
| 449 | /// |
| 450 | /// If \p LIS is not nullptr, the LiveIntervals info should be updated for |
| 451 | /// replacing \p MI with new instructions, even though this function does not |
| 452 | /// remove MI. |
| 453 | virtual MachineInstr *convertToThreeAddress(MachineInstr &MI, |
| 454 | LiveVariables *LV, |
| 455 | LiveIntervals *LIS) const { |
| 456 | return nullptr; |
| 457 | } |
| 458 | |
| 459 | // This constant can be used as an input value of operand index passed to |
| 460 | // the method findCommutedOpIndices() to tell the method that the |
| 461 | // corresponding operand index is not pre-defined and that the method |
| 462 | // can pick any commutable operand. |
| 463 | static const unsigned CommuteAnyOperandIndex = ~0U; |
| 464 | |
| 465 | /// This method commutes the operands of the given machine instruction MI. |
| 466 | /// |
| 467 | /// The operands to be commuted are specified by their indices OpIdx1 and |
| 468 | /// OpIdx2. OpIdx1 and OpIdx2 arguments may be set to a special value |
| 469 | /// 'CommuteAnyOperandIndex', which means that the method is free to choose |
| 470 | /// any arbitrarily chosen commutable operand. If both arguments are set to |
| 471 | /// 'CommuteAnyOperandIndex' then the method looks for 2 different commutable |
| 472 | /// operands; then commutes them if such operands could be found. |
| 473 | /// |
| 474 | /// If NewMI is false, MI is modified in place and returned; otherwise, a |
| 475 | /// new machine instruction is created and returned. |
| 476 | /// |
| 477 | /// Do not call this method for a non-commutable instruction or |
| 478 | /// for non-commuable operands. |
| 479 | /// Even though the instruction is commutable, the method may still |
| 480 | /// fail to commute the operands, null pointer is returned in such cases. |
| 481 | MachineInstr * |
| 482 | commuteInstruction(MachineInstr &MI, bool NewMI = false, |
| 483 | unsigned OpIdx1 = CommuteAnyOperandIndex, |
| 484 | unsigned OpIdx2 = CommuteAnyOperandIndex) const; |
| 485 | |
| 486 | /// Returns true iff the routine could find two commutable operands in the |
| 487 | /// given machine instruction. |
| 488 | /// The 'SrcOpIdx1' and 'SrcOpIdx2' are INPUT and OUTPUT arguments. |
| 489 | /// If any of the INPUT values is set to the special value |
| 490 | /// 'CommuteAnyOperandIndex' then the method arbitrarily picks a commutable |
| 491 | /// operand, then returns its index in the corresponding argument. |
| 492 | /// If both of INPUT values are set to 'CommuteAnyOperandIndex' then method |
| 493 | /// looks for 2 commutable operands. |
| 494 | /// If INPUT values refer to some operands of MI, then the method simply |
| 495 | /// returns true if the corresponding operands are commutable and returns |
| 496 | /// false otherwise. |
| 497 | /// |
| 498 | /// For example, calling this method this way: |
| 499 | /// unsigned Op1 = 1, Op2 = CommuteAnyOperandIndex; |
| 500 | /// findCommutedOpIndices(MI, Op1, Op2); |
| 501 | /// can be interpreted as a query asking to find an operand that would be |
| 502 | /// commutable with the operand#1. |
| 503 | virtual bool findCommutedOpIndices(const MachineInstr &MI, |
| 504 | unsigned &SrcOpIdx1, |
| 505 | unsigned &SrcOpIdx2) const; |
| 506 | |
| 507 | /// Returns true if the target has a preference on the operands order of |
| 508 | /// the given machine instruction. And specify if \p Commute is required to |
| 509 | /// get the desired operands order. |
| 510 | virtual bool hasCommutePreference(MachineInstr &MI, bool &Commute) const { |
| 511 | return false; |
| 512 | } |
| 513 | |
| 514 | /// If possible, converts the instruction to a simplified/canonical form. |
| 515 | /// Returns true if the instruction was modified. |
| 516 | /// |
| 517 | /// This function is only called after register allocation. The MI will be |
| 518 | /// modified in place. This is called by passes such as |
| 519 | /// MachineCopyPropagation, where their mutation of the MI operands may |
| 520 | /// expose opportunities to convert the instruction to a simpler form (e.g. |
| 521 | /// a load of 0). |
| 522 | virtual bool simplifyInstruction(MachineInstr &MI) const { return false; } |
| 523 | |
| 524 | /// A pair composed of a register and a sub-register index. |
| 525 | /// Used to give some type checking when modeling Reg:SubReg. |
| 526 | struct RegSubRegPair { |
| 527 | Register Reg; |
| 528 | unsigned SubReg; |
| 529 | |
| 530 | RegSubRegPair(Register Reg = Register(), unsigned SubReg = 0) |
| 531 | : Reg(Reg), SubReg(SubReg) {} |
| 532 | |
| 533 | bool operator==(const RegSubRegPair& P) const { |
| 534 | return Reg == P.Reg && SubReg == P.SubReg; |
| 535 | } |
| 536 | bool operator!=(const RegSubRegPair& P) const { |
| 537 | return !(*this == P); |
| 538 | } |
| 539 | }; |
| 540 | |
| 541 | /// A pair composed of a pair of a register and a sub-register index, |
| 542 | /// and another sub-register index. |
| 543 | /// Used to give some type checking when modeling Reg:SubReg1, SubReg2. |
| 544 | struct RegSubRegPairAndIdx : RegSubRegPair { |
| 545 | unsigned SubIdx; |
| 546 | |
| 547 | RegSubRegPairAndIdx(Register Reg = Register(), unsigned SubReg = 0, |
| 548 | unsigned SubIdx = 0) |
| 549 | : RegSubRegPair(Reg, SubReg), SubIdx(SubIdx) {} |
| 550 | }; |
| 551 | |
| 552 | /// Build the equivalent inputs of a REG_SEQUENCE for the given \p MI |
| 553 | /// and \p DefIdx. |
| 554 | /// \p [out] InputRegs of the equivalent REG_SEQUENCE. Each element of |
| 555 | /// the list is modeled as <Reg:SubReg, SubIdx>. Operands with the undef |
| 556 | /// flag are not added to this list. |
| 557 | /// E.g., REG_SEQUENCE %1:sub1, sub0, %2, sub1 would produce |
| 558 | /// two elements: |
| 559 | /// - %1:sub1, sub0 |
| 560 | /// - %2<:0>, sub1 |
| 561 | /// |
| 562 | /// \returns true if it is possible to build such an input sequence |
| 563 | /// with the pair \p MI, \p DefIdx. False otherwise. |
| 564 | /// |
| 565 | /// \pre MI.isRegSequence() or MI.isRegSequenceLike(). |
| 566 | /// |
| 567 | /// \note The generic implementation does not provide any support for |
| 568 | /// MI.isRegSequenceLike(). In other words, one has to override |
| 569 | /// getRegSequenceLikeInputs for target specific instructions. |
| 570 | bool |
| 571 | getRegSequenceInputs(const MachineInstr &MI, unsigned DefIdx, |
| 572 | SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const; |
| 573 | |
| 574 | /// Build the equivalent inputs of a EXTRACT_SUBREG for the given \p MI |
| 575 | /// and \p DefIdx. |
| 576 | /// \p [out] InputReg of the equivalent EXTRACT_SUBREG. |
| 577 | /// E.g., EXTRACT_SUBREG %1:sub1, sub0, sub1 would produce: |
| 578 | /// - %1:sub1, sub0 |
| 579 | /// |
| 580 | /// \returns true if it is possible to build such an input sequence |
| 581 | /// with the pair \p MI, \p DefIdx and the operand has no undef flag set. |
| 582 | /// False otherwise. |
| 583 | /// |
| 584 | /// \pre MI.isExtractSubreg() or MI.isExtractSubregLike(). |
| 585 | /// |
| 586 | /// \note The generic implementation does not provide any support for |
| 587 | /// MI.isExtractSubregLike(). In other words, one has to override |
| 588 | /// getExtractSubregLikeInputs for target specific instructions. |
| 589 | bool getExtractSubregInputs(const MachineInstr &MI, unsigned DefIdx, |
| 590 | RegSubRegPairAndIdx &InputReg) const; |
| 591 | |
| 592 | /// Build the equivalent inputs of a INSERT_SUBREG for the given \p MI |
| 593 | /// and \p DefIdx. |
| 594 | /// \p [out] BaseReg and \p [out] InsertedReg contain |
| 595 | /// the equivalent inputs of INSERT_SUBREG. |
| 596 | /// E.g., INSERT_SUBREG %0:sub0, %1:sub1, sub3 would produce: |
| 597 | /// - BaseReg: %0:sub0 |
| 598 | /// - InsertedReg: %1:sub1, sub3 |
| 599 | /// |
| 600 | /// \returns true if it is possible to build such an input sequence |
| 601 | /// with the pair \p MI, \p DefIdx and the operand has no undef flag set. |
| 602 | /// False otherwise. |
| 603 | /// |
| 604 | /// \pre MI.isInsertSubreg() or MI.isInsertSubregLike(). |
| 605 | /// |
| 606 | /// \note The generic implementation does not provide any support for |
| 607 | /// MI.isInsertSubregLike(). In other words, one has to override |
| 608 | /// getInsertSubregLikeInputs for target specific instructions. |
| 609 | bool getInsertSubregInputs(const MachineInstr &MI, unsigned DefIdx, |
| 610 | RegSubRegPair &BaseReg, |
| 611 | RegSubRegPairAndIdx &InsertedReg) const; |
| 612 | |
| 613 | /// Return true if two machine instructions would produce identical values. |
| 614 | /// By default, this is only true when the two instructions |
| 615 | /// are deemed identical except for defs. If this function is called when the |
| 616 | /// IR is still in SSA form, the caller can pass the MachineRegisterInfo for |
| 617 | /// aggressive checks. |
| 618 | virtual bool produceSameValue(const MachineInstr &MI0, |
| 619 | const MachineInstr &MI1, |
| 620 | const MachineRegisterInfo *MRI = nullptr) const; |
| 621 | |
| 622 | /// \returns true if a branch from an instruction with opcode \p BranchOpc |
| 623 | /// bytes is capable of jumping to a position \p BrOffset bytes away. |
| 624 | virtual bool isBranchOffsetInRange(unsigned BranchOpc, |
| 625 | int64_t BrOffset) const { |
| 626 | llvm_unreachable("target did not implement" ); |
| 627 | } |
| 628 | |
| 629 | /// \returns The block that branch instruction \p MI jumps to. |
| 630 | virtual MachineBasicBlock *getBranchDestBlock(const MachineInstr &MI) const { |
| 631 | llvm_unreachable("target did not implement" ); |
| 632 | } |
| 633 | |
| 634 | /// Insert an unconditional indirect branch at the end of \p MBB to \p |
| 635 | /// NewDestBB. Optionally, insert the clobbered register restoring in \p |
| 636 | /// RestoreBB. \p BrOffset indicates the offset of \p NewDestBB relative to |
| 637 | /// the offset of the position to insert the new branch. |
| 638 | virtual void insertIndirectBranch(MachineBasicBlock &MBB, |
| 639 | MachineBasicBlock &NewDestBB, |
| 640 | MachineBasicBlock &RestoreBB, |
| 641 | const DebugLoc &DL, int64_t BrOffset = 0, |
| 642 | RegScavenger *RS = nullptr) const { |
| 643 | llvm_unreachable("target did not implement" ); |
| 644 | } |
| 645 | |
| 646 | /// Analyze the branching code at the end of MBB, returning |
| 647 | /// true if it cannot be understood (e.g. it's a switch dispatch or isn't |
| 648 | /// implemented for a target). Upon success, this returns false and returns |
| 649 | /// with the following information in various cases: |
| 650 | /// |
| 651 | /// 1. If this block ends with no branches (it just falls through to its succ) |
| 652 | /// just return false, leaving TBB/FBB null. |
| 653 | /// 2. If this block ends with only an unconditional branch, it sets TBB to be |
| 654 | /// the destination block. |
| 655 | /// 3. If this block ends with a conditional branch and it falls through to a |
| 656 | /// successor block, it sets TBB to be the branch destination block and a |
| 657 | /// list of operands that evaluate the condition. These operands can be |
| 658 | /// passed to other TargetInstrInfo methods to create new branches. |
| 659 | /// 4. If this block ends with a conditional branch followed by an |
| 660 | /// unconditional branch, it returns the 'true' destination in TBB, the |
| 661 | /// 'false' destination in FBB, and a list of operands that evaluate the |
| 662 | /// condition. These operands can be passed to other TargetInstrInfo |
| 663 | /// methods to create new branches. |
| 664 | /// |
| 665 | /// Note that removeBranch and insertBranch must be implemented to support |
| 666 | /// cases where this method returns success. |
| 667 | /// |
| 668 | /// If AllowModify is true, then this routine is allowed to modify the basic |
| 669 | /// block (e.g. delete instructions after the unconditional branch). |
| 670 | /// |
| 671 | /// The CFG information in MBB.Predecessors and MBB.Successors must be valid |
| 672 | /// before calling this function. |
| 673 | virtual bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, |
| 674 | MachineBasicBlock *&FBB, |
| 675 | SmallVectorImpl<MachineOperand> &Cond, |
| 676 | bool AllowModify = false) const { |
| 677 | return true; |
| 678 | } |
| 679 | |
| 680 | /// Represents a predicate at the MachineFunction level. The control flow a |
| 681 | /// MachineBranchPredicate represents is: |
| 682 | /// |
| 683 | /// Reg = LHS `Predicate` RHS == ConditionDef |
| 684 | /// if Reg then goto TrueDest else goto FalseDest |
| 685 | /// |
| 686 | struct MachineBranchPredicate { |
| 687 | enum ComparePredicate { |
| 688 | PRED_EQ, // True if two values are equal |
