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_TARGET_TARGETINSTRINFO_H 14 #define LLVM_TARGET_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/None.h" 20 #include "llvm/CodeGen/LiveRegUnits.h" 21 #include "llvm/CodeGen/MIRFormatter.h" 22 #include "llvm/CodeGen/MachineBasicBlock.h" 23 #include "llvm/CodeGen/MachineCombinerPattern.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/PseudoSourceValue.h" 30 #include "llvm/CodeGen/VirtRegMap.h" 31 #include "llvm/MC/MCInstrInfo.h" 32 #include "llvm/Support/BranchProbability.h" 33 #include "llvm/Support/ErrorHandling.h" 34 #include <cassert> 35 #include <cstddef> 36 #include <cstdint> 37 #include <utility> 38 #include <vector> 39 40 namespace llvm { 41 42 class AAResults; 43 class DFAPacketizer; 44 class InstrItineraryData; 45 class LiveIntervals; 46 class LiveVariables; 47 class MachineLoop; 48 class MachineMemOperand; 49 class MachineRegisterInfo; 50 class MCAsmInfo; 51 class MCInst; 52 struct MCSchedModel; 53 class Module; 54 class ScheduleDAG; 55 class ScheduleDAGMI; 56 class ScheduleHazardRecognizer; 57 class SDNode; 58 class SelectionDAG; 59 class RegScavenger; 60 class TargetRegisterClass; 61 class TargetRegisterInfo; 62 class TargetSchedModel; 63 class TargetSubtargetInfo; 64 65 template <class T> class SmallVectorImpl; 66 67 using ParamLoadedValue = std::pair<MachineOperand, DIExpression*>; 68 69 struct DestSourcePair { 70 const MachineOperand *Destination; 71 const MachineOperand *Source; 72 DestSourcePairDestSourcePair73 DestSourcePair(const MachineOperand &Dest, const MachineOperand &Src) 74 : Destination(&Dest), Source(&Src) {} 75 }; 76 77 /// Used to describe a register and immediate addition. 78 struct RegImmPair { 79 Register Reg; 80 int64_t Imm; 81 RegImmPairRegImmPair82 RegImmPair(Register Reg, int64_t Imm) : Reg(Reg), Imm(Imm) {} 83 }; 84 85 //--------------------------------------------------------------------------- 86 /// 87 /// TargetInstrInfo - Interface to description of machine instruction set 88 /// 89 class TargetInstrInfo : public MCInstrInfo { 90 public: 91 TargetInstrInfo(unsigned CFSetupOpcode = ~0u, unsigned CFDestroyOpcode = ~0u, 92 unsigned CatchRetOpcode = ~0u, unsigned ReturnOpcode = ~0u) CallFrameSetupOpcode(CFSetupOpcode)93 : CallFrameSetupOpcode(CFSetupOpcode), 94 CallFrameDestroyOpcode(CFDestroyOpcode), CatchRetOpcode(CatchRetOpcode), 95 ReturnOpcode(ReturnOpcode) {} 96 TargetInstrInfo(const TargetInstrInfo &) = delete; 97 TargetInstrInfo &operator=(const TargetInstrInfo &) = delete; 98 virtual ~TargetInstrInfo(); 99 isGenericOpcode(unsigned Opc)100 static bool isGenericOpcode(unsigned Opc) { 101 return Opc <= TargetOpcode::GENERIC_OP_END; 102 } 103 104 /// Given a machine instruction descriptor, returns the register 105 /// class constraint for OpNum, or NULL. 106 virtual 107 const TargetRegisterClass *getRegClass(const MCInstrDesc &MCID, unsigned OpNum, 108 const TargetRegisterInfo *TRI, 109 const MachineFunction &MF) const; 110 111 /// Return true if the instruction is trivially rematerializable, meaning it 112 /// has no side effects and requires no operands that aren't always available. 113 /// This means the only allowed uses are constants and unallocatable physical 114 /// registers so that the instructions result is independent of the place 115 /// in the function. 116 bool isTriviallyReMaterializable(const MachineInstr &MI, 117 AAResults *AA = nullptr) const { 118 return MI.getOpcode() == TargetOpcode::IMPLICIT_DEF || 119 (MI.getDesc().isRematerializable() && 120 (isReallyTriviallyReMaterializable(MI, AA) || 121 isReallyTriviallyReMaterializableGeneric(MI, AA))); 122 } 123 124 protected: 125 /// For instructions with opcodes for which the M_REMATERIALIZABLE flag is 126 /// set, this hook lets the target specify whether the instruction is actually 127 /// trivially rematerializable, taking into consideration its operands. This 128 /// predicate must return false if the instruction has any side effects other 129 /// than producing a value, or if it requres any address registers that are 130 /// not always available. 131 /// Requirements must be check as stated in isTriviallyReMaterializable() . isReallyTriviallyReMaterializable(const MachineInstr & MI,AAResults * AA)132 virtual bool isReallyTriviallyReMaterializable(const MachineInstr &MI, 133 AAResults *AA) const { 134 return false; 135 } 136 137 /// This method commutes the operands of the given machine instruction MI. 138 /// The operands to be commuted are specified by their indices OpIdx1 and 139 /// OpIdx2. 140 /// 141 /// If a target has any instructions that are commutable but require 142 /// converting to different instructions or making non-trivial changes 143 /// to commute them, this method can be overloaded to do that. 144 /// The default implementation simply swaps the commutable operands. 145 /// 146 /// If NewMI is false, MI is modified in place and returned; otherwise, a 147 /// new machine instruction is created and returned. 148 /// 149 /// Do not call this method for a non-commutable instruction. 150 /// Even though the instruction is commutable, the method may still 151 /// fail to commute the operands, null pointer is returned in such cases. 152 virtual MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI, 153 unsigned OpIdx1, 154 unsigned OpIdx2) const; 155 156 /// Assigns the (CommutableOpIdx1, CommutableOpIdx2) pair of commutable 157 /// operand indices to (ResultIdx1, ResultIdx2). 158 /// One or both input values of the pair: (ResultIdx1, ResultIdx2) may be 159 /// predefined to some indices or be undefined (designated by the special 160 /// value 'CommuteAnyOperandIndex'). 161 /// The predefined result indices cannot be re-defined. 162 /// The function returns true iff after the result pair redefinition 163 /// the fixed result pair is equal to or equivalent to the source pair of 164 /// indices: (CommutableOpIdx1, CommutableOpIdx2). It is assumed here that 165 /// the pairs (x,y) and (y,x) are equivalent. 166 static bool fixCommutedOpIndices(unsigned &ResultIdx1, unsigned &ResultIdx2, 167 unsigned CommutableOpIdx1, 168 unsigned CommutableOpIdx2); 169 170 private: 171 /// For instructions with opcodes for which the M_REMATERIALIZABLE flag is 172 /// set and the target hook isReallyTriviallyReMaterializable returns false, 173 /// this function does target-independent tests to determine if the 174 /// instruction is really trivially rematerializable. 175 bool isReallyTriviallyReMaterializableGeneric(const MachineInstr &MI, 176 AAResults *AA) const; 177 178 public: 179 /// These methods return the opcode of the frame setup/destroy instructions 180 /// if they exist (-1 otherwise). Some targets use pseudo instructions in 181 /// order to abstract away the difference between operating with a frame 182 /// pointer and operating without, through the use of these two instructions. 183 /// getCallFrameSetupOpcode()184 unsigned getCallFrameSetupOpcode() const { return CallFrameSetupOpcode; } getCallFrameDestroyOpcode()185 unsigned getCallFrameDestroyOpcode() const { return CallFrameDestroyOpcode; } 186 187 /// Returns true if the argument is a frame pseudo instruction. isFrameInstr(const MachineInstr & I)188 bool isFrameInstr(const MachineInstr &I) const { 189 return I.getOpcode() == getCallFrameSetupOpcode() || 190 I.getOpcode() == getCallFrameDestroyOpcode(); 191 } 192 193 /// Returns true if the argument is a frame setup pseudo instruction. isFrameSetup(const MachineInstr & I)194 bool isFrameSetup(const MachineInstr &I) const { 195 return I.getOpcode() == getCallFrameSetupOpcode(); 196 } 197 198 /// Returns size of the frame associated with the given frame instruction. 199 /// For frame setup instruction this is frame that is set up space set up 200 /// after the instruction. For frame destroy instruction this is the frame 201 /// freed by the caller. 202 /// Note, in some cases a call frame (or a part of it) may be prepared prior 203 /// to the frame setup instruction. It occurs in the calls that involve 204 /// inalloca arguments. This function reports only the size of the frame part 205 /// that is set up between the frame setup and destroy pseudo instructions. getFrameSize(const MachineInstr & I)206 int64_t getFrameSize(const MachineInstr &I) const { 207 assert(isFrameInstr(I) && "Not a frame instruction"); 208 assert(I.getOperand(0).getImm() >= 0); 209 return I.getOperand(0).getImm(); 210 } 211 212 /// Returns the total frame size, which is made up of the space set up inside 213 /// the pair of frame start-stop instructions and the space that is set up 214 /// prior to the pair. getFrameTotalSize(const MachineInstr & I)215 int64_t getFrameTotalSize(const MachineInstr &I) const { 216 if (isFrameSetup(I)) { 217 assert(I.getOperand(1).getImm() >= 0 && 218 "Frame size must not be negative"); 219 return getFrameSize(I) + I.getOperand(1).getImm(); 220 } 221 return getFrameSize(I); 222 } 223 getCatchReturnOpcode()224 unsigned getCatchReturnOpcode() const { return CatchRetOpcode; } getReturnOpcode()225 unsigned getReturnOpcode() const { return ReturnOpcode; } 226 227 /// Returns the actual stack pointer adjustment made by an instruction 228 /// as part of a call sequence. By default, only call frame setup/destroy 229 /// instructions adjust the stack, but targets may want to override this 230 /// to enable more fine-grained adjustment, or adjust by a different value. 231 virtual int getSPAdjust(const MachineInstr &MI) const; 232 233 /// Return true if the instruction is a "coalescable" extension instruction. 234 /// That is, it's like a copy where it's legal for the source to overlap the 235 /// destination. e.g. X86::MOVSX64rr32. If this returns true, then it's 236 /// expected the pre-extension value is available as a subreg of the result 237 /// register. This also returns the sub-register index in SubIdx. isCoalescableExtInstr(const MachineInstr & MI,unsigned & SrcReg,unsigned & DstReg,unsigned & SubIdx)238 virtual bool isCoalescableExtInstr(const MachineInstr &MI, unsigned &SrcReg, 239 unsigned &DstReg, unsigned &SubIdx) const { 240 return false; 241 } 242 243 /// If the specified machine instruction is a direct 244 /// load from a stack slot, return the virtual or physical register number of 245 /// the destination along with the FrameIndex of the loaded stack slot. If 246 /// not, return 0. This predicate must return 0 if the instruction has 247 /// any side effects other than loading from the stack slot. isLoadFromStackSlot(const MachineInstr & MI,int & FrameIndex)248 virtual unsigned isLoadFromStackSlot(const MachineInstr &MI, 249 int &FrameIndex) const { 250 return 0; 251 } 252 253 /// Optional extension of isLoadFromStackSlot that returns the number of 254 /// bytes loaded from the stack. This must be implemented if a backend 255 /// supports partial stack slot spills/loads to further disambiguate 256 /// what the load does. isLoadFromStackSlot(const MachineInstr & MI,int & FrameIndex,unsigned & MemBytes)257 virtual unsigned isLoadFromStackSlot(const MachineInstr &MI, 258 int &FrameIndex, 259 unsigned &MemBytes) const { 260 MemBytes = 0; 261 return isLoadFromStackSlot(MI, FrameIndex); 262 } 263 264 /// Check for post-frame ptr elimination stack locations as well. 265 /// This uses a heuristic so it isn't reliable for correctness. isLoadFromStackSlotPostFE(const MachineInstr & MI,int & FrameIndex)266 virtual unsigned isLoadFromStackSlotPostFE(const MachineInstr &MI, 267 int &FrameIndex) const { 268 return 0; 269 } 270 271 /// If the specified machine instruction has a load from a stack slot, 272 /// return true along with the FrameIndices of the loaded stack slot and the 273 /// machine mem operands containing the reference. 274 /// If not, return false. Unlike isLoadFromStackSlot, this returns true for 275 /// any instructions that loads from the stack. This is just a hint, as some 276 /// cases may be missed. 277 virtual bool hasLoadFromStackSlot( 278 const MachineInstr &MI, 279 SmallVectorImpl<const MachineMemOperand *> &Accesses) const; 280 281 /// If the specified machine instruction is a direct 282 /// store to a stack slot, return the virtual or physical register number of 283 /// the source reg along with the FrameIndex of the loaded stack slot. If 284 /// not, return 0. This predicate must return 0 if the instruction has 285 /// any side effects other than storing to the stack slot. isStoreToStackSlot(const MachineInstr & MI,int & FrameIndex)286 virtual unsigned isStoreToStackSlot(const MachineInstr &MI, 287 int &FrameIndex) const { 288 return 0; 289 } 290 291 /// Optional extension of isStoreToStackSlot that returns the number of 292 /// bytes stored to the stack. This must be implemented if a backend 293 /// supports partial stack slot spills/loads to further disambiguate 294 /// what the store does. isStoreToStackSlot(const MachineInstr & MI,int & FrameIndex,unsigned & MemBytes)295 virtual unsigned isStoreToStackSlot(const MachineInstr &MI, 296 int &FrameIndex, 297 unsigned &MemBytes) const { 298 MemBytes = 0; 299 return isStoreToStackSlot(MI, FrameIndex); 300 } 301 302 /// Check for post-frame ptr elimination stack locations as well. 303 /// This uses a heuristic, so it isn't reliable for correctness. isStoreToStackSlotPostFE(const MachineInstr & MI,int & FrameIndex)304 virtual unsigned isStoreToStackSlotPostFE(const MachineInstr &MI, 305 int &FrameIndex) const { 306 return 0; 307 } 308 309 /// If the specified machine instruction has a store to a stack slot, 310 /// return true along with the FrameIndices of the loaded stack slot and the 311 /// machine mem operands containing the reference. 312 /// If not, return false. Unlike isStoreToStackSlot, 313 /// this returns true for any instructions that stores to the 314 /// stack. This is just a hint, as some cases may be missed. 315 virtual bool hasStoreToStackSlot( 316 const MachineInstr &MI, 317 SmallVectorImpl<const MachineMemOperand *> &Accesses) const; 318 319 /// Return true if the specified machine instruction 320 /// is a copy of one stack slot to another and has no other effect. 321 /// Provide the identity of the two frame indices. isStackSlotCopy(const MachineInstr & MI,int & DestFrameIndex,int & SrcFrameIndex)322 virtual bool isStackSlotCopy(const MachineInstr &MI, int &DestFrameIndex, 323 int &SrcFrameIndex) const { 324 return false; 325 } 326 327 /// Compute the size in bytes and offset within a stack slot of a spilled 328 /// register or subregister. 329 /// 330 /// \param [out] Size in bytes of the spilled value. 331 /// \param [out] Offset in bytes within the stack slot. 332 /// \returns true if both Size and Offset are successfully computed. 333 /// 334 /// Not all subregisters have computable spill slots. For example, 335 /// subregisters registers may not be byte-sized, and a pair of discontiguous 336 /// subregisters has no single offset. 337 /// 338 /// Targets with nontrivial bigendian implementations may need to override 339 /// this, particularly to support spilled vector registers. 340 virtual bool getStackSlotRange(const TargetRegisterClass *RC, unsigned SubIdx, 341 unsigned &Size, unsigned &Offset, 342 const MachineFunction &MF) const; 343 344 /// Returns the size in bytes of the specified MachineInstr, or ~0U 345 /// when this function is not implemented by a target. getInstSizeInBytes(const MachineInstr & MI)346 virtual unsigned getInstSizeInBytes(const MachineInstr &MI) const { 347 return ~0U; 348 } 349 350 /// Return true if the instruction is as cheap as a move instruction. 351 /// 352 /// Targets for different archs need to override this, and different 353 /// micro-architectures can also be finely tuned inside. isAsCheapAsAMove(const MachineInstr & MI)354 virtual bool isAsCheapAsAMove(const MachineInstr &MI) const { 355 return MI.isAsCheapAsAMove(); 356 } 357 358 /// Return true if the instruction should be sunk by MachineSink. 359 /// 360 /// MachineSink determines on its own whether the instruction is safe to sink; 361 /// this gives the target a hook to override the default behavior with regards 362 /// to which instructions should be sunk. shouldSink(const MachineInstr & MI)363 virtual bool shouldSink(const MachineInstr &MI) const { return true; } 364 365 /// Re-issue the specified 'original' instruction at the 366 /// specific location targeting a new destination register. 367 /// The register in Orig->getOperand(0).getReg() will be substituted by 368 /// DestReg:SubIdx. Any existing subreg index is preserved or composed with 369 /// SubIdx. 370 virtual void reMaterialize(MachineBasicBlock &MBB, 371 MachineBasicBlock::iterator MI, unsigned DestReg, 372 unsigned SubIdx, const MachineInstr &Orig, 373 const TargetRegisterInfo &TRI) const; 374 375 /// Clones instruction or the whole instruction bundle \p Orig and 376 /// insert into \p MBB before \p InsertBefore. The target may update operands 377 /// that are required to be unique. 378 /// 379 /// \p Orig must not return true for MachineInstr::isNotDuplicable(). 380 virtual MachineInstr &duplicate(MachineBasicBlock &MBB, 381 MachineBasicBlock::iterator InsertBefore, 382 const MachineInstr &Orig) const; 383 384 /// This method must be implemented by targets that 385 /// set the M_CONVERTIBLE_TO_3_ADDR flag. When this flag is set, the target 386 /// may be able to convert a two-address instruction into one or more true 387 /// three-address instructions on demand. This allows the X86 target (for 388 /// example) to convert ADD and SHL instructions into LEA instructions if they 389 /// would require register copies due to two-addressness. 390 /// 391 /// This method returns a null pointer if the transformation cannot be 392 /// performed, otherwise it returns the last new instruction. 393 /// convertToThreeAddress(MachineFunction::iterator & MFI,MachineInstr & MI,LiveVariables * LV)394 virtual MachineInstr *convertToThreeAddress(MachineFunction::iterator &MFI, 395 MachineInstr &MI, 396 LiveVariables *LV) const { 397 return nullptr; 398 } 399 400 // This constant can be used as an input value of operand index passed to 401 // the method findCommutedOpIndices() to tell the method that the 402 // corresponding operand index is not pre-defined and that the method 403 // can pick any commutable operand. 404 static const unsigned CommuteAnyOperandIndex = ~0U; 405 406 /// This method commutes the operands of the given machine instruction MI. 407 /// 408 /// The operands to be commuted are specified by their indices OpIdx1 and 409 /// OpIdx2. OpIdx1 and OpIdx2 arguments may be set to a special value 410 /// 'CommuteAnyOperandIndex', which means that the method is free to choose 411 /// any arbitrarily chosen commutable operand. If both arguments are set to 412 /// 'CommuteAnyOperandIndex' then the method looks for 2 different commutable 413 /// operands; then commutes them if such operands could be found. 414 /// 415 /// If NewMI is false, MI is modified in place and returned; otherwise, a 416 /// new machine instruction is created and returned. 417 /// 418 /// Do not call this method for a non-commutable instruction or 419 /// for non-commuable operands. 420 /// Even though the instruction is commutable, the method may still 421 /// fail to commute the operands, null pointer is returned in such cases. 422 MachineInstr * 423 commuteInstruction(MachineInstr &MI, bool NewMI = false, 424 unsigned OpIdx1 = CommuteAnyOperandIndex, 425 unsigned OpIdx2 = CommuteAnyOperandIndex) const; 426 427 /// Returns true iff the routine could find two commutable operands in the 428 /// given machine instruction. 