| 689 | PRED_NE, // True if two values are not equal |
| 690 | PRED_INVALID // Sentinel value |
| 691 | }; |
| 692 | |
| 693 | ComparePredicate Predicate = PRED_INVALID; |
| 694 | MachineOperand LHS = MachineOperand::CreateImm(Val: 0); |
| 695 | MachineOperand RHS = MachineOperand::CreateImm(Val: 0); |
| 696 | MachineBasicBlock *TrueDest = nullptr; |
| 697 | MachineBasicBlock *FalseDest = nullptr; |
| 698 | MachineInstr *ConditionDef = nullptr; |
| 699 | |
| 700 | /// SingleUseCondition is true if ConditionDef is dead except for the |
| 701 | /// branch(es) at the end of the basic block. |
| 702 | /// |
| 703 | bool SingleUseCondition = false; |
| 704 | |
| 705 | explicit MachineBranchPredicate() = default; |
| 706 | }; |
| 707 | |
| 708 | /// Analyze the branching code at the end of MBB and parse it into the |
| 709 | /// MachineBranchPredicate structure if possible. Returns false on success |
| 710 | /// and true on failure. |
| 711 | /// |
| 712 | /// If AllowModify is true, then this routine is allowed to modify the basic |
| 713 | /// block (e.g. delete instructions after the unconditional branch). |
| 714 | /// |
| 715 | virtual bool analyzeBranchPredicate(MachineBasicBlock &MBB, |
| 716 | MachineBranchPredicate &MBP, |
| 717 | bool AllowModify = false) const { |
| 718 | return true; |
| 719 | } |
| 720 | |
| 721 | /// Remove the branching code at the end of the specific MBB. |
| 722 | /// This is only invoked in cases where analyzeBranch returns success. It |
| 723 | /// returns the number of instructions that were removed. |
| 724 | /// If \p BytesRemoved is non-null, report the change in code size from the |
| 725 | /// removed instructions. |
| 726 | virtual unsigned removeBranch(MachineBasicBlock &MBB, |
| 727 | int *BytesRemoved = nullptr) const { |
| 728 | llvm_unreachable("Target didn't implement TargetInstrInfo::removeBranch!" ); |
| 729 | } |
| 730 | |
| 731 | /// Insert branch code into the end of the specified MachineBasicBlock. The |
| 732 | /// operands to this method are the same as those returned by analyzeBranch. |
| 733 | /// This is only invoked in cases where analyzeBranch returns success. It |
| 734 | /// returns the number of instructions inserted. If \p BytesAdded is non-null, |
| 735 | /// report the change in code size from the added instructions. |
| 736 | /// |
| 737 | /// It is also invoked by tail merging to add unconditional branches in |
| 738 | /// cases where analyzeBranch doesn't apply because there was no original |
| 739 | /// branch to analyze. At least this much must be implemented, else tail |
| 740 | /// merging needs to be disabled. |
| 741 | /// |
| 742 | /// The CFG information in MBB.Predecessors and MBB.Successors must be valid |
| 743 | /// before calling this function. |
| 744 | virtual unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, |
| 745 | MachineBasicBlock *FBB, |
| 746 | ArrayRef<MachineOperand> Cond, |
| 747 | const DebugLoc &DL, |
| 748 | int *BytesAdded = nullptr) const { |
| 749 | llvm_unreachable("Target didn't implement TargetInstrInfo::insertBranch!" ); |
| 750 | } |
| 751 | |
| 752 | unsigned insertUnconditionalBranch(MachineBasicBlock &MBB, |
| 753 | MachineBasicBlock *DestBB, |
| 754 | const DebugLoc &DL, |
| 755 | int *BytesAdded = nullptr) const { |
| 756 | return insertBranch(MBB, TBB: DestBB, FBB: nullptr, Cond: ArrayRef<MachineOperand>(), DL, |
| 757 | BytesAdded); |
| 758 | } |
| 759 | |
| 760 | /// Object returned by analyzeLoopForPipelining. Allows software pipelining |
| 761 | /// implementations to query attributes of the loop being pipelined and to |
| 762 | /// apply target-specific updates to the loop once pipelining is complete. |
| 763 | class LLVM_ABI PipelinerLoopInfo { |
| 764 | public: |
| 765 | virtual ~PipelinerLoopInfo(); |
| 766 | /// Return true if the given instruction should not be pipelined and should |
| 767 | /// be ignored. An example could be a loop comparison, or induction variable |
| 768 | /// update with no users being pipelined. |
| 769 | virtual bool shouldIgnoreForPipelining(const MachineInstr *MI) const = 0; |
| 770 | |
| 771 | /// Return true if the proposed schedule should used. Otherwise return |
| 772 | /// false to not pipeline the loop. This function should be used to ensure |
| 773 | /// that pipelined loops meet target-specific quality heuristics. |
| 774 | virtual bool shouldUseSchedule(SwingSchedulerDAG &SSD, SMSchedule &SMS) { |
| 775 | return true; |
| 776 | } |
| 777 | |
| 778 | /// Create a condition to determine if the trip count of the loop is greater |
| 779 | /// than TC, where TC is always one more than for the previous prologue or |
| 780 | /// 0 if this is being called for the outermost prologue. |
| 781 | /// |
| 782 | /// If the trip count is statically known to be greater than TC, return |
| 783 | /// true. If the trip count is statically known to be not greater than TC, |
| 784 | /// return false. Otherwise return nullopt and fill out Cond with the test |
| 785 | /// condition. |
| 786 | /// |
| 787 | /// Note: This hook is guaranteed to be called from the innermost to the |
| 788 | /// outermost prologue of the loop being software pipelined. |
| 789 | virtual std::optional<bool> |
| 790 | createTripCountGreaterCondition(int TC, MachineBasicBlock &MBB, |
| 791 | SmallVectorImpl<MachineOperand> &Cond) = 0; |
| 792 | |
| 793 | /// Create a condition to determine if the remaining trip count for a phase |
| 794 | /// is greater than TC. Some instructions such as comparisons may be |
| 795 | /// inserted at the bottom of MBB. All instructions expanded for the |
| 796 | /// phase must be inserted in MBB before calling this function. |
| 797 | /// LastStage0Insts is the map from the original instructions scheduled at |
| 798 | /// stage#0 to the expanded instructions for the last iteration of the |
| 799 | /// kernel. LastStage0Insts is intended to obtain the instruction that |
| 800 | /// refers the latest loop counter value. |
| 801 | /// |
| 802 | /// MBB can also be a predecessor of the prologue block. Then |
| 803 | /// LastStage0Insts must be empty and the compared value is the initial |
| 804 | /// value of the trip count. |
| 805 | virtual void createRemainingIterationsGreaterCondition( |
| 806 | int TC, MachineBasicBlock &MBB, SmallVectorImpl<MachineOperand> &Cond, |
| 807 | DenseMap<MachineInstr *, MachineInstr *> &LastStage0Insts) { |
| 808 | llvm_unreachable( |
| 809 | "Target didn't implement " |
| 810 | "PipelinerLoopInfo::createRemainingIterationsGreaterCondition!" ); |
| 811 | } |
| 812 | |
| 813 | /// Modify the loop such that the trip count is |
| 814 | /// OriginalTC + TripCountAdjust. |
| 815 | virtual void adjustTripCount(int TripCountAdjust) = 0; |
| 816 | |
| 817 | /// Called when the loop's preheader has been modified to NewPreheader. |
| 818 | virtual void (MachineBasicBlock *) = 0; |
| 819 | |
| 820 | /// Called when the loop is being removed. Any instructions in the preheader |
| 821 | /// should be removed. |
| 822 | /// |
| 823 | /// Once this function is called, no other functions on this object are |
| 824 | /// valid; the loop has been removed. |
| 825 | virtual void disposed(LiveIntervals *LIS = nullptr) {} |
| 826 | |
| 827 | /// Return true if the target can expand pipelined schedule with modulo |
| 828 | /// variable expansion. |
| 829 | virtual bool isMVEExpanderSupported() { return false; } |
| 830 | }; |
| 831 | |
| 832 | /// Analyze loop L, which must be a single-basic-block loop, and if the |
| 833 | /// conditions can be understood enough produce a PipelinerLoopInfo object. |
| 834 | virtual std::unique_ptr<PipelinerLoopInfo> |
| 835 | analyzeLoopForPipelining(MachineBasicBlock *LoopBB) const { |
| 836 | return nullptr; |
| 837 | } |
| 838 | |
| 839 | /// Analyze the loop code, return true if it cannot be understood. Upon |
| 840 | /// success, this function returns false and returns information about the |
| 841 | /// induction variable and compare instruction used at the end. |
| 842 | virtual bool analyzeLoop(MachineLoop &L, MachineInstr *&IndVarInst, |
| 843 | MachineInstr *&CmpInst) const { |
| 844 | return true; |
| 845 | } |
| 846 | |
| 847 | /// Generate code to reduce the loop iteration by one and check if the loop |
| 848 | /// is finished. Return the value/register of the new loop count. We need |
| 849 | /// this function when peeling off one or more iterations of a loop. This |
| 850 | /// function assumes the nth iteration is peeled first. |
| 851 | virtual unsigned reduceLoopCount(MachineBasicBlock &MBB, |
| 852 | MachineBasicBlock &, |
| 853 | MachineInstr *IndVar, MachineInstr &Cmp, |
| 854 | SmallVectorImpl<MachineOperand> &Cond, |
| 855 | SmallVectorImpl<MachineInstr *> &PrevInsts, |
| 856 | unsigned Iter, unsigned MaxIter) const { |
| 857 | llvm_unreachable("Target didn't implement ReduceLoopCount" ); |
| 858 | } |
| 859 | |
| 860 | /// Delete the instruction OldInst and everything after it, replacing it with |
| 861 | /// an unconditional branch to NewDest. This is used by the tail merging pass. |
| 862 | virtual void ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail, |
| 863 | MachineBasicBlock *NewDest) const; |
| 864 | |
| 865 | /// Return true if it's legal to split the given basic |
| 866 | /// block at the specified instruction (i.e. instruction would be the start |
| 867 | /// of a new basic block). |
| 868 | virtual bool isLegalToSplitMBBAt(MachineBasicBlock &MBB, |
| 869 | MachineBasicBlock::iterator MBBI) const { |
| 870 | return true; |
| 871 | } |
| 872 | |
| 873 | /// Return true if it's profitable to predicate |
| 874 | /// instructions with accumulated instruction latency of "NumCycles" |
| 875 | /// of the specified basic block, where the probability of the instructions |
| 876 | /// being executed is given by Probability, and Confidence is a measure |
| 877 | /// of our confidence that it will be properly predicted. |
| 878 | virtual bool isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumCycles, |
| 879 | unsigned , |
| 880 | BranchProbability Probability) const { |
| 881 | return false; |
| 882 | } |
| 883 | |
| 884 | /// Second variant of isProfitableToIfCvt. This one |
| 885 | /// checks for the case where two basic blocks from true and false path |
| 886 | /// of a if-then-else (diamond) are predicated on mutually exclusive |
| 887 | /// predicates, where the probability of the true path being taken is given |
| 888 | /// by Probability, and Confidence is a measure of our confidence that it |
| 889 | /// will be properly predicted. |
| 890 | virtual bool isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumTCycles, |
| 891 | unsigned , |
| 892 | MachineBasicBlock &FMBB, unsigned NumFCycles, |
| 893 | unsigned , |
| 894 | BranchProbability Probability) const { |
| 895 | return false; |
| 896 | } |
| 897 | |
| 898 | /// Return true if it's profitable for if-converter to duplicate instructions |
| 899 | /// of specified accumulated instruction latencies in the specified MBB to |
| 900 | /// enable if-conversion. |
| 901 | /// The probability of the instructions being executed is given by |
| 902 | /// Probability, and Confidence is a measure of our confidence that it |
| 903 | /// will be properly predicted. |
| 904 | virtual bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB, |
| 905 | unsigned NumCycles, |
| 906 | BranchProbability Probability) const { |
| 907 | return false; |
| 908 | } |
| 909 | |
| 910 | /// Return the increase in code size needed to predicate a contiguous run of |
| 911 | /// NumInsts instructions. |
| 912 | virtual unsigned (const MachineFunction &MF, |
| 913 | unsigned NumInsts) const { |
| 914 | return 0; |
| 915 | } |
| 916 | |
| 917 | /// Return an estimate for the code size reduction (in bytes) which will be |
| 918 | /// caused by removing the given branch instruction during if-conversion. |
| 919 | virtual unsigned predictBranchSizeForIfCvt(MachineInstr &MI) const { |
| 920 | return getInstSizeInBytes(MI); |
| 921 | } |
| 922 | |
| 923 | /// Return true if it's profitable to unpredicate |
| 924 | /// one side of a 'diamond', i.e. two sides of if-else predicated on mutually |
| 925 | /// exclusive predicates. |
| 926 | /// e.g. |
| 927 | /// subeq r0, r1, #1 |
| 928 | /// addne r0, r1, #1 |
| 929 | /// => |
| 930 | /// sub r0, r1, #1 |
| 931 | /// addne r0, r1, #1 |
| 932 | /// |
| 933 | /// This may be profitable is conditional instructions are always executed. |
| 934 | virtual bool isProfitableToUnpredicate(MachineBasicBlock &TMBB, |
| 935 | MachineBasicBlock &FMBB) const { |