429 /// The 'SrcOpIdx1' and 'SrcOpIdx2' are INPUT and OUTPUT arguments. 430 /// If any of the INPUT values is set to the special value 431 /// 'CommuteAnyOperandIndex' then the method arbitrarily picks a commutable 432 /// operand, then returns its index in the corresponding argument. 433 /// If both of INPUT values are set to 'CommuteAnyOperandIndex' then method 434 /// looks for 2 commutable operands. 435 /// If INPUT values refer to some operands of MI, then the method simply 436 /// returns true if the corresponding operands are commutable and returns 437 /// false otherwise. 438 /// 439 /// For example, calling this method this way: 440 /// unsigned Op1 = 1, Op2 = CommuteAnyOperandIndex; 441 /// findCommutedOpIndices(MI, Op1, Op2); 442 /// can be interpreted as a query asking to find an operand that would be 443 /// commutable with the operand#1. 444 virtual bool findCommutedOpIndices(const MachineInstr &MI, 445 unsigned &SrcOpIdx1, 446 unsigned &SrcOpIdx2) const; 447 448 /// A pair composed of a register and a sub-register index. 449 /// Used to give some type checking when modeling Reg:SubReg. 450 struct RegSubRegPair { 451 unsigned Reg; 452 unsigned SubReg; 453 454 RegSubRegPair(unsigned Reg = 0, unsigned SubReg = 0) RegRegSubRegPair455 : Reg(Reg), SubReg(SubReg) {} 456 457 bool operator==(const RegSubRegPair& P) const { 458 return Reg == P.Reg && SubReg == P.SubReg; 459 } 460 bool operator!=(const RegSubRegPair& P) const { 461 return !(*this == P); 462 } 463 }; 464 465 /// A pair composed of a pair of a register and a sub-register index, 466 /// and another sub-register index. 467 /// Used to give some type checking when modeling Reg:SubReg1, SubReg2. 468 struct RegSubRegPairAndIdx : RegSubRegPair { 469 unsigned SubIdx; 470 471 RegSubRegPairAndIdx(unsigned Reg = 0, unsigned SubReg = 0, 472 unsigned SubIdx = 0) RegSubRegPairRegSubRegPairAndIdx473 : RegSubRegPair(Reg, SubReg), SubIdx(SubIdx) {} 474 }; 475 476 /// Build the equivalent inputs of a REG_SEQUENCE for the given \p MI 477 /// and \p DefIdx. 478 /// \p [out] InputRegs of the equivalent REG_SEQUENCE. Each element of 479 /// the list is modeled as <Reg:SubReg, SubIdx>. Operands with the undef 480 /// flag are not added to this list. 481 /// E.g., REG_SEQUENCE %1:sub1, sub0, %2, sub1 would produce 482 /// two elements: 483 /// - %1:sub1, sub0 484 /// - %2<:0>, sub1 485 /// 486 /// \returns true if it is possible to build such an input sequence 487 /// with the pair \p MI, \p DefIdx. False otherwise. 488 /// 489 /// \pre MI.isRegSequence() or MI.isRegSequenceLike(). 490 /// 491 /// \note The generic implementation does not provide any support for 492 /// MI.isRegSequenceLike(). In other words, one has to override 493 /// getRegSequenceLikeInputs for target specific instructions. 494 bool 495 getRegSequenceInputs(const MachineInstr &MI, unsigned DefIdx, 496 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const; 497 498 /// Build the equivalent inputs of a EXTRACT_SUBREG for the given \p MI 499 /// and \p DefIdx. 500 /// \p [out] InputReg of the equivalent EXTRACT_SUBREG. 501 /// E.g., EXTRACT_SUBREG %1:sub1, sub0, sub1 would produce: 502 /// - %1:sub1, sub0 503 /// 504 /// \returns true if it is possible to build such an input sequence 505 /// with the pair \p MI, \p DefIdx and the operand has no undef flag set. 506 /// False otherwise. 507 /// 508 /// \pre MI.isExtractSubreg() or MI.isExtractSubregLike(). 509 /// 510 /// \note The generic implementation does not provide any support for 511 /// MI.isExtractSubregLike(). In other words, one has to override 512 /// getExtractSubregLikeInputs for target specific instructions. 513 bool getExtractSubregInputs(const MachineInstr &MI, unsigned DefIdx, 514 RegSubRegPairAndIdx &InputReg) const; 515 516 /// Build the equivalent inputs of a INSERT_SUBREG for the given \p MI 517 /// and \p DefIdx. 518 /// \p [out] BaseReg and \p [out] InsertedReg contain 519 /// the equivalent inputs of INSERT_SUBREG. 520 /// E.g., INSERT_SUBREG %0:sub0, %1:sub1, sub3 would produce: 521 /// - BaseReg: %0:sub0 522 /// - InsertedReg: %1:sub1, sub3 523 /// 524 /// \returns true if it is possible to build such an input sequence 525 /// with the pair \p MI, \p DefIdx and the operand has no undef flag set. 526 /// False otherwise. 527 /// 528 /// \pre MI.isInsertSubreg() or MI.isInsertSubregLike(). 529 /// 530 /// \note The generic implementation does not provide any support for 531 /// MI.isInsertSubregLike(). In other words, one has to override 532 /// getInsertSubregLikeInputs for target specific instructions. 533 bool getInsertSubregInputs(const MachineInstr &MI, unsigned DefIdx, 534 RegSubRegPair &BaseReg, 535 RegSubRegPairAndIdx &InsertedReg) const; 536 537 /// Return true if two machine instructions would produce identical values. 538 /// By default, this is only true when the two instructions 539 /// are deemed identical except for defs. If this function is called when the 540 /// IR is still in SSA form, the caller can pass the MachineRegisterInfo for 541 /// aggressive checks. 542 virtual bool produceSameValue(const MachineInstr &MI0, 543 const MachineInstr &MI1, 544 const MachineRegisterInfo *MRI = nullptr) const; 545 546 /// \returns true if a branch from an instruction with opcode \p BranchOpc 547 /// bytes is capable of jumping to a position \p BrOffset bytes away. isBranchOffsetInRange(unsigned BranchOpc,int64_t BrOffset)548 virtual bool isBranchOffsetInRange(unsigned BranchOpc, 549 int64_t BrOffset) const { 550 llvm_unreachable("target did not implement"); 551 } 552 553 /// \returns The block that branch instruction \p MI jumps to. getBranchDestBlock(const MachineInstr & MI)554 virtual MachineBasicBlock *getBranchDestBlock(const MachineInstr &MI) const { 555 llvm_unreachable("target did not implement"); 556 } 557 558 /// Insert an unconditional indirect branch at the end of \p MBB to \p 559 /// NewDestBB. \p BrOffset indicates the offset of \p NewDestBB relative to 560 /// the offset of the position to insert the new branch. 561 /// 562 /// \returns The number of bytes added to the block. 563 virtual unsigned insertIndirectBranch(MachineBasicBlock &MBB, 564 MachineBasicBlock &NewDestBB, 565 const DebugLoc &DL, 566 int64_t BrOffset = 0, 567 RegScavenger *RS = nullptr) const { 568 llvm_unreachable("target did not implement"); 569 } 570 571 /// Analyze the branching code at the end of MBB, returning 572 /// true if it cannot be understood (e.g. it's a switch dispatch or isn't 573 /// implemented for a target). Upon success, this returns false and returns 574 /// with the following information in various cases: 575 /// 576 /// 1. If this block ends with no branches (it just falls through to its succ) 577 /// just return false, leaving TBB/FBB null. 578 /// 2. If this block ends with only an unconditional branch, it sets TBB to be 579 /// the destination block. 580 /// 3. If this block ends with a conditional branch and it falls through to a 581 /// successor block, it sets TBB to be the branch destination block and a 582 /// list of operands that evaluate the condition. These operands can be 583 /// passed to other TargetInstrInfo methods to create new branches. 584 /// 4. If this block ends with a conditional branch followed by an 585 /// unconditional branch, it returns the 'true' destination in TBB, the 586 /// 'false' destination in FBB, and a list of operands that evaluate the 587 /// condition. These operands can be passed to other TargetInstrInfo 588 /// methods to create new branches. 589 /// 590 /// Note that removeBranch and insertBranch must be implemented to support 591 /// cases where this method returns success. 592 /// 593 /// If AllowModify is true, then this routine is allowed to modify the basic 594 /// block (e.g. delete instructions after the unconditional branch). 595 /// 596 /// The CFG information in MBB.Predecessors and MBB.Successors must be valid 597 /// before calling this function. 598 virtual bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, 599 MachineBasicBlock *&FBB, 600 SmallVectorImpl<MachineOperand> &Cond, 601 bool AllowModify = false) const { 602 return true; 603 } 604 605 /// Represents a predicate at the MachineFunction level. The control flow a 606 /// MachineBranchPredicate represents is: 607 /// 608 /// Reg = LHS `Predicate` RHS == ConditionDef 609 /// if Reg then goto TrueDest else goto FalseDest 610 /// 611 struct MachineBranchPredicate { 612 enum ComparePredicate { 613 PRED_EQ, // True if two values are equal 614 PRED_NE, // True if two values are not equal 615 PRED_INVALID // Sentinel value 616 }; 617 618 ComparePredicate Predicate = PRED_INVALID; 619 MachineOperand LHS = MachineOperand::CreateImm(0); 620 MachineOperand RHS = MachineOperand::CreateImm(0); 621 MachineBasicBlock *TrueDest = nullptr; 622 MachineBasicBlock *FalseDest = nullptr; 623 MachineInstr *ConditionDef = nullptr; 624 625 /// SingleUseCondition is true if ConditionDef is dead except for the 626 /// branch(es) at the end of the basic block. 627 /// 628 bool SingleUseCondition = false; 629 630 explicit MachineBranchPredicate() = default; 631 }; 632 633 /// Analyze the branching code at the end of MBB and parse it into the 634 /// MachineBranchPredicate structure if possible. Returns false on success 635 /// and true on failure. 636 /// 637 /// If AllowModify is true, then this routine is allowed to modify the basic 638 /// block (e.g. delete instructions after the unconditional branch). 