| 936 | return false; |
| 937 | } |
| 938 | |
| 939 | /// Return true if it is possible to insert a select |
| 940 | /// instruction that chooses between TrueReg and FalseReg based on the |
| 941 | /// condition code in Cond. |
| 942 | /// |
| 943 | /// When successful, also return the latency in cycles from TrueReg, |
| 944 | /// FalseReg, and Cond to the destination register. In most cases, a select |
| 945 | /// instruction will be 1 cycle, so CondCycles = TrueCycles = FalseCycles = 1 |
| 946 | /// |
| 947 | /// Some x86 implementations have 2-cycle cmov instructions. |
| 948 | /// |
| 949 | /// @param MBB Block where select instruction would be inserted. |
| 950 | /// @param Cond Condition returned by analyzeBranch. |
| 951 | /// @param DstReg Virtual dest register that the result should write to. |
| 952 | /// @param TrueReg Virtual register to select when Cond is true. |
| 953 | /// @param FalseReg Virtual register to select when Cond is false. |
| 954 | /// @param CondCycles Latency from Cond+Branch to select output. |
| 955 | /// @param TrueCycles Latency from TrueReg to select output. |
| 956 | /// @param FalseCycles Latency from FalseReg to select output. |
| 957 | virtual bool canInsertSelect(const MachineBasicBlock &MBB, |
| 958 | ArrayRef<MachineOperand> Cond, Register DstReg, |
| 959 | Register TrueReg, Register FalseReg, |
| 960 | int &CondCycles, int &TrueCycles, |
| 961 | int &FalseCycles) const { |
| 962 | return false; |
| 963 | } |
| 964 | |
| 965 | /// Insert a select instruction into MBB before I that will copy TrueReg to |
| 966 | /// DstReg when Cond is true, and FalseReg to DstReg when Cond is false. |
| 967 | /// |
| 968 | /// This function can only be called after canInsertSelect() returned true. |
| 969 | /// The condition in Cond comes from analyzeBranch, and it can be assumed |
| 970 | /// that the same flags or registers required by Cond are available at the |
| 971 | /// insertion point. |
| 972 | /// |
| 973 | /// @param MBB Block where select instruction should be inserted. |
| 974 | /// @param I Insertion point. |
| 975 | /// @param DL Source location for debugging. |
| 976 | /// @param DstReg Virtual register to be defined by select instruction. |
| 977 | /// @param Cond Condition as computed by analyzeBranch. |
| 978 | /// @param TrueReg Virtual register to copy when Cond is true. |
| 979 | /// @param FalseReg Virtual register to copy when Cons is false. |
| 980 | virtual void insertSelect(MachineBasicBlock &MBB, |
| 981 | MachineBasicBlock::iterator I, const DebugLoc &DL, |
| 982 | Register DstReg, ArrayRef<MachineOperand> Cond, |
| 983 | Register TrueReg, Register FalseReg) const { |
| 984 | llvm_unreachable("Target didn't implement TargetInstrInfo::insertSelect!" ); |
| 985 | } |
| 986 | |
| 987 | /// Analyze the given select instruction, returning true if |
| 988 | /// it cannot be understood. It is assumed that MI->isSelect() is true. |
| 989 | /// |
| 990 | /// When successful, return the controlling condition and the operands that |
| 991 | /// determine the true and false result values. |
| 992 | /// |
| 993 | /// Result = SELECT Cond, TrueOp, FalseOp |
| 994 | /// |
| 995 | /// Some targets can optimize select instructions, for example by predicating |
| 996 | /// the instruction defining one of the operands. Such targets should set |
| 997 | /// Optimizable. |
| 998 | /// |
| 999 | /// @param MI Select instruction to analyze. |
| 1000 | /// @param Cond Condition controlling the select. |
| 1001 | /// @param TrueOp Operand number of the value selected when Cond is true. |
| 1002 | /// @param FalseOp Operand number of the value selected when Cond is false. |
| 1003 | /// @param Optimizable Returned as true if MI is optimizable. |
| 1004 | /// @returns False on success. |
| 1005 | virtual bool analyzeSelect(const MachineInstr &MI, |
| 1006 | SmallVectorImpl<MachineOperand> &Cond, |
| 1007 | unsigned &TrueOp, unsigned &FalseOp, |
| 1008 | bool &Optimizable) const { |
| 1009 | assert(MI.getDesc().isSelect() && "MI must be a select instruction" ); |
| 1010 | return true; |
| 1011 | } |
| 1012 | |
| 1013 | /// Given a select instruction that was understood by |
| 1014 | /// analyzeSelect and returned Optimizable = true, attempt to optimize MI by |
| 1015 | /// merging it with one of its operands. Returns NULL on failure. |
| 1016 | /// |
| 1017 | /// When successful, returns the new select instruction. The client is |
| 1018 | /// responsible for deleting MI. |
| 1019 | /// |
| 1020 | /// If both sides of the select can be optimized, PreferFalse is used to pick |
| 1021 | /// a side. |
| 1022 | /// |
| 1023 | /// @param MI Optimizable select instruction. |
| 1024 | /// @param NewMIs Set that record all MIs in the basic block up to \p |
| 1025 | /// MI. Has to be updated with any newly created MI or deleted ones. |
| 1026 | /// @param PreferFalse Try to optimize FalseOp instead of TrueOp. |
| 1027 | /// @returns Optimized instruction or NULL. |
| 1028 | virtual MachineInstr *optimizeSelect(MachineInstr &MI, |
| 1029 | SmallPtrSetImpl<MachineInstr *> &NewMIs, |
| 1030 | bool PreferFalse = false) const { |
| 1031 | // This function must be implemented if Optimizable is ever set. |
| 1032 | llvm_unreachable("Target must implement TargetInstrInfo::optimizeSelect!" ); |
| 1033 | } |
| 1034 | |
| 1035 | /// Emit instructions to copy a pair of physical registers. |
| 1036 | /// |
| 1037 | /// This function should support copies within any legal register class as |
| 1038 | /// well as any cross-class copies created during instruction selection. |
| 1039 | /// |
| 1040 | /// The source and destination registers may overlap, which may require a |
| 1041 | /// careful implementation when multiple copy instructions are required for |
| 1042 | /// large registers. See for example the ARM target. |
| 1043 | /// |
| 1044 | /// If RenamableDest is true, the copy instruction's destination operand is |
| 1045 | /// marked renamable. |
| 1046 | /// If RenamableSrc is true, the copy instruction's source operand is |
| 1047 | /// marked renamable. |
| 1048 | virtual void copyPhysReg(MachineBasicBlock &MBB, |
| 1049 | MachineBasicBlock::iterator MI, const DebugLoc &DL, |
| 1050 | Register DestReg, Register SrcReg, bool KillSrc, |
| 1051 | bool RenamableDest = false, |
| 1052 | bool RenamableSrc = false) const { |
| 1053 | llvm_unreachable("Target didn't implement TargetInstrInfo::copyPhysReg!" ); |
| 1054 | } |
| 1055 | |
| 1056 | /// Allow targets to tell MachineVerifier whether a specific register |
| 1057 | /// MachineOperand can be used as part of PC-relative addressing. |
| 1058 | /// PC-relative addressing modes in many CISC architectures contain |
| 1059 | /// (non-PC) registers as offsets or scaling values, which inherently |
| 1060 | /// tags the corresponding MachineOperand with OPERAND_PCREL. |
| 1061 | /// |
| 1062 | /// @param MO The MachineOperand in question. MO.isReg() should always |
| 1063 | /// be true. |
| 1064 | /// @return Whether this operand is allowed to be used PC-relatively. |
| 1065 | virtual bool isPCRelRegisterOperandLegal(const MachineOperand &MO) const { |
| 1066 | return false; |
| 1067 | } |
| 1068 | |
| 1069 | /// Return an index for MachineJumpTableInfo if \p insn is an indirect jump |
| 1070 | /// using a jump table, otherwise -1. |
| 1071 | virtual int getJumpTableIndex(const MachineInstr &MI) const { return -1; } |
| 1072 | |
| 1073 | protected: |
| 1074 | /// Target-dependent implementation for IsCopyInstr. |
| 1075 | /// If the specific machine instruction is a instruction that moves/copies |
| 1076 | /// value from one register to another register return destination and source |
| 1077 | /// registers as machine operands. |
| 1078 | virtual std::optional<DestSourcePair> |
| 1079 | isCopyInstrImpl(const MachineInstr &MI) const { |
| 1080 | return std::nullopt; |
| 1081 | } |
| 1082 | |
| 1083 | virtual std::optional<DestSourcePair> |
| 1084 | isCopyLikeInstrImpl(const MachineInstr &MI) const { |
| 1085 | return std::nullopt; |
| 1086 | } |
| 1087 | |
| 1088 | /// Return true if the given terminator MI is not expected to spill. This |
| 1089 | /// sets the live interval as not spillable and adjusts phi node lowering to |
| 1090 | /// not introduce copies after the terminator. Use with care, these are |
| 1091 | /// currently used for hardware loop intrinsics in very controlled situations, |
| 1092 | /// created prior to registry allocation in loops that only have single phi |
| 1093 | /// users for the terminators value. They may run out of registers if not used |
| 1094 | /// carefully. |
| 1095 | virtual bool isUnspillableTerminatorImpl(const MachineInstr *MI) const { |
| 1096 | return false; |
| 1097 | } |
| 1098 | |
| 1099 | public: |
| 1100 | /// If the specific machine instruction is a instruction that moves/copies |
| 1101 | /// value from one register to another register return destination and source |
| 1102 | /// registers as machine operands. |
| 1103 | /// For COPY-instruction the method naturally returns destination and source |
| 1104 | /// registers as machine operands, for all other instructions the method calls |
| 1105 | /// target-dependent implementation. |
| 1106 | std::optional<DestSourcePair> isCopyInstr(const MachineInstr &MI) const { |
| 1107 | if (MI.isCopy()) { |
| 1108 | return DestSourcePair{MI.getOperand(i: 0), MI.getOperand(i: 1)}; |
| 1109 | } |
| 1110 | return isCopyInstrImpl(MI); |
| 1111 | } |
| 1112 | |
| 1113 | // Similar to `isCopyInstr`, but adds non-copy semantics on MIR, but |
| 1114 | // ultimately generates a copy instruction. |
| 1115 | std::optional<DestSourcePair> isCopyLikeInstr(const MachineInstr &MI) const { |
| 1116 | if (auto IsCopyInstr = isCopyInstr(MI)) |
| 1117 | return IsCopyInstr; |
| 1118 | return isCopyLikeInstrImpl(MI); |
| 1119 | } |
| 1120 | |
| 1121 | bool isFullCopyInstr(const MachineInstr &MI) const { |
| 1122 | auto DestSrc = isCopyInstr(MI); |
| 1123 | if (!DestSrc) |
| 1124 | return false; |
| 1125 | |
| 1126 | const MachineOperand *DestRegOp = DestSrc->Destination; |
| 1127 | const MachineOperand *SrcRegOp = DestSrc->Source; |
| 1128 | return !DestRegOp->getSubReg() && !SrcRegOp->getSubReg(); |
| 1129 | } |
| 1130 | |
| 1131 | /// If the specific machine instruction is an instruction that adds an |
| 1132 | /// immediate value and a register, and stores the result in the given |
| 1133 | /// register \c Reg, return a pair of the source register and the offset |
| 1134 | /// which has been added. |
| 1135 | virtual std::optional<RegImmPair> isAddImmediate(const MachineInstr &MI, |
| 1136 | Register Reg) const { |
| 1137 | return std::nullopt; |
| 1138 | } |
| 1139 | |
| 1140 | /// Returns true if MI is an instruction that defines Reg to have a constant |
| 1141 | /// value and the value is recorded in ImmVal. The ImmVal is a result that |
| 1142 | /// should be interpreted as modulo size of Reg. |
| 1143 | virtual bool getConstValDefinedInReg(const MachineInstr &MI, |
| 1144 | const Register Reg, |
| 1145 | int64_t &ImmVal) const { |
| 1146 | return false; |
| 1147 | } |
| 1148 | |
| 1149 | /// Store the specified register of the given register class to the specified |
| 1150 | /// stack frame index. The store instruction is to be added to the given |
| 1151 | /// machine basic block before the specified machine instruction. If isKill |
| 1152 | /// is true, the register operand is the last use and must be marked kill. If |
| 1153 | /// \p SrcReg is being directly spilled as part of assigning a virtual |
| 1154 | /// register, \p VReg is the register being assigned. This additional register |
| 1155 | /// argument is needed for certain targets when invoked from RegAllocFast to |
| 1156 | /// map the spilled physical register to its virtual register. A null register |
| 1157 | /// can be passed elsewhere. The \p Flags is used to set appropriate machine |
| 1158 | /// flags on the spill instruction e.g. FrameSetup flag on a callee saved |
| 1159 | /// register spill instruction, part of prologue, during the frame lowering. |
| 1160 | virtual void storeRegToStackSlot( |
| 1161 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, Register SrcReg, |
| 1162 | bool isKill, int FrameIndex, const TargetRegisterClass *RC, |
| 1163 | const TargetRegisterInfo *TRI, Register VReg, |
| 1164 | MachineInstr::MIFlag Flags = MachineInstr::NoFlags) const { |
| 1165 | llvm_unreachable("Target didn't implement " |
| 1166 | "TargetInstrInfo::storeRegToStackSlot!" ); |
| 1167 | } |
| 1168 | |
| 1169 | /// Load the specified register of the given register class from the specified |