639 /// 640 virtual bool analyzeBranchPredicate(MachineBasicBlock &MBB, 641 MachineBranchPredicate &MBP, 642 bool AllowModify = false) const { 643 return true; 644 } 645 646 /// Remove the branching code at the end of the specific MBB. 647 /// This is only invoked in cases where AnalyzeBranch returns success. It 648 /// returns the number of instructions that were removed. 649 /// If \p BytesRemoved is non-null, report the change in code size from the 650 /// removed instructions. 651 virtual unsigned removeBranch(MachineBasicBlock &MBB, 652 int *BytesRemoved = nullptr) const { 653 llvm_unreachable("Target didn't implement TargetInstrInfo::removeBranch!"); 654 } 655 656 /// Insert branch code into the end of the specified MachineBasicBlock. The 657 /// operands to this method are the same as those returned by AnalyzeBranch. 658 /// This is only invoked in cases where AnalyzeBranch returns success. It 659 /// returns the number of instructions inserted. If \p BytesAdded is non-null, 660 /// report the change in code size from the added instructions. 661 /// 662 /// It is also invoked by tail merging to add unconditional branches in 663 /// cases where AnalyzeBranch doesn't apply because there was no original 664 /// branch to analyze. At least this much must be implemented, else tail 665 /// merging needs to be disabled. 666 /// 667 /// The CFG information in MBB.Predecessors and MBB.Successors must be valid 668 /// before calling this function. 669 virtual unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, 670 MachineBasicBlock *FBB, 671 ArrayRef<MachineOperand> Cond, 672 const DebugLoc &DL, 673 int *BytesAdded = nullptr) const { 674 llvm_unreachable("Target didn't implement TargetInstrInfo::insertBranch!"); 675 } 676 677 unsigned insertUnconditionalBranch(MachineBasicBlock &MBB, 678 MachineBasicBlock *DestBB, 679 const DebugLoc &DL, 680 int *BytesAdded = nullptr) const { 681 return insertBranch(MBB, DestBB, nullptr, ArrayRef<MachineOperand>(), DL, 682 BytesAdded); 683 } 684 685 /// Object returned by analyzeLoopForPipelining. Allows software pipelining 686 /// implementations to query attributes of the loop being pipelined and to 687 /// apply target-specific updates to the loop once pipelining is complete. 688 class PipelinerLoopInfo { 689 public: 690 virtual ~PipelinerLoopInfo(); 691 /// Return true if the given instruction should not be pipelined and should 692 /// be ignored. An example could be a loop comparison, or induction variable 693 /// update with no users being pipelined. 694 virtual bool shouldIgnoreForPipelining(const MachineInstr *MI) const = 0; 695 696 /// Create a condition to determine if the trip count of the loop is greater 697 /// than TC. 698 /// 699 /// If the trip count is statically known to be greater than TC, return 700 /// true. If the trip count is statically known to be not greater than TC, 701 /// return false. Otherwise return nullopt and fill out Cond with the test 702 /// condition. 703 virtual Optional<bool> 704 createTripCountGreaterCondition(int TC, MachineBasicBlock &MBB, 705 SmallVectorImpl<MachineOperand> &Cond) = 0; 706 707 /// Modify the loop such that the trip count is 708 /// OriginalTC + TripCountAdjust. 709 virtual void adjustTripCount(int TripCountAdjust) = 0; 710 711 /// Called when the loop's preheader has been modified to NewPreheader. 712 virtual void setPreheader(MachineBasicBlock *NewPreheader) = 0; 713 714 /// Called when the loop is being removed. Any instructions in the preheader 715 /// should be removed. 716 /// 717 /// Once this function is called, no other functions on this object are 718 /// valid; the loop has been removed. 719 virtual void disposed() = 0; 720 }; 721 722 /// Analyze loop L, which must be a single-basic-block loop, and if the 723 /// conditions can be understood enough produce a PipelinerLoopInfo object. 724 virtual std::unique_ptr<PipelinerLoopInfo> analyzeLoopForPipelining(MachineBasicBlock * LoopBB)725 analyzeLoopForPipelining(MachineBasicBlock *LoopBB) const { 726 return nullptr; 727 } 728 729 /// Analyze the loop code, return true if it cannot be understoo. Upon 730 /// success, this function returns false and returns information about the 731 /// induction variable and compare instruction used at the end. analyzeLoop(MachineLoop & L,MachineInstr * & IndVarInst,MachineInstr * & CmpInst)732 virtual bool analyzeLoop(MachineLoop &L, MachineInstr *&IndVarInst, 733 MachineInstr *&CmpInst) const { 734 return true; 735 } 736 737 /// Generate code to reduce the loop iteration by one and check if the loop 738 /// is finished. Return the value/register of the new loop count. We need 739 /// this function when peeling off one or more iterations of a loop. This 740 /// function assumes the nth iteration is peeled first. reduceLoopCount(MachineBasicBlock & MBB,MachineBasicBlock & PreHeader,MachineInstr * IndVar,MachineInstr & Cmp,SmallVectorImpl<MachineOperand> & Cond,SmallVectorImpl<MachineInstr * > & PrevInsts,unsigned Iter,unsigned MaxIter)741 virtual unsigned reduceLoopCount(MachineBasicBlock &MBB, 742 MachineBasicBlock &PreHeader, 743 MachineInstr *IndVar, MachineInstr &Cmp, 744 SmallVectorImpl<MachineOperand> &Cond, 745 SmallVectorImpl<MachineInstr *> &PrevInsts, 746 unsigned Iter, unsigned MaxIter) const { 747 llvm_unreachable("Target didn't implement ReduceLoopCount"); 748 } 749 750 /// Delete the instruction OldInst and everything after it, replacing it with 751 /// an unconditional branch to NewDest. This is used by the tail merging pass. 752 virtual void ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail, 753 MachineBasicBlock *NewDest) const; 754 755 /// Return true if it's legal to split the given basic 756 /// block at the specified instruction (i.e. instruction would be the start 757 /// of a new basic block). isLegalToSplitMBBAt(MachineBasicBlock & MBB,MachineBasicBlock::iterator MBBI)758 virtual bool isLegalToSplitMBBAt(MachineBasicBlock &MBB, 759 MachineBasicBlock::iterator MBBI) const { 760 return true; 761 } 762 763 /// Return true if it's profitable to predicate 764 /// instructions with accumulated instruction latency of "NumCycles" 765 /// of the specified basic block, where the probability of the instructions 766 /// being executed is given by Probability, and Confidence is a measure 767 /// of our confidence that it will be properly predicted. isProfitableToIfCvt(MachineBasicBlock & MBB,unsigned NumCycles,unsigned ExtraPredCycles,BranchProbability Probability)768 virtual bool isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumCycles, 769 unsigned ExtraPredCycles, 770 BranchProbability Probability) const { 771 return false; 772 } 773 774 /// Second variant of isProfitableToIfCvt. This one 775 /// checks for the case where two basic blocks from true and false path 776 /// of a if-then-else (diamond) are predicated on mutally exclusive 777 /// predicates, where the probability of the true path being taken is given 778 /// by Probability, and Confidence is a measure of our confidence that it 779 /// will be properly predicted. isProfitableToIfCvt(MachineBasicBlock & TMBB,unsigned NumTCycles,unsigned ExtraTCycles,MachineBasicBlock & FMBB,unsigned NumFCycles,unsigned ExtraFCycles,BranchProbability Probability)780 virtual bool isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumTCycles, 781 unsigned ExtraTCycles, 782 MachineBasicBlock &FMBB, unsigned NumFCycles, 783 unsigned ExtraFCycles, 784 BranchProbability Probability) const { 785 return false; 786 } 787 788 /// Return true if it's profitable for if-converter to duplicate instructions 789 /// of specified accumulated instruction latencies in the specified MBB to 790 /// enable if-conversion. 791 /// The probability of the instructions being executed is given by 792 /// Probability, and Confidence is a measure of our confidence that it 793 /// will be properly predicted. isProfitableToDupForIfCvt(MachineBasicBlock & MBB,unsigned NumCycles,BranchProbability Probability)794 virtual bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB, 795 unsigned NumCycles, 796 BranchProbability Probability) const { 797 return false; 798 } 799 800 /// Return the increase in code size needed to predicate a contiguous run of 801 /// NumInsts instructions. extraSizeToPredicateInstructions(const MachineFunction & MF,unsigned NumInsts)802 virtual unsigned extraSizeToPredicateInstructions(const MachineFunction &MF, 803 unsigned NumInsts) const { 804 return 0; 805 } 806 807 /// Return an estimate for the code size reduction (in bytes) which will be 808 /// caused by removing the given branch instruction during if-conversion. predictBranchSizeForIfCvt(MachineInstr & MI)809 virtual unsigned predictBranchSizeForIfCvt(MachineInstr &MI) const { 810 return getInstSizeInBytes(MI); 811 } 812 813 /// Return true if it's profitable to unpredicate 814 /// one side of a 'diamond', i.e. two sides of if-else predicated on mutually 815 /// exclusive predicates. 816 /// e.g. 817 /// subeq r0, r1, #1 818 /// addne r0, r1, #1 819 /// => 820 /// sub r0, r1, #1 821 /// addne r0, r1, #1 822 /// 823 /// This may be profitable is conditional instructions are always executed. isProfitableToUnpredicate(MachineBasicBlock & TMBB,MachineBasicBlock & FMBB)824 virtual bool isProfitableToUnpredicate(MachineBasicBlock &TMBB, 825 MachineBasicBlock &FMBB) const { 826 return false; 827 } 828 829 /// Return true if it is possible to insert a select 830 /// instruction that chooses between TrueReg and FalseReg based on the 831 /// condition code in Cond. 832 /// 833 /// When successful, also return the latency in cycles from TrueReg, 834 /// FalseReg, and Cond to the destination register. In most cases, a select 835 /// instruction will be 1 cycle, so CondCycles = TrueCycles = FalseCycles = 1 836 /// 837 /// Some x86 implementations have 2-cycle cmov instructions. 838 /// 839 /// @param MBB Block where select instruction would be inserted. 840 /// @param Cond Condition returned by AnalyzeBranch. 841 /// @param TrueReg Virtual register to select when Cond is true. 842 /// @param FalseReg Virtual register to select when Cond is false. 843 /// @param CondCycles Latency from Cond+Branch to select output. 844 /// @param TrueCycles Latency from TrueReg to select output. 845 /// @param FalseCycles Latency from FalseReg to select output. canInsertSelect(const MachineBasicBlock & MBB,ArrayRef<MachineOperand> Cond,unsigned TrueReg,unsigned FalseReg,int & CondCycles,int & TrueCycles,int & FalseCycles)846 virtual bool canInsertSelect(const MachineBasicBlock &MBB, 847 ArrayRef<MachineOperand> Cond, unsigned TrueReg, 848 unsigned FalseReg, int &CondCycles, 849 int &TrueCycles, int &FalseCycles) const { 850 return false; 851 } 852 853 /// Insert a select instruction into MBB before I that will copy TrueReg to 854 /// DstReg when Cond is true, and FalseReg to DstReg when Cond is false. 855 /// 856 /// This function can only be called after canInsertSelect() returned true. 857 /// The condition in Cond comes from AnalyzeBranch, and it can be assumed 858 /// that the same flags or registers required by Cond are available at the 859 /// insertion point. 860 /// 861 /// @param MBB Block where select instruction should be inserted. 862 /// @param I Insertion point. 863 /// @param DL Source location for debugging. 864 /// @param DstReg Virtual register to be defined by select instruction. 865 /// @param Cond Condition as computed by AnalyzeBranch. 866 /// @param TrueReg Virtual register to copy when Cond is true. 867 /// @param FalseReg Virtual register to copy when Cons is false. insertSelect(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,const DebugLoc & DL,unsigned DstReg,ArrayRef<MachineOperand> Cond,unsigned TrueReg,unsigned FalseReg)868 virtual void insertSelect(MachineBasicBlock &MBB, 869 MachineBasicBlock::iterator I, const DebugLoc &DL, 870 unsigned DstReg, ArrayRef<MachineOperand> Cond, 871 unsigned TrueReg, unsigned FalseReg) const { 872 llvm_unreachable("Target didn't implement TargetInstrInfo::insertSelect!"); 873 } 874 875 /// Analyze the given select instruction, returning true if 876 /// it cannot be understood. It is assumed that MI->isSelect() is true. 877 /// 878 /// When successful, return the controlling condition and the operands that 879 /// determine the true and false result values. 880 /// 881 /// Result = SELECT Cond, TrueOp, FalseOp 882 /// 883 /// Some targets can optimize select instructions, for example by predicating 884 /// the instruction defining one of the operands. Such targets should set 885 /// Optimizable. 886 /// 887 /// @param MI Select instruction to analyze. 888 /// @param Cond Condition controlling the select. 889 /// @param TrueOp Operand number of the value selected when Cond is true. 890 /// @param FalseOp Operand number of the value selected when Cond is false. 891 /// @param Optimizable Returned as true if MI is optimizable. 892 /// @returns False on success. analyzeSelect(const MachineInstr & MI,SmallVectorImpl<MachineOperand> & Cond,unsigned & TrueOp,unsigned & FalseOp,bool & Optimizable)893 virtual bool analyzeSelect(const MachineInstr &MI, 894 SmallVectorImpl<MachineOperand> &Cond, 895 unsigned &TrueOp, unsigned &FalseOp, 896 bool &Optimizable) const { 897 assert(MI.getDesc().isSelect() && "MI must be a select instruction"); 898 return true; 899 } 900 901 /// Given a select instruction that was understood by 902 /// analyzeSelect and returned Optimizable = true, attempt to optimize MI by 903 /// merging it with one of its operands. Returns NULL on failure. 904 /// 905 /// When successful, returns the new select instruction. The client is 906 /// responsible for deleting MI. 907 /// 908 /// If both sides of the select can be optimized, PreferFalse is used to pick 909 /// a side. 910 /// 911 /// @param MI Optimizable select instruction. 912 /// @param NewMIs Set that record all MIs in the basic block up to \p 913 /// MI. Has to be updated with any newly created MI or deleted ones. 914 /// @param PreferFalse Try to optimize FalseOp instead of TrueOp. 915 /// @returns Optimized instruction or NULL. 916 virtual MachineInstr *optimizeSelect(MachineInstr &MI, 917 SmallPtrSetImpl<MachineInstr *> &NewMIs, 918 bool PreferFalse = false) const { 919 // This function must be implemented if Optimizable is ever set. 920 llvm_unreachable("Target must implement TargetInstrInfo::optimizeSelect!"); 921 } 922 923 /// Emit instructions to copy a pair of physical registers. 924 /// 925 /// This function should support copies within any legal register class as 926 /// well as any cross-class copies created during instruction selection. 927 /// 928 /// The source and destination registers may overlap, which may require a 929 /// careful implementation when multiple copy instructions are required for 930 /// large registers. See for example the ARM target. copyPhysReg(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,const DebugLoc & DL,MCRegister DestReg,MCRegister SrcReg,bool KillSrc)931 virtual void copyPhysReg(MachineBasicBlock &MBB, 932 MachineBasicBlock::iterator MI, const DebugLoc &DL, 933 MCRegister DestReg, MCRegister SrcReg, 934 bool KillSrc) const { 935 llvm_unreachable("Target didn't implement TargetInstrInfo::copyPhysReg!"); 936 } 937 938 protected: 939 /// Target-dependent implementation for IsCopyInstr. 940 /// If the specific machine instruction is a instruction that moves/copies 941 /// value from one register to another register return destination and source 942 /// registers as machine operands. 943 virtual Optional<DestSourcePair> isCopyInstrImpl(const MachineInstr & MI)944 isCopyInstrImpl(const MachineInstr &MI) const { 945 return None; 946 } 947 948 public: 949 /// If the specific machine instruction is a instruction that moves/copies 950 /// value from one register to another register return destination and source 951 /// registers as machine operands. 952 /// For COPY-instruction the method naturally returns destination and source 953 /// registers as machine operands, for all other instructions the method calls 954 /// target-dependent implementation. isCopyInstr(const MachineInstr & MI)955 Optional<DestSourcePair> isCopyInstr(const MachineInstr &MI) const { 956 if (MI.isCopy()) { 957 return DestSourcePair{MI.getOperand(0), MI.getOperand(1)}; 958 } 959 return isCopyInstrImpl(MI); 960 } 961 962 /// If the specific machine instruction is an instruction that adds an 963 /// immediate value and a physical register, and stores the result in 964 /// the given physical register \c Reg, return a pair of the source 965 /// register and the offset which has been added. isAddImmediate(const MachineInstr & MI,Register Reg)966 virtual Optional<RegImmPair> isAddImmediate(const MachineInstr &MI, 967 Register Reg) const { 968 return None; 969 } 970 971 /// Store the specified register of the given register class to the specified 972 /// stack frame index. The store instruction is to be added to the given 973 /// machine basic block before the specified machine instruction. If isKill 974 /// is true, the register operand is the last use and must be marked kill. storeRegToStackSlot(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,unsigned SrcReg,bool isKill,int FrameIndex,const TargetRegisterClass * RC,const TargetRegisterInfo * TRI)975 virtual void storeRegToStackSlot(MachineBasicBlock &MBB, 976 MachineBasicBlock::iterator MI, 977 unsigned SrcReg, bool isKill, int FrameIndex, 978 const TargetRegisterClass *RC, 979 const TargetRegisterInfo *TRI) const { 980 llvm_unreachable("Target didn't implement " 981 "TargetInstrInfo::storeRegToStackSlot!"); 982 } 983 984 /// Load the specified register of the given register class from the specified 985 /// stack frame index. The load instruction is to be added to the given 986 /// machine basic block before the specified machine instruction. loadRegFromStackSlot(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI,unsigned DestReg,int FrameIndex,const TargetRegisterClass * RC,const TargetRegisterInfo * TRI)987 virtual void loadRegFromStackSlot(MachineBasicBlock &MBB, 988 MachineBasicBlock::iterator MI, 989 unsigned DestReg, int FrameIndex, 990 const TargetRegisterClass *RC, 991 const TargetRegisterInfo *TRI) const { 992 llvm_unreachable("Target didn't implement " 993 "TargetInstrInfo::loadRegFromStackSlot!"); 994 } 995 996 /// This function is called for all pseudo instructions 997 /// that remain after register allocation. Many pseudo instructions are 998 /// created to help register allocation. This is the place to convert them 999 /// into real instructions. The target can edit MI in place, or it can insert 1000 /// new instructions and erase MI. The function should return true if 1001 /// anything was changed. expandPostRAPseudo(MachineInstr & MI)1002 virtual bool expandPostRAPseudo(MachineInstr &MI) const { return false; } 1003 1004 /// Check whether the target can fold a load that feeds a subreg operand 1005 /// (or a subreg operand that feeds a store). 1006 /// For example, X86 may want to return true if it can fold 1007 /// movl (%esp), %eax 1008 /// subb, %al, ... 1009 /// Into: 1010 /// subb (%esp), ... 1011 /// 1012 /// Ideally, we'd like the target implementation of foldMemoryOperand() to 1013 /// reject subregs - but since this behavior used to be enforced in the 1014 /// target-independent code, moving this responsibility to the targets 1015 /// has the potential of causing nasty silent breakage in out-of-tree targets. isSubregFoldable()1016 virtual bool isSubregFoldable() const { return false; } 1017 1018 /// Attempt to fold a load or store of the specified stack 1019 /// slot into the specified machine instruction for the specified operand(s). 