| 1170 | /// stack frame index. The load instruction is to be added to the given |
| 1171 | /// machine basic block before the specified machine instruction. If \p |
| 1172 | /// DestReg is being directly reloaded as part of assigning a virtual |
| 1173 | /// register, \p VReg is the register being assigned. This additional register |
| 1174 | /// argument is needed for certain targets when invoked from RegAllocFast to |
| 1175 | /// map the loaded physical register to its virtual register. A null register |
| 1176 | /// can be passed elsewhere. The \p Flags is used to set appropriate machine |
| 1177 | /// flags on the spill instruction e.g. FrameDestroy flag on a callee saved |
| 1178 | /// register reload instruction, part of epilogue, during the frame lowering. |
| 1179 | virtual void loadRegFromStackSlot( |
| 1180 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, Register DestReg, |
| 1181 | int FrameIndex, const TargetRegisterClass *RC, |
| 1182 | const TargetRegisterInfo *TRI, Register VReg, |
| 1183 | MachineInstr::MIFlag Flags = MachineInstr::NoFlags) const { |
| 1184 | llvm_unreachable("Target didn't implement " |
| 1185 | "TargetInstrInfo::loadRegFromStackSlot!" ); |
| 1186 | } |
| 1187 | |
| 1188 | /// This function is called for all pseudo instructions |
| 1189 | /// that remain after register allocation. Many pseudo instructions are |
| 1190 | /// created to help register allocation. This is the place to convert them |
| 1191 | /// into real instructions. The target can edit MI in place, or it can insert |
| 1192 | /// new instructions and erase MI. The function should return true if |
| 1193 | /// anything was changed. |
| 1194 | virtual bool expandPostRAPseudo(MachineInstr &MI) const { return false; } |
| 1195 | |
| 1196 | /// Check whether the target can fold a load that feeds a subreg operand |
| 1197 | /// (or a subreg operand that feeds a store). |
| 1198 | /// For example, X86 may want to return true if it can fold |
| 1199 | /// movl (%esp), %eax |
| 1200 | /// subb, %al, ... |
| 1201 | /// Into: |
| 1202 | /// subb (%esp), ... |
| 1203 | /// |
| 1204 | /// Ideally, we'd like the target implementation of foldMemoryOperand() to |
| 1205 | /// reject subregs - but since this behavior used to be enforced in the |
| 1206 | /// target-independent code, moving this responsibility to the targets |
| 1207 | /// has the potential of causing nasty silent breakage in out-of-tree targets. |
| 1208 | virtual bool isSubregFoldable() const { return false; } |
| 1209 | |
| 1210 | /// For a patchpoint, stackmap, or statepoint intrinsic, return the range of |
| 1211 | /// operands which can't be folded into stack references. Operands outside |
| 1212 | /// of the range are most likely foldable but it is not guaranteed. |
| 1213 | /// These instructions are unique in that stack references for some operands |
| 1214 | /// have the same execution cost (e.g. none) as the unfolded register forms. |
| 1215 | /// The ranged return is guaranteed to include all operands which can't be |
| 1216 | /// folded at zero cost. |
| 1217 | virtual std::pair<unsigned, unsigned> |
| 1218 | getPatchpointUnfoldableRange(const MachineInstr &MI) const; |
| 1219 | |
| 1220 | /// Attempt to fold a load or store of the specified stack |
| 1221 | /// slot into the specified machine instruction for the specified operand(s). |
| 1222 | /// If this is possible, a new instruction is returned with the specified |
| 1223 | /// operand folded, otherwise NULL is returned. |
| 1224 | /// The new instruction is inserted before MI, and the client is responsible |
| 1225 | /// for removing the old instruction. |
| 1226 | /// If VRM is passed, the assigned physregs can be inspected by target to |
| 1227 | /// decide on using an opcode (note that those assignments can still change). |
| 1228 | MachineInstr *foldMemoryOperand(MachineInstr &MI, ArrayRef<unsigned> Ops, |
| 1229 | int FI, |
| 1230 | LiveIntervals *LIS = nullptr, |
| 1231 | VirtRegMap *VRM = nullptr) const; |
| 1232 | |
| 1233 | /// Same as the previous version except it allows folding of any load and |
| 1234 | /// store from / to any address, not just from a specific stack slot. |
| 1235 | MachineInstr *foldMemoryOperand(MachineInstr &MI, ArrayRef<unsigned> Ops, |
| 1236 | MachineInstr &LoadMI, |
| 1237 | LiveIntervals *LIS = nullptr) const; |
| 1238 | |
| 1239 | /// This function defines the logic to lower COPY instruction to |
| 1240 | /// target specific instruction(s). |
| 1241 | void lowerCopy(MachineInstr *MI, const TargetRegisterInfo *TRI) const; |
| 1242 | |
| 1243 | /// Return true when there is potentially a faster code sequence |
| 1244 | /// for an instruction chain ending in \p Root. All potential patterns are |
| 1245 | /// returned in the \p Pattern vector. Pattern should be sorted in priority |
| 1246 | /// order since the pattern evaluator stops checking as soon as it finds a |
| 1247 | /// faster sequence. |
| 1248 | /// \param Root - Instruction that could be combined with one of its operands |
| 1249 | /// \param Patterns - Vector of possible combination patterns |
| 1250 | virtual bool getMachineCombinerPatterns(MachineInstr &Root, |
| 1251 | SmallVectorImpl<unsigned> &Patterns, |
| 1252 | bool DoRegPressureReduce) const; |
| 1253 | |
| 1254 | /// Return true if target supports reassociation of instructions in machine |
| 1255 | /// combiner pass to reduce register pressure for a given BB. |
| 1256 | virtual bool |
| 1257 | shouldReduceRegisterPressure(const MachineBasicBlock *MBB, |
| 1258 | const RegisterClassInfo *RegClassInfo) const { |
| 1259 | return false; |
| 1260 | } |
| 1261 | |
| 1262 | /// Fix up the placeholder we may add in genAlternativeCodeSequence(). |
| 1263 | virtual void |
| 1264 | finalizeInsInstrs(MachineInstr &Root, unsigned &Pattern, |
| 1265 | SmallVectorImpl<MachineInstr *> &InsInstrs) const {} |
| 1266 | |
| 1267 | /// Return true when a code sequence can improve throughput. It |
| 1268 | /// should be called only for instructions in loops. |
| 1269 | /// \param Pattern - combiner pattern |
| 1270 | virtual bool isThroughputPattern(unsigned Pattern) const; |
| 1271 | |
| 1272 | /// Return the objective of a combiner pattern. |
| 1273 | /// \param Pattern - combiner pattern |
| 1274 | virtual CombinerObjective getCombinerObjective(unsigned Pattern) const; |
| 1275 | |
| 1276 | /// Return true if the input \P Inst is part of a chain of dependent ops |
| 1277 | /// that are suitable for reassociation, otherwise return false. |
| 1278 | /// If the instruction's operands must be commuted to have a previous |
| 1279 | /// instruction of the same type define the first source operand, \P Commuted |
| 1280 | /// will be set to true. |
| 1281 | bool isReassociationCandidate(const MachineInstr &Inst, bool &Commuted) const; |
| 1282 | |
| 1283 | /// Return true when \P Inst is both associative and commutative. If \P Invert |
| 1284 | /// is true, then the inverse of \P Inst operation must be tested. |
| 1285 | virtual bool isAssociativeAndCommutative(const MachineInstr &Inst, |
| 1286 | bool Invert = false) const { |
| 1287 | return false; |
| 1288 | } |
| 1289 | |
| 1290 | /// Find chains of accumulations that can be rewritten as a tree for increased |
| 1291 | /// ILP. |
| 1292 | bool getAccumulatorReassociationPatterns( |
| 1293 | MachineInstr &Root, SmallVectorImpl<unsigned> &Patterns) const; |
| 1294 | |
| 1295 | /// Find the chain of accumulator instructions in \P MBB and return them in |
| 1296 | /// \P Chain. |
| 1297 | void getAccumulatorChain(MachineInstr *CurrentInstr, |
| 1298 | SmallVectorImpl<Register> &Chain) const; |
| 1299 | |
| 1300 | /// Return true when \P OpCode is an instruction which performs |
| 1301 | /// accumulation into one of its operand registers. |
| 1302 | virtual bool isAccumulationOpcode(unsigned Opcode) const { return false; } |
| 1303 | |
| 1304 | /// Returns an opcode which defines the accumulator used by \P Opcode. |
| 1305 | virtual unsigned getAccumulationStartOpcode(unsigned Opcode) const { |
| 1306 | llvm_unreachable("Function not implemented for target!" ); |
| 1307 | return 0; |
| 1308 | } |
| 1309 | |
| 1310 | /// Returns the opcode that should be use to reduce accumulation registers. |
| 1311 | virtual unsigned |
| 1312 | getReduceOpcodeForAccumulator(unsigned int AccumulatorOpCode) const { |
| 1313 | llvm_unreachable("Function not implemented for target!" ); |
| 1314 | return 0; |
| 1315 | } |
| 1316 | |
| 1317 | /// Reduces branches of the accumulator tree into a single register. |
| 1318 | void reduceAccumulatorTree(SmallVectorImpl<Register> &RegistersToReduce, |
| 1319 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 1320 | MachineFunction &MF, MachineInstr &Root, |
| 1321 | MachineRegisterInfo &MRI, |
| 1322 | DenseMap<Register, unsigned> &InstrIdxForVirtReg, |
| 1323 | Register ResultReg) const; |
| 1324 | |
| 1325 | /// Return the inverse operation opcode if it exists for \P Opcode (e.g. add |
| 1326 | /// for sub and vice versa). |
| 1327 | virtual std::optional<unsigned> getInverseOpcode(unsigned Opcode) const { |
| 1328 | return std::nullopt; |
| 1329 | } |
| 1330 | |
| 1331 | /// Return true when \P Opcode1 or its inversion is equal to \P Opcode2. |
| 1332 | bool areOpcodesEqualOrInverse(unsigned Opcode1, unsigned Opcode2) const; |
| 1333 | |
| 1334 | /// Return true when \P Inst has reassociable operands in the same \P MBB. |
| 1335 | virtual bool hasReassociableOperands(const MachineInstr &Inst, |
| 1336 | const MachineBasicBlock *MBB) const; |
| 1337 | |
| 1338 | /// Return true when \P Inst has reassociable sibling. |
| 1339 | virtual bool hasReassociableSibling(const MachineInstr &Inst, |
| 1340 | bool &Commuted) const; |
| 1341 | |
| 1342 | /// When getMachineCombinerPatterns() finds patterns, this function generates |
| 1343 | /// the instructions that could replace the original code sequence. The client |
| 1344 | /// has to decide whether the actual replacement is beneficial or not. |
| 1345 | /// \param Root - Instruction that could be combined with one of its operands |
| 1346 | /// \param Pattern - Combination pattern for Root |
| 1347 | /// \param InsInstrs - Vector of new instructions that implement P |
| 1348 | /// \param DelInstrs - Old instructions, including Root, that could be |
| 1349 | /// replaced by InsInstr |
| 1350 | /// \param InstIdxForVirtReg - map of virtual register to instruction in |
| 1351 | /// InsInstr that defines it |
| 1352 | virtual void genAlternativeCodeSequence( |
| 1353 | MachineInstr &Root, unsigned Pattern, |
| 1354 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 1355 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
| 1356 | DenseMap<Register, unsigned> &InstIdxForVirtReg) const; |
| 1357 | |
| 1358 | /// When calculate the latency of the root instruction, accumulate the |
| 1359 | /// latency of the sequence to the root latency. |
| 1360 | /// \param Root - Instruction that could be combined with one of its operands |
| 1361 | virtual bool accumulateInstrSeqToRootLatency(MachineInstr &Root) const { |
| 1362 | return true; |
| 1363 | } |
| 1364 | |
| 1365 | /// The returned array encodes the operand index for each parameter because |
| 1366 | /// the operands may be commuted; the operand indices for associative |
| 1367 | /// operations might also be target-specific. Each element specifies the index |
| 1368 | /// of {Prev, A, B, X, Y}. |
| 1369 | virtual void |
| 1370 | getReassociateOperandIndices(const MachineInstr &Root, unsigned Pattern, |
| 1371 | std::array<unsigned, 5> &OperandIndices) const; |
| 1372 | |
| 1373 | /// Attempt to reassociate \P Root and \P Prev according to \P Pattern to |
| 1374 | /// reduce critical path length. |
| 1375 | void reassociateOps(MachineInstr &Root, MachineInstr &Prev, unsigned Pattern, |
| 1376 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
| 1377 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
| 1378 | ArrayRef<unsigned> OperandIndices, |
| 1379 | DenseMap<Register, unsigned> &InstrIdxForVirtReg) const; |
| 1380 | |
| 1381 | /// Reassociation of some instructions requires inverse operations (e.g. |
| 1382 | /// (X + A) - Y => (X - Y) + A). This method returns a pair of new opcodes |
| 1383 | /// (new root opcode, new prev opcode) that must be used to reassociate \P |
| 1384 | /// Root and \P Prev accoring to \P Pattern. |
| 1385 | std::pair<unsigned, unsigned> |
| 1386 | getReassociationOpcodes(unsigned Pattern, const MachineInstr &Root, |
| 1387 | const MachineInstr &Prev) const; |