1020 /// If this is possible, a new instruction is returned with the specified 1021 /// operand folded, otherwise NULL is returned. 1022 /// The new instruction is inserted before MI, and the client is responsible 1023 /// for removing the old instruction. 1024 /// If VRM is passed, the assigned physregs can be inspected by target to 1025 /// decide on using an opcode (note that those assignments can still change). 1026 MachineInstr *foldMemoryOperand(MachineInstr &MI, ArrayRef<unsigned> Ops, 1027 int FI, 1028 LiveIntervals *LIS = nullptr, 1029 VirtRegMap *VRM = nullptr) const; 1030 1031 /// Same as the previous version except it allows folding of any load and 1032 /// store from / to any address, not just from a specific stack slot. 1033 MachineInstr *foldMemoryOperand(MachineInstr &MI, ArrayRef<unsigned> Ops, 1034 MachineInstr &LoadMI, 1035 LiveIntervals *LIS = nullptr) const; 1036 1037 /// Return true when there is potentially a faster code sequence 1038 /// for an instruction chain ending in \p Root. All potential patterns are 1039 /// returned in the \p Pattern vector. Pattern should be sorted in priority 1040 /// order since the pattern evaluator stops checking as soon as it finds a 1041 /// faster sequence. 1042 /// \param Root - Instruction that could be combined with one of its operands 1043 /// \param Patterns - Vector of possible combination patterns 1044 virtual bool getMachineCombinerPatterns( 1045 MachineInstr &Root, 1046 SmallVectorImpl<MachineCombinerPattern> &Patterns) const; 1047 1048 /// Return true when a code sequence can improve throughput. It 1049 /// should be called only for instructions in loops. 1050 /// \param Pattern - combiner pattern 1051 virtual bool isThroughputPattern(MachineCombinerPattern Pattern) const; 1052 1053 /// Return true if the input \P Inst is part of a chain of dependent ops 1054 /// that are suitable for reassociation, otherwise return false. 1055 /// If the instruction's operands must be commuted to have a previous 1056 /// instruction of the same type define the first source operand, \P Commuted 1057 /// will be set to true. 1058 bool isReassociationCandidate(const MachineInstr &Inst, bool &Commuted) const; 1059 1060 /// Return true when \P Inst is both associative and commutative. isAssociativeAndCommutative(const MachineInstr & Inst)1061 virtual bool isAssociativeAndCommutative(const MachineInstr &Inst) const { 1062 return false; 1063 } 1064 1065 /// Return true when \P Inst has reassociable operands in the same \P MBB. 1066 virtual bool hasReassociableOperands(const MachineInstr &Inst, 1067 const MachineBasicBlock *MBB) const; 1068 1069 /// Return true when \P Inst has reassociable sibling. 1070 bool hasReassociableSibling(const MachineInstr &Inst, bool &Commuted) const; 1071 1072 /// When getMachineCombinerPatterns() finds patterns, this function generates 1073 /// the instructions that could replace the original code sequence. The client 1074 /// has to decide whether the actual replacement is beneficial or not. 1075 /// \param Root - Instruction that could be combined with one of its operands 1076 /// \param Pattern - Combination pattern for Root 1077 /// \param InsInstrs - Vector of new instructions that implement P 1078 /// \param DelInstrs - Old instructions, including Root, that could be 1079 /// replaced by InsInstr 1080 /// \param InstIdxForVirtReg - map of virtual register to instruction in 1081 /// InsInstr that defines it 1082 virtual void genAlternativeCodeSequence( 1083 MachineInstr &Root, MachineCombinerPattern Pattern, 1084 SmallVectorImpl<MachineInstr *> &InsInstrs, 1085 SmallVectorImpl<MachineInstr *> &DelInstrs, 1086 DenseMap<unsigned, unsigned> &InstIdxForVirtReg) const; 1087 1088 /// Attempt to reassociate \P Root and \P Prev according to \P Pattern to 1089 /// reduce critical path length. 1090 void reassociateOps(MachineInstr &Root, MachineInstr &Prev, 1091 MachineCombinerPattern Pattern, 1092 SmallVectorImpl<MachineInstr *> &InsInstrs, 1093 SmallVectorImpl<MachineInstr *> &DelInstrs, 1094 DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const; 1095 1096 /// This is an architecture-specific helper function of reassociateOps. 1097 /// Set special operand attributes for new instructions after reassociation. setSpecialOperandAttr(MachineInstr & OldMI1,MachineInstr & OldMI2,MachineInstr & NewMI1,MachineInstr & NewMI2)1098 virtual void setSpecialOperandAttr(MachineInstr &OldMI1, MachineInstr &OldMI2, 1099 MachineInstr &NewMI1, 1100 MachineInstr &NewMI2) const {} 1101 1102 /// Return true when a target supports MachineCombiner. useMachineCombiner()1103 virtual bool useMachineCombiner() const { return false; } 1104 1105 /// Return true if the given SDNode can be copied during scheduling 1106 /// even if it has glue. canCopyGluedNodeDuringSchedule(SDNode * N)1107 virtual bool canCopyGluedNodeDuringSchedule(SDNode *N) const { return false; } 1108 1109 protected: 1110 /// Target-dependent implementation for foldMemoryOperand. 1111 /// Target-independent code in foldMemoryOperand will 1112 /// take care of adding a MachineMemOperand to the newly created instruction. 1113 /// The instruction and any auxiliary instructions necessary will be inserted 1114 /// at InsertPt. 1115 virtual MachineInstr * 1116 foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI, 1117 ArrayRef<unsigned> Ops, 1118 MachineBasicBlock::iterator InsertPt, int FrameIndex, 1119 LiveIntervals *LIS = nullptr, 1120 VirtRegMap *VRM = nullptr) const { 1121 return nullptr; 1122 } 1123 1124 /// Target-dependent implementation for foldMemoryOperand. 1125 /// Target-independent code in foldMemoryOperand will 1126 /// take care of adding a MachineMemOperand to the newly created instruction. 1127 /// The instruction and any auxiliary instructions necessary will be inserted 1128 /// at InsertPt. 1129 virtual MachineInstr *foldMemoryOperandImpl( 1130 MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, 1131 MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, 1132 LiveIntervals *LIS = nullptr) const { 1133 return nullptr; 1134 } 1135 1136 /// Target-dependent implementation of getRegSequenceInputs. 1137 /// 1138 /// \returns true if it is possible to build the equivalent 1139 /// REG_SEQUENCE inputs with the pair \p MI, \p DefIdx. False otherwise. 1140 /// 1141 /// \pre MI.isRegSequenceLike(). 1142 /// 1143 /// \see TargetInstrInfo::getRegSequenceInputs. getRegSequenceLikeInputs(const MachineInstr & MI,unsigned DefIdx,SmallVectorImpl<RegSubRegPairAndIdx> & InputRegs)1144 virtual bool getRegSequenceLikeInputs( 1145 const MachineInstr &MI, unsigned DefIdx, 1146 SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const { 1147 return false; 1148 } 1149 1150 /// Target-dependent implementation of getExtractSubregInputs. 1151 /// 1152 /// \returns true if it is possible to build the equivalent 1153 /// EXTRACT_SUBREG inputs with the pair \p MI, \p DefIdx. False otherwise. 1154 /// 1155 /// \pre MI.isExtractSubregLike(). 1156 /// 1157 /// \see TargetInstrInfo::getExtractSubregInputs. getExtractSubregLikeInputs(const MachineInstr & MI,unsigned DefIdx,RegSubRegPairAndIdx & InputReg)1158 virtual bool getExtractSubregLikeInputs(const MachineInstr &MI, 1159 unsigned DefIdx, 1160 RegSubRegPairAndIdx &InputReg) const { 1161 return false; 1162 } 1163 1164 /// Target-dependent implementation of getInsertSubregInputs. 1165 /// 1166 /// \returns true if it is possible to build the equivalent 1167 /// INSERT_SUBREG inputs with the pair \p MI, \p DefIdx. False otherwise. 1168 /// 1169 /// \pre MI.isInsertSubregLike(). 1170 /// 1171 /// \see TargetInstrInfo::getInsertSubregInputs. 1172 virtual bool getInsertSubregLikeInputs(const MachineInstr & MI,unsigned DefIdx,RegSubRegPair & BaseReg,RegSubRegPairAndIdx & InsertedReg)1173 getInsertSubregLikeInputs(const MachineInstr &MI, unsigned DefIdx, 1174 RegSubRegPair &BaseReg, 1175 RegSubRegPairAndIdx &InsertedReg) const { 1176 return false; 1177 } 1178 1179 public: 1180 /// getAddressSpaceForPseudoSourceKind - Given the kind of memory 1181 /// (e.g. stack) the target returns the corresponding address space. 1182 virtual unsigned getAddressSpaceForPseudoSourceKind(unsigned Kind)1183 getAddressSpaceForPseudoSourceKind(unsigned Kind) const { 1184 return 0; 1185 } 1186 1187 /// unfoldMemoryOperand - Separate a single instruction which folded a load or 1188 /// a store or a load and a store into two or more instruction. If this is 1189 /// possible, returns true as well as the new instructions by reference. 1190 virtual bool unfoldMemoryOperand(MachineFunction & MF,MachineInstr & MI,unsigned Reg,bool UnfoldLoad,bool UnfoldStore,SmallVectorImpl<MachineInstr * > & NewMIs)1191 unfoldMemoryOperand(MachineFunction &MF, MachineInstr &MI, unsigned Reg, 1192 bool UnfoldLoad, bool UnfoldStore, 1193 SmallVectorImpl<MachineInstr *> &NewMIs) const { 1194 return false; 1195 } 1196 unfoldMemoryOperand(SelectionDAG & DAG,SDNode * N,SmallVectorImpl<SDNode * > & NewNodes)1197 virtual bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N, 1198 SmallVectorImpl<SDNode *> &NewNodes) const { 1199 return false; 1200 } 1201 1202 /// Returns the opcode of the would be new 1203 /// instruction after load / store are unfolded from an instruction of the 1204 /// specified opcode. It returns zero if the specified unfolding is not 1205 /// possible. If LoadRegIndex is non-null, it is filled in with the operand 1206 /// index of the operand which will hold the register holding the loaded 1207 /// value. 1208 virtual unsigned 1209 getOpcodeAfterMemoryUnfold(unsigned Opc, bool UnfoldLoad, bool UnfoldStore, 1210 unsigned *LoadRegIndex = nullptr) const { 1211 return 0; 1212 } 1213 1214 /// This is used by the pre-regalloc scheduler to determine if two loads are 1215 /// loading from the same base address. It should only return true if the base 1216 /// pointers are the same and the only differences between the two addresses 1217 /// are the offset. It also returns the offsets by reference. areLoadsFromSameBasePtr(SDNode * Load1,SDNode * Load2,int64_t & Offset1,int64_t & Offset2)1218 virtual bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, 1219 int64_t &Offset1, 1220 int64_t &Offset2) const { 1221 return false; 1222 } 1223 1224 /// This is a used by the pre-regalloc scheduler to determine (in conjunction 1225 /// with areLoadsFromSameBasePtr) if two loads should be scheduled together. 1226 /// On some targets if two loads are loading from 1227 /// addresses in the same cache line, it's better if they are scheduled 1228 /// together. This function takes two integers that represent the load offsets 1229 /// from the common base address. It returns true if it decides it's desirable 1230 /// to schedule the two loads together. "NumLoads" is the number of loads that 1231 /// have already been scheduled after Load1. shouldScheduleLoadsNear(SDNode * Load1,SDNode * Load2,int64_t Offset1,int64_t Offset2,unsigned NumLoads)1232 virtual bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2, 1233 int64_t Offset1, int64_t Offset2, 1234 unsigned NumLoads) const { 1235 return false; 1236 } 1237 1238 /// Get the base operand and byte offset of an instruction that reads/writes 1239 /// memory. 1240 /// It returns false if MI does not read/write memory. 1241 /// It returns false if no base operand and offset was found. 1242 /// It is not guaranteed to always recognize base operand and offsets in all 1243 /// cases. getMemOperandWithOffset(const MachineInstr & MI,const MachineOperand * & BaseOp,int64_t & Offset,const TargetRegisterInfo * TRI)1244 virtual bool getMemOperandWithOffset(const MachineInstr &MI, 1245 const MachineOperand *&BaseOp, 1246 int64_t &Offset, 1247 const TargetRegisterInfo *TRI) const { 1248 return false; 1249 } 1250 1251 /// Return true if the instruction contains a base register and offset. If 1252 /// true, the function also sets the operand position in the instruction 1253 /// for the base register and offset. getBaseAndOffsetPosition(const MachineInstr & MI,unsigned & BasePos,unsigned & OffsetPos)1254 virtual bool getBaseAndOffsetPosition(const MachineInstr &MI, 1255 unsigned &BasePos, 1256 unsigned &OffsetPos) const { 1257 return false; 1258 } 1259 1260 /// If the instruction is an increment of a constant value, return the amount. getIncrementValue(const MachineInstr & MI,int & Value)1261 virtual bool getIncrementValue(const MachineInstr &MI, int &Value) const { 1262 return false; 1263 } 1264 1265 /// Returns true if the two given memory operations should be scheduled 1266 /// adjacent. Note that you have to add: 1267 /// DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI)); 1268 /// or 1269 /// DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI)); 1270 /// to TargetPassConfig::createMachineScheduler() to have an effect. shouldClusterMemOps(const MachineOperand & BaseOp1,const MachineOperand & BaseOp2,unsigned NumLoads)1271 virtual bool shouldClusterMemOps(const MachineOperand &BaseOp1, 1272 const MachineOperand &BaseOp2, 1273 unsigned NumLoads) const { 1274 llvm_unreachable("target did not implement shouldClusterMemOps()"); 1275 } 1276 1277 /// Reverses the branch condition of the specified condition list, 1278 /// returning false on success and true if it cannot be reversed. 1279 virtual bool reverseBranchCondition(SmallVectorImpl<MachineOperand> & Cond)1280 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 1281 return true; 1282 } 1283 1284 /// Insert a noop into the instruction stream at the specified point. 1285 virtual void insertNoop(MachineBasicBlock &MBB, 1286 MachineBasicBlock::iterator MI) const; 1287 1288 /// Return the noop instruction to use for a noop. 1289 virtual void getNoop(MCInst &NopInst) const; 1290 1291 /// Return true for post-incremented instructions. isPostIncrement(const MachineInstr & MI)1292 virtual bool isPostIncrement(const MachineInstr &MI) const { return false; } 1293 1294 /// Returns true if the instruction is already predicated. isPredicated(const MachineInstr & MI)1295 virtual bool isPredicated(const MachineInstr &MI) const { return false; } 1296 1297 /// Returns true if the instruction is a 1298 /// terminator instruction that has not been predicated. 1299 virtual bool isUnpredicatedTerminator(const MachineInstr &MI) const; 1300 1301 /// Returns true if MI is an unconditional tail call. isUnconditionalTailCall(const MachineInstr & MI)1302 virtual bool isUnconditionalTailCall(const MachineInstr &MI) const { 1303 return false; 1304 } 1305 1306 /// Returns true if the tail call can be made conditional on BranchCond. canMakeTailCallConditional(SmallVectorImpl<MachineOperand> & Cond,const MachineInstr & TailCall)1307 virtual bool canMakeTailCallConditional(SmallVectorImpl<MachineOperand> &Cond, 1308 const MachineInstr &TailCall) const { 1309 return false; 1310 } 1311 1312 /// Replace the conditional branch in MBB with a conditional tail call. replaceBranchWithTailCall(MachineBasicBlock & MBB,SmallVectorImpl<MachineOperand> & Cond,const MachineInstr & TailCall)1313 virtual void replaceBranchWithTailCall(MachineBasicBlock &MBB, 1314 SmallVectorImpl<MachineOperand> &Cond, 1315 const MachineInstr &TailCall) const { 1316 llvm_unreachable("Target didn't implement replaceBranchWithTailCall!"); 1317 } 1318 1319 /// Convert the instruction into a predicated instruction. 1320 /// It returns true if the operation was successful. 1321 virtual bool PredicateInstruction(MachineInstr &MI, 1322 ArrayRef<MachineOperand> Pred) const; 1323 1324 /// Returns true if the first specified predicate 1325 /// subsumes the second, e.g. GE subsumes GT. SubsumesPredicate(ArrayRef<MachineOperand> Pred1,ArrayRef<MachineOperand> Pred2)1326 virtual bool SubsumesPredicate(ArrayRef<MachineOperand> Pred1, 1327 ArrayRef<MachineOperand> Pred2) const { 1328 return false; 1329 } 1330 1331 /// If the specified instruction defines any predicate 1332 /// or condition code register(s) used for predication, returns true as well 1333 /// as the definition predicate(s) by reference. DefinesPredicate(MachineInstr & MI,std::vector<MachineOperand> & Pred)1334 virtual bool DefinesPredicate(MachineInstr &MI, 1335 std::vector<MachineOperand> &Pred) const { 1336 return false; 1337 } 1338 1339 /// Return true if the specified instruction can be predicated. 1340 /// By default, this returns true for every instruction with a 1341 /// PredicateOperand. isPredicable(const MachineInstr & MI)1342 virtual bool isPredicable(const MachineInstr &MI) const { 1343 return MI.getDesc().isPredicable(); 1344 } 1345 1346 /// Return true if it's safe to move a machine 1347 /// instruction that defines the specified register class. isSafeToMoveRegClassDefs(const TargetRegisterClass * RC)1348 virtual bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const { 1349 return true; 1350 } 1351 1352 /// Test if the given instruction should be considered a scheduling boundary. 1353 /// This primarily includes labels and terminators. 1354 virtual bool isSchedulingBoundary(const MachineInstr &MI, 1355 const MachineBasicBlock *MBB, 1356 const MachineFunction &MF) const; 1357 1358 /// Measure the specified inline asm to determine an approximation of its 1359 /// length. 1360 virtual unsigned getInlineAsmLength( 1361 const char *Str, const MCAsmInfo &MAI, 1362 const TargetSubtargetInfo *STI = nullptr) const; 1363 1364 /// Allocate and return a hazard recognizer to use for this target when 1365 /// scheduling the machine instructions before register allocation. 1366 virtual ScheduleHazardRecognizer * 1367 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, 1368 const ScheduleDAG *DAG) const; 1369 1370 /// Allocate and return a hazard recognizer to use for this target when 1371 /// scheduling the machine instructions before register allocation. 1372 virtual ScheduleHazardRecognizer * 1373 CreateTargetMIHazardRecognizer(const InstrItineraryData *, 1374 const ScheduleDAGMI *DAG) const; 1375 1376 /// Allocate and return a hazard recognizer to use for this target when 1377 /// scheduling the machine instructions after register allocation. 1378 virtual ScheduleHazardRecognizer * 1379 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *, 1380 const ScheduleDAG *DAG) const; 1381 1382 /// Allocate and return a hazard recognizer to use for by non-scheduling 1383 /// passes. 1384 virtual ScheduleHazardRecognizer * CreateTargetPostRAHazardRecognizer(const MachineFunction & MF)1385 CreateTargetPostRAHazardRecognizer(const MachineFunction &MF) const { 1386 return nullptr; 1387 } 1388 1389 /// Provide a global flag for disabling the PreRA hazard recognizer that 1390 /// targets may choose to honor. 