| 1388 | |
| 1389 | /// The limit on resource length extension we accept in MachineCombiner Pass. |
| 1390 | virtual int getExtendResourceLenLimit() const { return 0; } |
| 1391 | |
| 1392 | /// This is an architecture-specific helper function of reassociateOps. |
| 1393 | /// Set special operand attributes for new instructions after reassociation. |
| 1394 | virtual void setSpecialOperandAttr(MachineInstr &OldMI1, MachineInstr &OldMI2, |
| 1395 | MachineInstr &NewMI1, |
| 1396 | MachineInstr &NewMI2) const {} |
| 1397 | |
| 1398 | /// Return true when a target supports MachineCombiner. |
| 1399 | virtual bool useMachineCombiner() const { return false; } |
| 1400 | |
| 1401 | /// Return a strategy that MachineCombiner must use when creating traces. |
| 1402 | virtual MachineTraceStrategy getMachineCombinerTraceStrategy() const; |
| 1403 | |
| 1404 | /// Return true if the given SDNode can be copied during scheduling |
| 1405 | /// even if it has glue. |
| 1406 | virtual bool canCopyGluedNodeDuringSchedule(SDNode *N) const { return false; } |
| 1407 | |
| 1408 | protected: |
| 1409 | /// Target-dependent implementation for foldMemoryOperand. |
| 1410 | /// Target-independent code in foldMemoryOperand will |
| 1411 | /// take care of adding a MachineMemOperand to the newly created instruction. |
| 1412 | /// The instruction and any auxiliary instructions necessary will be inserted |
| 1413 | /// at InsertPt. |
| 1414 | virtual MachineInstr * |
| 1415 | foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI, |
| 1416 | ArrayRef<unsigned> Ops, |
| 1417 | MachineBasicBlock::iterator InsertPt, int FrameIndex, |
| 1418 | LiveIntervals *LIS = nullptr, |
| 1419 | VirtRegMap *VRM = nullptr) const { |
| 1420 | return nullptr; |
| 1421 | } |
| 1422 | |
| 1423 | /// Target-dependent implementation for foldMemoryOperand. |
| 1424 | /// Target-independent code in foldMemoryOperand will |
| 1425 | /// take care of adding a MachineMemOperand to the newly created instruction. |
| 1426 | /// The instruction and any auxiliary instructions necessary will be inserted |
| 1427 | /// at InsertPt. |
| 1428 | virtual MachineInstr *foldMemoryOperandImpl( |
| 1429 | MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, |
| 1430 | MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, |
| 1431 | LiveIntervals *LIS = nullptr) const { |
| 1432 | return nullptr; |
| 1433 | } |
| 1434 | |
| 1435 | /// Target-dependent implementation of getRegSequenceInputs. |
| 1436 | /// |
| 1437 | /// \returns true if it is possible to build the equivalent |
| 1438 | /// REG_SEQUENCE inputs with the pair \p MI, \p DefIdx. False otherwise. |
| 1439 | /// |
| 1440 | /// \pre MI.isRegSequenceLike(). |
| 1441 | /// |
| 1442 | /// \see TargetInstrInfo::getRegSequenceInputs. |
| 1443 | virtual bool getRegSequenceLikeInputs( |
| 1444 | const MachineInstr &MI, unsigned DefIdx, |
| 1445 | SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const { |
| 1446 | return false; |
| 1447 | } |
| 1448 | |
| 1449 | /// Target-dependent implementation of getExtractSubregInputs. |
| 1450 | /// |
| 1451 | /// \returns true if it is possible to build the equivalent |
| 1452 | /// EXTRACT_SUBREG inputs with the pair \p MI, \p DefIdx. False otherwise. |
| 1453 | /// |
| 1454 | /// \pre MI.isExtractSubregLike(). |
| 1455 | /// |
| 1456 | /// \see TargetInstrInfo::getExtractSubregInputs. |
| 1457 | virtual bool getExtractSubregLikeInputs(const MachineInstr &MI, |
| 1458 | unsigned DefIdx, |
| 1459 | RegSubRegPairAndIdx &InputReg) const { |
| 1460 | return false; |
| 1461 | } |
| 1462 | |
| 1463 | /// Target-dependent implementation of getInsertSubregInputs. |
| 1464 | /// |
| 1465 | /// \returns true if it is possible to build the equivalent |
| 1466 | /// INSERT_SUBREG inputs with the pair \p MI, \p DefIdx. False otherwise. |
| 1467 | /// |
| 1468 | /// \pre MI.isInsertSubregLike(). |
| 1469 | /// |
| 1470 | /// \see TargetInstrInfo::getInsertSubregInputs. |
| 1471 | virtual bool |
| 1472 | getInsertSubregLikeInputs(const MachineInstr &MI, unsigned DefIdx, |
| 1473 | RegSubRegPair &BaseReg, |
| 1474 | RegSubRegPairAndIdx &InsertedReg) const { |
| 1475 | return false; |
| 1476 | } |
| 1477 | |
| 1478 | public: |
| 1479 | /// unfoldMemoryOperand - Separate a single instruction which folded a load or |
| 1480 | /// a store or a load and a store into two or more instruction. If this is |
| 1481 | /// possible, returns true as well as the new instructions by reference. |
| 1482 | virtual bool |
| 1483 | unfoldMemoryOperand(MachineFunction &MF, MachineInstr &MI, Register Reg, |
| 1484 | bool UnfoldLoad, bool UnfoldStore, |
| 1485 | SmallVectorImpl<MachineInstr *> &NewMIs) const { |
| 1486 | return false; |
| 1487 | } |
| 1488 | |
| 1489 | virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N, |
| 1490 | SmallVectorImpl<SDNode *> &NewNodes) const { |
| 1491 | return false; |
| 1492 | } |
| 1493 | |
| 1494 | /// Returns the opcode of the would be new |
| 1495 | /// instruction after load / store are unfolded from an instruction of the |
| 1496 | /// specified opcode. It returns zero if the specified unfolding is not |
| 1497 | /// possible. If LoadRegIndex is non-null, it is filled in with the operand |
| 1498 | /// index of the operand which will hold the register holding the loaded |
| 1499 | /// value. |
| 1500 | virtual unsigned |
| 1501 | getOpcodeAfterMemoryUnfold(unsigned Opc, bool UnfoldLoad, bool UnfoldStore, |
| 1502 | unsigned *LoadRegIndex = nullptr) const { |
| 1503 | return 0; |
| 1504 | } |
| 1505 | |
| 1506 | /// This is used by the pre-regalloc scheduler to determine if two loads are |
| 1507 | /// loading from the same base address. It should only return true if the base |
| 1508 | /// pointers are the same and the only differences between the two addresses |
| 1509 | /// are the offset. It also returns the offsets by reference. |
| 1510 | virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, |
| 1511 | int64_t &Offset1, |
| 1512 | int64_t &Offset2) const { |
| 1513 | return false; |
| 1514 | } |
| 1515 | |
| 1516 | /// This is a used by the pre-regalloc scheduler to determine (in conjunction |
| 1517 | /// with areLoadsFromSameBasePtr) if two loads should be scheduled together. |
| 1518 | /// On some targets if two loads are loading from |
| 1519 | /// addresses in the same cache line, it's better if they are scheduled |
| 1520 | /// together. This function takes two integers that represent the load offsets |
| 1521 | /// from the common base address. It returns true if it decides it's desirable |
| 1522 | /// to schedule the two loads together. "NumLoads" is the number of loads that |
| 1523 | /// have already been scheduled after Load1. |
| 1524 | virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2, |
| 1525 | int64_t Offset1, int64_t Offset2, |
| 1526 | unsigned NumLoads) const { |
| 1527 | return false; |
| 1528 | } |
| 1529 | |
| 1530 | /// Get the base operand and byte offset of an instruction that reads/writes |
| 1531 | /// memory. This is a convenience function for callers that are only prepared |
| 1532 | /// to handle a single base operand. |
| 1533 | /// FIXME: Move Offset and OffsetIsScalable to some ElementCount-style |
| 1534 | /// abstraction that supports negative offsets. |
| 1535 | bool getMemOperandWithOffset(const MachineInstr &MI, |
| 1536 | const MachineOperand *&BaseOp, int64_t &Offset, |
| 1537 | bool &OffsetIsScalable, |
| 1538 | const TargetRegisterInfo *TRI) const; |
| 1539 | |
| 1540 | /// Get zero or more base operands and the byte offset of an instruction that |
| 1541 | /// reads/writes memory. Note that there may be zero base operands if the |
| 1542 | /// instruction accesses a constant address. |
| 1543 | /// It returns false if MI does not read/write memory. |
| 1544 | /// It returns false if base operands and offset could not be determined. |
| 1545 | /// It is not guaranteed to always recognize base operands and offsets in all |
| 1546 | /// cases. |
| 1547 | /// FIXME: Move Offset and OffsetIsScalable to some ElementCount-style |
| 1548 | /// abstraction that supports negative offsets. |
| 1549 | virtual bool getMemOperandsWithOffsetWidth( |
| 1550 | const MachineInstr &MI, SmallVectorImpl<const MachineOperand *> &BaseOps, |
| 1551 | int64_t &Offset, bool &OffsetIsScalable, LocationSize &Width, |
| 1552 | const TargetRegisterInfo *TRI) const { |
| 1553 | return false; |
| 1554 | } |
| 1555 | |
| 1556 | /// Return true if the instruction contains a base register and offset. If |
| 1557 | /// true, the function also sets the operand position in the instruction |
| 1558 | /// for the base register and offset. |
| 1559 | virtual bool getBaseAndOffsetPosition(const MachineInstr &MI, |
| 1560 | unsigned &BasePos, |
| 1561 | unsigned &OffsetPos) const { |
| 1562 | return false; |
| 1563 | } |
| 1564 | |
| 1565 | /// Target dependent implementation to get the values constituting the address |
| 1566 | /// MachineInstr that is accessing memory. These values are returned as a |
| 1567 | /// struct ExtAddrMode which contains all relevant information to make up the |
| 1568 | /// address. |
| 1569 | virtual std::optional<ExtAddrMode> |
| 1570 | getAddrModeFromMemoryOp(const MachineInstr &MemI, |
| 1571 | const TargetRegisterInfo *TRI) const { |
| 1572 | return std::nullopt; |
| 1573 | } |
| 1574 | |
| 1575 | /// Check if it's possible and beneficial to fold the addressing computation |
| 1576 | /// `AddrI` into the addressing mode of the load/store instruction `MemI`. The |
| 1577 | /// memory instruction is a user of the virtual register `Reg`, which in turn |
| 1578 | /// is the ultimate destination of zero or more COPY instructions from the |
| 1579 | /// output register of `AddrI`. |
| 1580 | /// Return the adddressing mode after folding in `AM`. |
| 1581 | virtual bool canFoldIntoAddrMode(const MachineInstr &MemI, Register Reg, |
| 1582 | const MachineInstr &AddrI, |
| 1583 | ExtAddrMode &AM) const { |
| 1584 | return false; |
| 1585 | } |
| 1586 | |
| 1587 | /// Emit a load/store instruction with the same value register as `MemI`, but |
| 1588 | /// using the address from `AM`. The addressing mode must have been obtained |
| 1589 | /// from `canFoldIntoAddr` for the same memory instruction. |
| 1590 | virtual MachineInstr *emitLdStWithAddr(MachineInstr &MemI, |
| 1591 | const ExtAddrMode &AM) const { |
| 1592 | llvm_unreachable("target did not implement emitLdStWithAddr()" ); |
| 1593 | } |
| 1594 | |
| 1595 | /// Returns true if MI's Def is NullValueReg, and the MI |
| 1596 | /// does not change the Zero value. i.e. cases such as rax = shr rax, X where |
| 1597 | /// NullValueReg = rax. Note that if the NullValueReg is non-zero, this |
| 1598 | /// function can return true even if becomes zero. Specifically cases such as |
| 1599 | /// NullValueReg = shl NullValueReg, 63. |
| 1600 | virtual bool preservesZeroValueInReg(const MachineInstr *MI, |
| 1601 | const Register NullValueReg, |
| 1602 | const TargetRegisterInfo *TRI) const { |
| 1603 | return false; |
| 1604 | } |
| 1605 | |
| 1606 | /// If the instruction is an increment of a constant value, return the amount. |
| 1607 | virtual bool getIncrementValue(const MachineInstr &MI, int &Value) const { |
| 1608 | return false; |
| 1609 | } |
| 1610 | |
| 1611 | /// Returns true if the two given memory operations should be scheduled |
| 1612 | /// adjacent. Note that you have to add: |
| 1613 | /// DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); |
| 1614 | /// or |
| 1615 | /// DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); |
| 1616 | /// to TargetMachine::createMachineScheduler() to have an effect. |
| 1617 | /// |
| 1618 | /// \p BaseOps1 and \p BaseOps2 are memory operands of two memory operations. |
| 1619 | /// \p Offset1 and \p Offset2 are the byte offsets for the memory |
| 1620 | /// operations. |
| 1621 | /// \p OffsetIsScalable1 and \p OffsetIsScalable2 indicate if the offset is |
| 1622 | /// scaled by a runtime quantity. |
| 1623 | /// \p ClusterSize is the number of operations in the resulting load/store |
| 1624 | /// cluster if this hook returns true. |
| 1625 | /// \p NumBytes is the number of bytes that will be loaded from all the |
| 1626 | /// clustered loads if this hook returns true. |
| 1627 | virtual bool shouldClusterMemOps(ArrayRef<const MachineOperand *> BaseOps1, |
| 1628 | int64_t Offset1, bool OffsetIsScalable1, |
| 1629 | ArrayRef<const MachineOperand *> BaseOps2, |
| 1630 | int64_t Offset2, bool OffsetIsScalable2, |
| 1631 | unsigned ClusterSize, |
| 1632 | unsigned NumBytes) const { |
| 1633 | llvm_unreachable("target did not implement shouldClusterMemOps()" ); |
| 1634 | } |
| 1635 | |
| 1636 | /// Reverses the branch condition of the specified condition list, |
| 1637 | /// returning false on success and true if it cannot be reversed. |
| 1638 | virtual bool |
| 1639 | reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { |
| 1640 | return true; |
| 1641 | } |
| 1642 | |
| 1643 | /// Insert a noop into the instruction stream at the specified point. |
| 1644 | virtual void insertNoop(MachineBasicBlock &MBB, |
| 1645 | MachineBasicBlock::iterator MI) const; |
| 1646 | |
| 1647 | /// Insert noops into the instruction stream at the specified point. |
| 1648 | virtual void insertNoops(MachineBasicBlock &MBB, |
| 1649 | MachineBasicBlock::iterator MI, |
| 1650 | unsigned Quantity) const; |
| 1651 | |
| 1652 | /// Return the noop instruction to use for a noop. |
| 1653 | virtual MCInst getNop() const; |
| 1654 | |
| 1655 | /// Return true for post-incremented instructions. |
| 1656 | virtual bool isPostIncrement(const MachineInstr &MI) const { return false; } |
| 1657 | |
| 1658 | /// Returns true if the instruction is already predicated. |
| 1659 | virtual bool isPredicated(const MachineInstr &MI) const { return false; } |
| 1660 | |
| 1661 | /// Assumes the instruction is already predicated and returns true if the |
| 1662 | /// instruction can be predicated again. |
| 1663 | virtual bool canPredicatePredicatedInstr(const MachineInstr &MI) const { |
| 1664 | assert(isPredicated(MI) && "Instruction is not predicated" ); |
| 1665 | return false; |
| 1666 | } |
| 1667 | |
| 1668 | // Returns a MIRPrinter comment for this machine operand. |
| 1669 | virtual std::string |
| 1670 | createMIROperandComment(const MachineInstr &MI, const MachineOperand &Op, |
| 1671 | unsigned OpIdx, const TargetRegisterInfo *TRI) const; |
| 1672 | |
| 1673 | /// Returns true if the instruction is a |
| 1674 | /// terminator instruction that has not been predicated. |
| 1675 | bool isUnpredicatedTerminator(const MachineInstr &MI) const; |
| 1676 | |
| 1677 | /// Returns true if MI is an unconditional tail call. |
| 1678 | virtual bool isUnconditionalTailCall(const MachineInstr &MI) const { |
| 1679 | return false; |
| 1680 | } |
| 1681 | |
| 1682 | /// Returns true if the tail call can be made conditional on BranchCond. |
| 1683 | virtual bool canMakeTailCallConditional(SmallVectorImpl<MachineOperand> &Cond, |
| 1684 | const MachineInstr &TailCall) const { |
| 1685 | return false; |
| 1686 | } |
| 1687 | |
| 1688 | /// Replace the conditional branch in MBB with a conditional tail call. |
| 1689 | virtual void replaceBranchWithTailCall(MachineBasicBlock &MBB, |
| 1690 | SmallVectorImpl<MachineOperand> &Cond, |
| 1691 | const MachineInstr &TailCall) const { |
| 1692 | llvm_unreachable("Target didn't implement replaceBranchWithTailCall!" ); |
| 1693 | } |
| 1694 | |
| 1695 | /// Convert the instruction into a predicated instruction. |
| 1696 | /// It returns true if the operation was successful. |
| 1697 | virtual bool PredicateInstruction(MachineInstr &MI, |
| 1698 | ArrayRef<MachineOperand> Pred) const; |
| 1699 | |
| 1700 | /// Returns true if the first specified predicate |
| 1701 | /// subsumes the second, e.g. GE subsumes GT. |
| 1702 | virtual bool SubsumesPredicate(ArrayRef<MachineOperand> Pred1, |
| 1703 | ArrayRef<MachineOperand> Pred2) const { |
| 1704 | return false; |
| 1705 | } |
| 1706 | |
| 1707 | /// If the specified instruction defines any predicate |
| 1708 | /// or condition code register(s) used for predication, returns true as well |
| 1709 | /// as the definition predicate(s) by reference. |
| 1710 | /// SkipDead should be set to false at any point that dead |
| 1711 | /// predicate instructions should be considered as being defined. |
| 1712 | /// A dead predicate instruction is one that is guaranteed to be removed |
| 1713 | /// after a call to PredicateInstruction. |
| 1714 | virtual bool ClobbersPredicate(MachineInstr &MI, |
| 1715 | std::vector<MachineOperand> &Pred, |
| 1716 | bool SkipDead) const { |
| 1717 | return false; |
| 1718 | } |
| 1719 | |
| 1720 | /// Return true if the specified instruction can be predicated. |
| 1721 | /// By default, this returns true for every instruction with a |
| 1722 | /// PredicateOperand. |
| 1723 | virtual bool isPredicable(const MachineInstr &MI) const { |
| 1724 | return MI.getDesc().isPredicable(); |
| 1725 | } |
| 1726 | |
| 1727 | /// Return true if it's safe to move a machine |
| 1728 | /// instruction that defines the specified register class. |
| 1729 | virtual bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const { |
| 1730 | return true; |
| 1731 | } |
| 1732 | |
| 1733 | /// Return true if it's safe to move a machine instruction. |
| 1734 | /// This allows the backend to prevent certain special instruction |
| 1735 | /// sequences from being broken by instruction motion in optimization |
| 1736 | /// passes. |
| 1737 | /// By default, this returns true for every instruction. |
| 1738 | virtual bool isSafeToMove(const MachineInstr &MI, |
| 1739 | const MachineBasicBlock *MBB, |
| 1740 | const MachineFunction &MF) const { |
| 1741 | return true; |
| 1742 | } |
| 1743 | |
| 1744 | /// Test if the given instruction should be considered a scheduling boundary. |
| 1745 | /// This primarily includes labels and terminators. |
| 1746 | virtual bool isSchedulingBoundary(const MachineInstr &MI, |
| 1747 | const MachineBasicBlock *MBB, |
| 1748 | const MachineFunction &MF) const; |
| 1749 | |
| 1750 | /// Measure the specified inline asm to determine an approximation of its |
| 1751 | /// length. |
| 1752 | virtual unsigned getInlineAsmLength( |
| 1753 | const char *Str, const MCAsmInfo &MAI, |
| 1754 | const TargetSubtargetInfo *STI = nullptr) const; |
| 1755 | |
| 1756 | /// Allocate and return a hazard recognizer to use for this target when |
| 1757 | /// scheduling the machine instructions before register allocation. |
| 1758 | virtual ScheduleHazardRecognizer * |
| 1759 | CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, |
| 1760 | const ScheduleDAG *DAG) const; |
| 1761 | |
| 1762 | /// Allocate and return a hazard recognizer to use for this target when |
| 1763 | /// scheduling the machine instructions before register allocation. |
| 1764 | virtual ScheduleHazardRecognizer * |
| 1765 | CreateTargetMIHazardRecognizer(const InstrItineraryData *, |
| 1766 | const ScheduleDAGMI *DAG) const; |
| 1767 | |
| 1768 | /// Allocate and return a hazard recognizer to use for this target when |
| 1769 | /// scheduling the machine instructions after register allocation. |
| 1770 | virtual ScheduleHazardRecognizer * |
| 1771 | CreateTargetPostRAHazardRecognizer(const InstrItineraryData *, |
| 1772 | const ScheduleDAG *DAG) const; |
| 1773 | |
| 1774 | /// Allocate and return a hazard recognizer to use for by non-scheduling |
| 1775 | /// passes. |
| 1776 | virtual ScheduleHazardRecognizer * |
| 1777 | CreateTargetPostRAHazardRecognizer(const MachineFunction &MF) const { |
| 1778 | return nullptr; |
| 1779 | } |
| 1780 | |
| 1781 | /// Provide a global flag for disabling the PreRA hazard recognizer that |
| 1782 | /// targets may choose to honor. |
| 1783 | bool usePreRAHazardRecognizer() const; |
| 1784 | |
| 1785 | /// For a comparison instruction, return the source registers |
| 1786 | /// in SrcReg and SrcReg2 if having two register operands, and the value it |
| 1787 | /// compares against in CmpValue. Return true if the comparison instruction |
| 1788 | /// can be analyzed. |
| 1789 | virtual bool analyzeCompare(const MachineInstr &MI, Register &SrcReg, |
| 1790 | Register &SrcReg2, int64_t &Mask, |
| 1791 | int64_t &Value) const { |
| 1792 | return false; |
| 1793 | } |
| 1794 | |
| 1795 | /// See if the comparison instruction can be converted |
| 1796 | /// into something more efficient. E.g., on ARM most instructions can set the |
| 1797 | /// flags register, obviating the need for a separate CMP. |
| 1798 | virtual bool optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg, |
| 1799 | Register SrcReg2, int64_t Mask, |
| 1800 | int64_t Value, |
| 1801 | const MachineRegisterInfo *MRI) const { |
| 1802 | return false; |
| 1803 | } |
| 1804 | virtual bool optimizeCondBranch(MachineInstr &MI) const { return false; } |
| 1805 | |
| 1806 | /// Try to remove the load by folding it to a register operand at the use. |
| 1807 | /// We fold the load instructions if and only if the |
| 1808 | /// def and use are in the same BB. We only look at one load and see |
| 1809 | /// whether it can be folded into MI. FoldAsLoadDefReg is the virtual register |
| 1810 | /// defined by the load we are trying to fold. DefMI returns the machine |
| 1811 | /// instruction that defines FoldAsLoadDefReg, and the function returns |
| 1812 | /// the machine instruction generated due to folding. |
| 1813 | virtual MachineInstr *optimizeLoadInstr(MachineInstr &MI, |
| 1814 | const MachineRegisterInfo *MRI, |
| 1815 | Register &FoldAsLoadDefReg, |
| 1816 | MachineInstr *&DefMI) const; |
| 1817 | |
| 1818 | /// 'Reg' is known to be defined by a move immediate instruction, |
| 1819 | /// try to fold the immediate into the use instruction. |
| 1820 | /// If MRI->hasOneNonDBGUse(Reg) is true, and this function returns true, |
| 1821 | /// then the caller may assume that DefMI has been erased from its parent |
| 1822 | /// block. The caller may assume that it will not be erased by this |
| 1823 | /// function otherwise. |
| 1824 | virtual bool foldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, |
| 1825 | Register Reg, MachineRegisterInfo *MRI) const { |
| 1826 | return false; |
| 1827 | } |
| 1828 | |
| 1829 | /// Return the number of u-operations the given machine |
| 1830 | /// instruction will be decoded to on the target cpu. The itinerary's |
| 1831 | /// IssueWidth is the number of microops that can be dispatched each |
| 1832 | /// cycle. An instruction with zero microops takes no dispatch resources. |
| 1833 | virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData, |
| 1834 | const MachineInstr &MI) const; |
| 1835 | |
| 1836 | /// Return true for pseudo instructions that don't consume any |
| 1837 | /// machine resources in their current form. These are common cases that the |
| 1838 | /// scheduler should consider free, rather than conservatively handling them |
| 1839 | /// as instructions with no itinerary. |
| 1840 | bool isZeroCost(unsigned Opcode) const { |
| 1841 | return Opcode <= TargetOpcode::COPY; |
| 1842 | } |
| 1843 | |
| 1844 | virtual std::optional<unsigned> |
| 1845 | getOperandLatency(const InstrItineraryData *ItinData, SDNode *DefNode, |
| 1846 | unsigned DefIdx, SDNode *UseNode, unsigned UseIdx) const; |
| 1847 | |
| 1848 | /// Compute and return the use operand latency of a given pair of def and use. |
| 1849 | /// In most cases, the static scheduling itinerary was enough to determine the |
| 1850 | /// operand latency. But it may not be possible for instructions with variable |
| 1851 | /// number of defs / uses. |
| 1852 | /// |
| 1853 | /// This is a raw interface to the itinerary that may be directly overridden |
| 1854 | /// by a target. Use computeOperandLatency to get the best estimate of |
| 1855 | /// latency. |
| 1856 | virtual std::optional<unsigned> |
| 1857 | getOperandLatency(const InstrItineraryData *ItinData, |
| 1858 | const MachineInstr &DefMI, unsigned DefIdx, |
| 1859 | const MachineInstr &UseMI, unsigned UseIdx) const; |
| 1860 | |
| 1861 | /// Compute the instruction latency of a given instruction. |
| 1862 | /// If the instruction has higher cost when predicated, it's returned via |
| 1863 | /// PredCost. |
| 1864 | virtual unsigned getInstrLatency(const InstrItineraryData *ItinData, |
| 1865 | const MachineInstr &MI, |
| 1866 | unsigned *PredCost = nullptr) const; |
| 1867 | |
| 1868 | virtual unsigned getPredicationCost(const MachineInstr &MI) const; |
| 1869 | |
| 1870 | virtual unsigned getInstrLatency(const InstrItineraryData *ItinData, |
| 1871 | SDNode *Node) const; |
| 1872 | |
| 1873 | /// Return the default expected latency for a def based on its opcode. |
| 1874 | unsigned defaultDefLatency(const MCSchedModel &SchedModel, |
| 1875 | const MachineInstr &DefMI) const; |
| 1876 | |
| 1877 | /// Return true if this opcode has high latency to its result. |
| 1878 | virtual bool isHighLatencyDef(int opc) const { return false; } |
| 1879 | |
| 1880 | /// Compute operand latency between a def of 'Reg' |
| 1881 | /// and a use in the current loop. Return true if the target considered |
| 1882 | /// it 'high'. This is used by optimization passes such as machine LICM to |
| 1883 | /// determine whether it makes sense to hoist an instruction out even in a |
| 1884 | /// high register pressure situation. |
| 1885 | virtual bool hasHighOperandLatency(const TargetSchedModel &SchedModel, |
| 1886 | const MachineRegisterInfo *MRI, |
| 1887 | const MachineInstr &DefMI, unsigned DefIdx, |
| 1888 | const MachineInstr &UseMI, |
| 1889 | unsigned UseIdx) const { |
| 1890 | return false; |
| 1891 | } |
| 1892 | |
| 1893 | /// Compute operand latency of a def of 'Reg'. Return true |
| 1894 | /// if the target considered it 'low'. |
| 1895 | virtual bool hasLowDefLatency(const TargetSchedModel &SchedModel, |
| 1896 | const MachineInstr &DefMI, |
| 1897 | unsigned DefIdx) const; |
| 1898 | |
| 1899 | /// Perform target-specific instruction verification. |
| 1900 | virtual bool verifyInstruction(const MachineInstr &MI, |
| 1901 | StringRef &ErrInfo) const { |
| 1902 | return true; |
| 1903 | } |
| 1904 | |
| 1905 | /// Return the current execution domain and bit mask of |
| 1906 | /// possible domains for instruction. |
| 1907 | /// |
| 1908 | /// Some micro-architectures have multiple execution domains, and multiple |
| 1909 | /// opcodes that perform the same operation in different domains. For |
| 1910 | /// example, the x86 architecture provides the por, orps, and orpd |
| 1911 | /// instructions that all do the same thing. There is a latency penalty if a |
| 1912 | /// register is written in one domain and read in another. |
| 1913 | /// |
| 1914 | /// This function returns a pair (domain, mask) containing the execution |
| 1915 | /// domain of MI, and a bit mask of possible domains. The setExecutionDomain |
| 1916 | /// function can be used to change the opcode to one of the domains in the |
| 1917 | /// bit mask. Instructions whose execution domain can't be changed should |
| 1918 | /// return a 0 mask. |
| 1919 | /// |
| 1920 | /// The execution domain numbers don't have any special meaning except domain |
| 1921 | /// 0 is used for instructions that are not associated with any interesting |
| 1922 | /// execution domain. |
| 1923 | /// |
| 1924 | virtual std::pair<uint16_t, uint16_t> |
| 1925 | getExecutionDomain(const MachineInstr &MI) const { |
| 1926 | return std::make_pair(x: 0, y: 0); |
| 1927 | } |
| 1928 | |
| 1929 | /// Change the opcode of MI to execute in Domain. |
| 1930 | /// |
| 1931 | /// The bit (1 << Domain) must be set in the mask returned from |
| 1932 | /// getExecutionDomain(MI). |
| 1933 | virtual void setExecutionDomain(MachineInstr &MI, unsigned Domain) const {} |
| 1934 | |
| 1935 | /// Returns the preferred minimum clearance |
| 1936 | /// before an instruction with an unwanted partial register update. |
| 1937 | /// |
| 1938 | /// Some instructions only write part of a register, and implicitly need to |
| 1939 | /// read the other parts of the register. This may cause unwanted stalls |
| 1940 | /// preventing otherwise unrelated instructions from executing in parallel in |
| 1941 | /// an out-of-order CPU. |
| 1942 | /// |
| 1943 | /// For example, the x86 instruction cvtsi2ss writes its result to bits |
| 1944 | /// [31:0] of the destination xmm register. Bits [127:32] are unaffected, so |
| 1945 | /// the instruction needs to wait for the old value of the register to become |
| 1946 | /// available: |
| 1947 | /// |
| 1948 | /// addps %xmm1, %xmm0 |
| 1949 | /// movaps %xmm0, (%rax) |
| 1950 | /// cvtsi2ss %rbx, %xmm0 |
| 1951 | /// |
| 1952 | /// In the code above, the cvtsi2ss instruction needs to wait for the addps |
| 1953 | /// instruction before it can issue, even though the high bits of %xmm0 |
| 1954 | /// probably aren't needed. |
| 1955 | /// |
| 1956 | /// This hook returns the preferred clearance before MI, measured in |
| 1957 | /// instructions. Other defs of MI's operand OpNum are avoided in the last N |
| 1958 | /// instructions before MI. It should only return a positive value for |
| 1959 | /// unwanted dependencies. If the old bits of the defined register have |
| 1960 | /// useful values, or if MI is determined to otherwise read the dependency, |
| 1961 | /// the hook should return 0. |
| 1962 | /// |
| 1963 | /// The unwanted dependency may be handled by: |
| 1964 | /// |
| 1965 | /// 1. Allocating the same register for an MI def and use. That makes the |
| 1966 | /// unwanted dependency identical to a required dependency. |
| 1967 | /// |
| 1968 | /// 2. Allocating a register for the def that has no defs in the previous N |
| 1969 | /// instructions. |
| 1970 | /// |
| 1971 | /// 3. Calling breakPartialRegDependency() with the same arguments. This |
| 1972 | /// allows the target to insert a dependency breaking instruction. |
| 1973 | /// |
| 1974 | virtual unsigned |
| 1975 | getPartialRegUpdateClearance(const MachineInstr &MI, unsigned OpNum, |
| 1976 | const TargetRegisterInfo *TRI) const { |
| 1977 | // The default implementation returns 0 for no partial register dependency. |
| 1978 | return 0; |
| 1979 | } |
| 1980 | |
| 1981 | /// Return the minimum clearance before an instruction that reads an |
| 1982 | /// unused register. |
| 1983 | /// |
| 1984 | /// For example, AVX instructions may copy part of a register operand into |
| 1985 | /// the unused high bits of the destination register. |
| 1986 | /// |
| 1987 | /// vcvtsi2sdq %rax, undef %xmm0, %xmm14 |
| 1988 | /// |
| 1989 | /// In the code above, vcvtsi2sdq copies %xmm0[127:64] into %xmm14 creating a |
| 1990 | /// false dependence on any previous write to %xmm0. |
| 1991 | /// |
| 1992 | /// This hook works similarly to getPartialRegUpdateClearance, except that it |
| 1993 | /// does not take an operand index. Instead sets \p OpNum to the index of the |
| 1994 | /// unused register. |
| 1995 | virtual unsigned getUndefRegClearance(const MachineInstr &MI, unsigned OpNum, |
| 1996 | const TargetRegisterInfo *TRI) const { |
| 1997 | // The default implementation returns 0 for no undef register dependency. |
| 1998 | return 0; |
| 1999 | } |
| 2000 | |
| 2001 | /// Insert a dependency-breaking instruction |
| 2002 | /// before MI to eliminate an unwanted dependency on OpNum. |
| 2003 | /// |
| 2004 | /// If it wasn't possible to avoid a def in the last N instructions before MI |
| 2005 | /// (see getPartialRegUpdateClearance), this hook will be called to break the |
| 2006 | /// unwanted dependency. |
| 2007 | /// |
| 2008 | /// On x86, an xorps instruction can be used as a dependency breaker: |
| 2009 | /// |
| 2010 | /// addps %xmm1, %xmm0 |
| 2011 | /// movaps %xmm0, (%rax) |
| 2012 | /// xorps %xmm0, %xmm0 |
| 2013 | /// cvtsi2ss %rbx, %xmm0 |
| 2014 | /// |
| 2015 | /// An <imp-kill> operand should be added to MI if an instruction was |
| 2016 | /// inserted. This ties the instructions together in the post-ra scheduler. |
| 2017 | /// |
| 2018 | virtual void breakPartialRegDependency(MachineInstr &MI, unsigned OpNum, |
| 2019 | const TargetRegisterInfo *TRI) const {} |
| 2020 | |
| 2021 | /// Create machine specific model for scheduling. |
| 2022 | virtual DFAPacketizer * |
| 2023 | CreateTargetScheduleState(const TargetSubtargetInfo &) const { |
| 2024 | return nullptr; |
| 2025 | } |
| 2026 | |
| 2027 | /// Sometimes, it is possible for the target |
| 2028 | /// to tell, even without aliasing information, that two MIs access different |
| 2029 | /// memory addresses. This function returns true if two MIs access different |
| 2030 | /// memory addresses and false otherwise. |
| 2031 | /// |
| 2032 | /// Assumes any physical registers used to compute addresses have the same |
| 2033 | /// value for both instructions. (This is the most useful assumption for |
| 2034 | /// post-RA scheduling.) |
| 2035 | /// |
| 2036 | /// See also MachineInstr::mayAlias, which is implemented on top of this |
| 2037 | /// function. |
| 2038 | virtual bool |
| 2039 | areMemAccessesTriviallyDisjoint(const MachineInstr &MIa, |
| 2040 | const MachineInstr &MIb) const { |
| 2041 | assert(MIa.mayLoadOrStore() && |
| 2042 | "MIa must load from or modify a memory location" ); |
| 2043 | assert(MIb.mayLoadOrStore() && |
| 2044 | "MIb must load from or modify a memory location" ); |
| 2045 | return false; |
| 2046 | } |
| 2047 | |
| 2048 | /// Return the value to use for the MachineCSE's LookAheadLimit, |
| 2049 | /// which is a heuristic used for CSE'ing phys reg defs. |
| 2050 | virtual unsigned getMachineCSELookAheadLimit() const { |
| 2051 | // The default lookahead is small to prevent unprofitable quadratic |
| 2052 | // behavior. |
| 2053 | return 5; |
| 2054 | } |
| 2055 | |
| 2056 | /// Return the maximal number of alias checks on memory operands. For |
| 2057 | /// instructions with more than one memory operands, the alias check on a |
| 2058 | /// single MachineInstr pair has quadratic overhead and results in |
| 2059 | /// unacceptable performance in the worst case. The limit here is to clamp |
| 2060 | /// that maximal checks performed. Usually, that's the product of memory |
| 2061 | /// operand numbers from that pair of MachineInstr to be checked. For |
| 2062 | /// instance, with two MachineInstrs with 4 and 5 memory operands |
| 2063 | /// correspondingly, a total of 20 checks are required. With this limit set to |
| 2064 | /// 16, their alias check is skipped. We choose to limit the product instead |
| 2065 | /// of the individual instruction as targets may have special MachineInstrs |
| 2066 | /// with a considerably high number of memory operands, such as `ldm` in ARM. |
| 2067 | /// Setting this limit per MachineInstr would result in either too high |
| 2068 | /// overhead or too rigid restriction. |
| 2069 | virtual unsigned getMemOperandAACheckLimit() const { return 16; } |
| 2070 | |
| 2071 | /// Return an array that contains the ids of the target indices (used for the |
| 2072 | /// TargetIndex machine operand) and their names. |
| 2073 | /// |
| 2074 | /// MIR Serialization is able to serialize only the target indices that are |
| 2075 | /// defined by this method. |
| 2076 | virtual ArrayRef<std::pair<int, const char *>> |
| 2077 | getSerializableTargetIndices() const { |
| 2078 | return {}; |
| 2079 | } |
| 2080 | |
| 2081 | /// Decompose the machine operand's target flags into two values - the direct |
| 2082 | /// target flag value and any of bit flags that are applied. |
| 2083 | virtual std::pair<unsigned, unsigned> |
| 2084 | decomposeMachineOperandsTargetFlags(unsigned /*TF*/) const { |
| 2085 | return std::make_pair(x: 0u, y: 0u); |
| 2086 | } |
| 2087 | |
| 2088 | /// Return an array that contains the direct target flag values and their |
| 2089 | /// names. |
| 2090 | /// |
| 2091 | /// MIR Serialization is able to serialize only the target flags that are |
| 2092 | /// defined by this method. |
| 2093 | virtual ArrayRef<std::pair<unsigned, const char *>> |
| 2094 | getSerializableDirectMachineOperandTargetFlags() const { |
| 2095 | return {}; |
| 2096 | } |
| 2097 | |
| 2098 | /// Return an array that contains the bitmask target flag values and their |
| 2099 | /// names. |
| 2100 | /// |
| 2101 | /// MIR Serialization is able to serialize only the target flags that are |
| 2102 | /// defined by this method. |
| 2103 | virtual ArrayRef<std::pair<unsigned, const char *>> |
| 2104 | getSerializableBitmaskMachineOperandTargetFlags() const { |
| 2105 | return {}; |
| 2106 | } |
| 2107 | |
| 2108 | /// Return an array that contains the MMO target flag values and their |
| 2109 | /// names. |
| 2110 | /// |
| 2111 | /// MIR Serialization is able to serialize only the MMO target flags that are |
| 2112 | /// defined by this method. |
| 2113 | virtual ArrayRef<std::pair<MachineMemOperand::Flags, const char *>> |
| 2114 | getSerializableMachineMemOperandTargetFlags() const { |
| 2115 | return {}; |
| 2116 | } |
| 2117 | |
| 2118 | /// Determines whether \p Inst is a tail call instruction. Override this |
| 2119 | /// method on targets that do not properly set MCID::Return and MCID::Call on |
| 2120 | /// tail call instructions." |
| 2121 | virtual bool isTailCall(const MachineInstr &Inst) const { |
| 2122 | return Inst.isReturn() && Inst.isCall(); |
| 2123 | } |
| 2124 | |
| 2125 | /// True if the instruction is bound to the top of its basic block and no |
| 2126 | /// other instructions shall be inserted before it. This can be implemented |
| 2127 | /// to prevent register allocator to insert spills for \p Reg before such |
| 2128 | /// instructions. |
| 2129 | virtual bool isBasicBlockPrologue(const MachineInstr &MI, |
| 2130 | Register Reg = Register()) const { |
| 2131 | return false; |
| 2132 | } |
| 2133 | |
| 2134 | /// Allows targets to use appropriate copy instruction while spilitting live |
| 2135 | /// range of a register in register allocation. |
| 2136 | virtual unsigned getLiveRangeSplitOpcode(Register Reg, |
| 2137 | const MachineFunction &MF) const { |
| 2138 | return TargetOpcode::COPY; |
| 2139 | } |
| 2140 | |
| 2141 | /// During PHI eleimination lets target to make necessary checks and |
| 2142 | /// insert the copy to the PHI destination register in a target specific |
| 2143 | /// manner. |
| 2144 | virtual MachineInstr *createPHIDestinationCopy( |
| 2145 | MachineBasicBlock &MBB, MachineBasicBlock::iterator InsPt, |
| 2146 | const DebugLoc &DL, Register Src, Register Dst) const { |
| 2147 | return BuildMI(BB&: MBB, I: InsPt, MIMD: DL, MCID: get(Opcode: TargetOpcode::COPY), DestReg: Dst) |
| 2148 | .addReg(RegNo: Src); |
| 2149 | } |
| 2150 | |
| 2151 | /// During PHI eleimination lets target to make necessary checks and |
| 2152 | /// insert the copy to the PHI destination register in a target specific |
| 2153 | /// manner. |
| 2154 | virtual MachineInstr *createPHISourceCopy(MachineBasicBlock &MBB, |
| 2155 | MachineBasicBlock::iterator InsPt, |
| 2156 | const DebugLoc &DL, Register Src, |
| 2157 | unsigned SrcSubReg, |
| 2158 | Register Dst) const { |
| 2159 | return BuildMI(BB&: MBB, I: InsPt, MIMD: DL, MCID: get(Opcode: TargetOpcode::COPY), DestReg: Dst) |
| 2160 | .addReg(RegNo: Src, flags: 0, SubReg: SrcSubReg); |
| 2161 | } |
| 2162 | |
| 2163 | /// Returns a \p outliner::OutlinedFunction struct containing target-specific |
| 2164 | /// information for a set of outlining candidates. Returns std::nullopt if the |
| 2165 | /// candidates are not suitable for outlining. \p MinRepeats is the minimum |
| 2166 | /// number of times the instruction sequence must be repeated. |
| 2167 | virtual std::optional<std::unique_ptr<outliner::OutlinedFunction>> |
| 2168 | getOutliningCandidateInfo( |
| 2169 | const MachineModuleInfo &MMI, |
| 2170 | std::vector<outliner::Candidate> &RepeatedSequenceLocs, |
| 2171 | unsigned MinRepeats) const { |
| 2172 | llvm_unreachable( |
| 2173 | "Target didn't implement TargetInstrInfo::getOutliningCandidateInfo!" ); |
| 2174 | } |
| 2175 | |
| 2176 | /// Optional target hook to create the LLVM IR attributes for the outlined |
| 2177 | /// function. If overridden, the overriding function must call the default |
| 2178 | /// implementation. |
| 2179 | virtual void mergeOutliningCandidateAttributes( |
| 2180 | Function &F, std::vector<outliner::Candidate> &Candidates) const; |
| 2181 | |
| 2182 | protected: |
| 2183 | /// Target-dependent implementation for getOutliningTypeImpl. |
| 2184 | virtual outliner::InstrType |
| 2185 | getOutliningTypeImpl(const MachineModuleInfo &MMI, |
| 2186 | MachineBasicBlock::iterator &MIT, unsigned Flags) const { |
| 2187 | llvm_unreachable( |
| 2188 | "Target didn't implement TargetInstrInfo::getOutliningTypeImpl!" ); |
| 2189 | } |
| 2190 | |
| 2191 | public: |
| 2192 | /// Returns how or if \p MIT should be outlined. \p Flags is the |
| 2193 | /// target-specific information returned by isMBBSafeToOutlineFrom. |
| 2194 | outliner::InstrType getOutliningType(const MachineModuleInfo &MMI, |
| 2195 | MachineBasicBlock::iterator &MIT, |
| 2196 | unsigned Flags) const; |
| 2197 | |
| 2198 | /// Optional target hook that returns true if \p MBB is safe to outline from, |
| 2199 | /// and returns any target-specific information in \p Flags. |
| 2200 | virtual bool isMBBSafeToOutlineFrom(MachineBasicBlock &MBB, |
| 2201 | unsigned &Flags) const; |
| 2202 | |
| 2203 | /// Optional target hook which partitions \p MBB into outlinable ranges for |
| 2204 | /// instruction mapping purposes. Each range is defined by two iterators: |
| 2205 | /// [start, end). |
| 2206 | /// |
| 2207 | /// Ranges are expected to be ordered top-down. That is, ranges closer to the |
| 2208 | /// top of the block should come before ranges closer to the end of the block. |
| 2209 | /// |
| 2210 | /// Ranges cannot overlap. |
| 2211 | /// |
| 2212 | /// If an entire block is mappable, then its range is [MBB.begin(), MBB.end()) |
| 2213 | /// |
| 2214 | /// All instructions not present in an outlinable range are considered |
| 2215 | /// illegal. |
| 2216 | virtual SmallVector< |
| 2217 | std::pair<MachineBasicBlock::iterator, MachineBasicBlock::iterator>> |
| 2218 | getOutlinableRanges(MachineBasicBlock &MBB, unsigned &Flags) const { |
| 2219 | return {std::make_pair(x: MBB.begin(), y: MBB.end())}; |
| 2220 | } |
| 2221 | |
| 2222 | /// Insert a custom frame for outlined functions. |
| 2223 | virtual void buildOutlinedFrame(MachineBasicBlock &MBB, MachineFunction &MF, |
| 2224 | const outliner::OutlinedFunction &OF) const { |
| 2225 | llvm_unreachable( |
| 2226 | "Target didn't implement TargetInstrInfo::buildOutlinedFrame!" ); |
| 2227 | } |
| 2228 | |
| 2229 | /// Insert a call to an outlined function into the program. |
| 2230 | /// Returns an iterator to the spot where we inserted the call. This must be |
| 2231 | /// implemented by the target. |
| 2232 | virtual MachineBasicBlock::iterator |
| 2233 | insertOutlinedCall(Module &M, MachineBasicBlock &MBB, |
| 2234 | MachineBasicBlock::iterator &It, MachineFunction &MF, |
| 2235 | outliner::Candidate &C) const { |
| 2236 | llvm_unreachable( |
| 2237 | "Target didn't implement TargetInstrInfo::insertOutlinedCall!" ); |
| 2238 | } |
| 2239 | |
| 2240 | /// Insert an architecture-specific instruction to clear a register. If you |
| 2241 | /// need to avoid sideeffects (e.g. avoid XOR on x86, which sets EFLAGS), set |
| 2242 | /// \p AllowSideEffects to \p false. |
| 2243 | virtual void buildClearRegister(Register Reg, MachineBasicBlock &MBB, |
| 2244 | MachineBasicBlock::iterator Iter, |
| 2245 | DebugLoc &DL, |
| 2246 | bool AllowSideEffects = true) const { |
| 2247 | #if 0 |
| 2248 | // FIXME: This should exist once all platforms that use stack protectors |
| 2249 | // implements it. |
| 2250 | llvm_unreachable( |
| 2251 | "Target didn't implement TargetInstrInfo::buildClearRegister!" ); |
| 2252 | #endif |
| 2253 | } |
| 2254 | |
| 2255 | /// Return true if the function can safely be outlined from. |
| 2256 | /// A function \p MF is considered safe for outlining if an outlined function |
| 2257 | /// produced from instructions in F will produce a program which produces the |
| 2258 | /// same output for any set of given inputs. |
| 2259 | virtual bool isFunctionSafeToOutlineFrom(MachineFunction &MF, |
| 2260 | bool OutlineFromLinkOnceODRs) const { |
| 2261 | llvm_unreachable("Target didn't implement " |
| 2262 | "TargetInstrInfo::isFunctionSafeToOutlineFrom!" ); |
| 2263 | } |
| 2264 | |
| 2265 | /// Return true if the function should be outlined from by default. |
| 2266 | virtual bool shouldOutlineFromFunctionByDefault(MachineFunction &MF) const { |
| 2267 | return false; |
| 2268 | } |
| 2269 | |
| 2270 | /// Return true if the function is a viable candidate for machine function |
| 2271 | /// splitting. The criteria for if a function can be split may vary by target. |
| 2272 | virtual bool isFunctionSafeToSplit(const MachineFunction &MF) const; |
| 2273 | |
| 2274 | /// Return true if the MachineBasicBlock can safely be split to the cold |
| 2275 | /// section. On AArch64, certain instructions may cause a block to be unsafe |
| 2276 | /// to split to the cold section. |
| 2277 | virtual bool isMBBSafeToSplitToCold(const MachineBasicBlock &MBB) const { |
| 2278 | return true; |
| 2279 | } |
| 2280 | |
| 2281 | /// Produce the expression describing the \p MI loading a value into |
| 2282 | /// the physical register \p Reg. This hook should only be used with |
| 2283 | /// \p MIs belonging to VReg-less functions. |
| 2284 | virtual std::optional<ParamLoadedValue> |
| 2285 | describeLoadedValue(const MachineInstr &MI, Register Reg) const; |
| 2286 | |
| 2287 | /// Given the generic extension instruction \p ExtMI, returns true if this |
| 2288 | /// extension is a likely candidate for being folded into an another |
| 2289 | /// instruction. |
| 2290 | virtual bool isExtendLikelyToBeFolded(MachineInstr &ExtMI, |
| 2291 | MachineRegisterInfo &MRI) const { |
| 2292 | return false; |
| 2293 | } |
| 2294 | |
| 2295 | /// Return MIR formatter to format/parse MIR operands. Target can override |
| 2296 | /// this virtual function and return target specific MIR formatter. |
| 2297 | virtual const MIRFormatter *getMIRFormatter() const { |
| 2298 | if (!Formatter) |
| 2299 | Formatter = std::make_unique<MIRFormatter>(); |
| 2300 | return Formatter.get(); |
| 2301 | } |
| 2302 | |
| 2303 | /// Returns the target-specific default value for tail duplication. |
| 2304 | /// This value will be used if the tail-dup-placement-threshold argument is |
| 2305 | /// not provided. |
| 2306 | virtual unsigned getTailDuplicateSize(CodeGenOptLevel OptLevel) const { |
| 2307 | return OptLevel >= CodeGenOptLevel::Aggressive ? 4 : 2; |
| 2308 | } |
| 2309 | |
| 2310 | /// Returns the target-specific default value for tail merging. |
| 2311 | /// This value will be used if the tail-merge-size argument is not provided. |
| 2312 | virtual unsigned getTailMergeSize(const MachineFunction &MF) const { |
| 2313 | return 3; |
| 2314 | } |
| 2315 | |
| 2316 | /// Returns the callee operand from the given \p MI. |
| 2317 | virtual const MachineOperand &getCalleeOperand(const MachineInstr &MI) const { |
| 2318 | return MI.getOperand(i: 0); |
| 2319 | } |
| 2320 | |
| 2321 | /// Return the uniformity behavior of the given instruction. |
| 2322 | virtual InstructionUniformity |
| 2323 | getInstructionUniformity(const MachineInstr &MI) const { |
| 2324 | return InstructionUniformity::Default; |
| 2325 | } |
| 2326 | |
| 2327 | /// Returns true if the given \p MI defines a TargetIndex operand that can be |
| 2328 | /// tracked by their offset, can have values, and can have debug info |
| 2329 | /// associated with it. If so, sets \p Index and \p Offset of the target index |
| 2330 | /// operand. |
| 2331 | virtual bool isExplicitTargetIndexDef(const MachineInstr &MI, int &Index, |
| 2332 | int64_t &Offset) const { |
| 2333 | return false; |
| 2334 | } |
| 2335 | |
| 2336 | // Get the call frame size just before MI. |
| 2337 | unsigned getCallFrameSizeAt(MachineInstr &MI) const; |
| 2338 | |
| 2339 | /// Fills in the necessary MachineOperands to refer to a frame index. |
| 2340 | /// The best way to understand this is to print `asm(""::"m"(x));` after |
| 2341 | /// finalize-isel. Example: |
| 2342 | /// INLINEASM ... 262190 /* mem:m */, %stack.0.x.addr, 1, $noreg, 0, $noreg |
| 2343 | /// we would add placeholders for: ^ ^ ^ ^ |
| 2344 | virtual void getFrameIndexOperands(SmallVectorImpl<MachineOperand> &Ops, |
| 2345 | int FI) const { |
| 2346 | llvm_unreachable("unknown number of operands necessary" ); |
| 2347 | } |
| 2348 | |
| 2349 | private: |
| 2350 | mutable std::unique_ptr<MIRFormatter> Formatter; |
| 2351 | unsigned CallFrameSetupOpcode, CallFrameDestroyOpcode; |
| 2352 | unsigned CatchRetOpcode; |
| 2353 | unsigned ReturnOpcode; |
| 2354 | }; |
| 2355 | |
| 2356 | /// Provide DenseMapInfo for TargetInstrInfo::RegSubRegPair. |
| 2357 | template <> struct DenseMapInfo<TargetInstrInfo::RegSubRegPair> { |
| 2358 | using RegInfo = DenseMapInfo<Register>; |
| 2359 | using SubRegInfo = DenseMapInfo<unsigned>; |
| 2360 | |
| 2361 | static inline TargetInstrInfo::RegSubRegPair getEmptyKey() { |
| 2362 | return TargetInstrInfo::RegSubRegPair(RegInfo::getEmptyKey(), |
| 2363 | SubRegInfo::getEmptyKey()); |
| 2364 | } |
| 2365 | |
| 2366 | static inline TargetInstrInfo::RegSubRegPair getTombstoneKey() { |
| 2367 | return TargetInstrInfo::RegSubRegPair(RegInfo::getTombstoneKey(), |
| 2368 | SubRegInfo::getTombstoneKey()); |
| 2369 | } |
| 2370 | |
| 2371 | /// Reuse getHashValue implementation from |
| 2372 | /// std::pair<unsigned, unsigned>. |
| 2373 | static unsigned getHashValue(const TargetInstrInfo::RegSubRegPair &Val) { |
| 2374 | return DenseMapInfo<std::pair<Register, unsigned>>::getHashValue( |
| 2375 | PairVal: std::make_pair(x: Val.Reg, y: Val.SubReg)); |
| 2376 | } |
| 2377 | |
| 2378 | static bool isEqual(const TargetInstrInfo::RegSubRegPair &LHS, |
| 2379 | const TargetInstrInfo::RegSubRegPair &RHS) { |
| 2380 | return LHS == RHS; |
| 2381 | } |
| 2382 | }; |
| 2383 | |
| 2384 | } // end namespace llvm |
| 2385 | |
| 2386 | #endif // LLVM_CODEGEN_TARGETINSTRINFO_H |
| 2387 | |