1391 bool usePreRAHazardRecognizer() const; 1392 1393 /// For a comparison instruction, return the source registers 1394 /// in SrcReg and SrcReg2 if having two register operands, and the value it 1395 /// compares against in CmpValue. Return true if the comparison instruction 1396 /// can be analyzed. analyzeCompare(const MachineInstr & MI,unsigned & SrcReg,unsigned & SrcReg2,int & Mask,int & Value)1397 virtual bool analyzeCompare(const MachineInstr &MI, unsigned &SrcReg, 1398 unsigned &SrcReg2, int &Mask, int &Value) const { 1399 return false; 1400 } 1401 1402 /// See if the comparison instruction can be converted 1403 /// into something more efficient. E.g., on ARM most instructions can set the 1404 /// flags register, obviating the need for a separate CMP. optimizeCompareInstr(MachineInstr & CmpInstr,unsigned SrcReg,unsigned SrcReg2,int Mask,int Value,const MachineRegisterInfo * MRI)1405 virtual bool optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg, 1406 unsigned SrcReg2, int Mask, int Value, 1407 const MachineRegisterInfo *MRI) const { 1408 return false; 1409 } optimizeCondBranch(MachineInstr & MI)1410 virtual bool optimizeCondBranch(MachineInstr &MI) const { return false; } 1411 1412 /// Try to remove the load by folding it to a register operand at the use. 1413 /// We fold the load instructions if and only if the 1414 /// def and use are in the same BB. We only look at one load and see 1415 /// whether it can be folded into MI. FoldAsLoadDefReg is the virtual register 1416 /// defined by the load we are trying to fold. DefMI returns the machine 1417 /// instruction that defines FoldAsLoadDefReg, and the function returns 1418 /// the machine instruction generated due to folding. optimizeLoadInstr(MachineInstr & MI,const MachineRegisterInfo * MRI,unsigned & FoldAsLoadDefReg,MachineInstr * & DefMI)1419 virtual MachineInstr *optimizeLoadInstr(MachineInstr &MI, 1420 const MachineRegisterInfo *MRI, 1421 unsigned &FoldAsLoadDefReg, 1422 MachineInstr *&DefMI) const { 1423 return nullptr; 1424 } 1425 1426 /// 'Reg' is known to be defined by a move immediate instruction, 1427 /// try to fold the immediate into the use instruction. 1428 /// If MRI->hasOneNonDBGUse(Reg) is true, and this function returns true, 1429 /// then the caller may assume that DefMI has been erased from its parent 1430 /// block. The caller may assume that it will not be erased by this 1431 /// function otherwise. FoldImmediate(MachineInstr & UseMI,MachineInstr & DefMI,unsigned Reg,MachineRegisterInfo * MRI)1432 virtual bool FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, 1433 unsigned Reg, MachineRegisterInfo *MRI) const { 1434 return false; 1435 } 1436 1437 /// Return the number of u-operations the given machine 1438 /// instruction will be decoded to on the target cpu. The itinerary's 1439 /// IssueWidth is the number of microops that can be dispatched each 1440 /// cycle. An instruction with zero microops takes no dispatch resources. 1441 virtual unsigned getNumMicroOps(const InstrItineraryData *ItinData, 1442 const MachineInstr &MI) const; 1443 1444 /// Return true for pseudo instructions that don't consume any 1445 /// machine resources in their current form. These are common cases that the 1446 /// scheduler should consider free, rather than conservatively handling them 1447 /// as instructions with no itinerary. isZeroCost(unsigned Opcode)1448 bool isZeroCost(unsigned Opcode) const { 1449 return Opcode <= TargetOpcode::COPY; 1450 } 1451 1452 virtual int getOperandLatency(const InstrItineraryData *ItinData, 1453 SDNode *DefNode, unsigned DefIdx, 1454 SDNode *UseNode, unsigned UseIdx) const; 1455 1456 /// Compute and return the use operand latency of a given pair of def and use. 1457 /// In most cases, the static scheduling itinerary was enough to determine the 1458 /// operand latency. But it may not be possible for instructions with variable 1459 /// number of defs / uses. 1460 /// 1461 /// This is a raw interface to the itinerary that may be directly overridden 1462 /// by a target. Use computeOperandLatency to get the best estimate of 1463 /// latency. 1464 virtual int getOperandLatency(const InstrItineraryData *ItinData, 1465 const MachineInstr &DefMI, unsigned DefIdx, 1466 const MachineInstr &UseMI, 1467 unsigned UseIdx) const; 1468 1469 /// Compute the instruction latency of a given instruction. 1470 /// If the instruction has higher cost when predicated, it's returned via 1471 /// PredCost. 1472 virtual unsigned getInstrLatency(const InstrItineraryData *ItinData, 1473 const MachineInstr &MI, 1474 unsigned *PredCost = nullptr) const; 1475 1476 virtual unsigned getPredicationCost(const MachineInstr &MI) const; 1477 1478 virtual int getInstrLatency(const InstrItineraryData *ItinData, 1479 SDNode *Node) const; 1480 1481 /// Return the default expected latency for a def based on its opcode. 1482 unsigned defaultDefLatency(const MCSchedModel &SchedModel, 1483 const MachineInstr &DefMI) const; 1484 1485 int computeDefOperandLatency(const InstrItineraryData *ItinData, 1486 const MachineInstr &DefMI) const; 1487 1488 /// Return true if this opcode has high latency to its result. isHighLatencyDef(int opc)1489 virtual bool isHighLatencyDef(int opc) const { return false; } 1490 1491 /// Compute operand latency between a def of 'Reg' 1492 /// and a use in the current loop. Return true if the target considered 1493 /// it 'high'. This is used by optimization passes such as machine LICM to 1494 /// determine whether it makes sense to hoist an instruction out even in a 1495 /// high register pressure situation. hasHighOperandLatency(const TargetSchedModel & SchedModel,const MachineRegisterInfo * MRI,const MachineInstr & DefMI,unsigned DefIdx,const MachineInstr & UseMI,unsigned UseIdx)1496 virtual bool hasHighOperandLatency(const TargetSchedModel &SchedModel, 1497 const MachineRegisterInfo *MRI, 1498 const MachineInstr &DefMI, unsigned DefIdx, 1499 const MachineInstr &UseMI, 1500 unsigned UseIdx) const { 1501 return false; 1502 } 1503 1504 /// Compute operand latency of a def of 'Reg'. Return true 1505 /// if the target considered it 'low'. 1506 virtual bool hasLowDefLatency(const TargetSchedModel &SchedModel, 1507 const MachineInstr &DefMI, 1508 unsigned DefIdx) const; 1509 1510 /// Perform target-specific instruction verification. verifyInstruction(const MachineInstr & MI,StringRef & ErrInfo)1511 virtual bool verifyInstruction(const MachineInstr &MI, 1512 StringRef &ErrInfo) const { 1513 return true; 1514 } 1515 1516 /// Return the current execution domain and bit mask of 1517 /// possible domains for instruction. 1518 /// 1519 /// Some micro-architectures have multiple execution domains, and multiple 1520 /// opcodes that perform the same operation in different domains. For 1521 /// example, the x86 architecture provides the por, orps, and orpd 1522 /// instructions that all do the same thing. There is a latency penalty if a 1523 /// register is written in one domain and read in another. 1524 /// 1525 /// This function returns a pair (domain, mask) containing the execution 1526 /// domain of MI, and a bit mask of possible domains. The setExecutionDomain 1527 /// function can be used to change the opcode to one of the domains in the 1528 /// bit mask. Instructions whose execution domain can't be changed should 1529 /// return a 0 mask. 1530 /// 1531 /// The execution domain numbers don't have any special meaning except domain 1532 /// 0 is used for instructions that are not associated with any interesting 1533 /// execution domain. 1534 /// 1535 virtual std::pair<uint16_t, uint16_t> getExecutionDomain(const MachineInstr & MI)1536 getExecutionDomain(const MachineInstr &MI) const { 1537 return std::make_pair(0, 0); 1538 } 1539 1540 /// Change the opcode of MI to execute in Domain. 1541 /// 1542 /// The bit (1 << Domain) must be set in the mask returned from 1543 /// getExecutionDomain(MI). setExecutionDomain(MachineInstr & MI,unsigned Domain)1544 virtual void setExecutionDomain(MachineInstr &MI, unsigned Domain) const {} 1545 1546 /// Returns the preferred minimum clearance 1547 /// before an instruction with an unwanted partial register update. 1548 /// 1549 /// Some instructions only write part of a register, and implicitly need to 1550 /// read the other parts of the register. This may cause unwanted stalls 1551 /// preventing otherwise unrelated instructions from executing in parallel in 1552 /// an out-of-order CPU. 1553 /// 1554 /// For example, the x86 instruction cvtsi2ss writes its result to bits 1555 /// [31:0] of the destination xmm register. Bits [127:32] are unaffected, so 1556 /// the instruction needs to wait for the old value of the register to become 1557 /// available: 1558 /// 1559 /// addps %xmm1, %xmm0 1560 /// movaps %xmm0, (%rax) 1561 /// cvtsi2ss %rbx, %xmm0 1562 /// 1563 /// In the code above, the cvtsi2ss instruction needs to wait for the addps 1564 /// instruction before it can issue, even though the high bits of %xmm0 1565 /// probably aren't needed. 1566 /// 1567 /// This hook returns the preferred clearance before MI, measured in 1568 /// instructions. Other defs of MI's operand OpNum are avoided in the last N 1569 /// instructions before MI. It should only return a positive value for 1570 /// unwanted dependencies. If the old bits of the defined register have 1571 /// useful values, or if MI is determined to otherwise read the dependency, 1572 /// the hook should return 0. 1573 /// 1574 /// The unwanted dependency may be handled by: 1575 /// 1576 /// 1. Allocating the same register for an MI def and use. That makes the 1577 /// unwanted dependency identical to a required dependency. 1578 /// 1579 /// 2. Allocating a register for the def that has no defs in the previous N 1580 /// instructions. 1581 /// 1582 /// 3. Calling breakPartialRegDependency() with the same arguments. This 1583 /// allows the target to insert a dependency breaking instruction. 1584 /// 1585 virtual unsigned getPartialRegUpdateClearance(const MachineInstr & MI,unsigned OpNum,const TargetRegisterInfo * TRI)1586 getPartialRegUpdateClearance(const MachineInstr &MI, unsigned OpNum, 1587 const TargetRegisterInfo *TRI) const { 1588 // The default implementation returns 0 for no partial register dependency. 1589 return 0; 1590 } 1591 1592 /// Return the minimum clearance before an instruction that reads an 1593 /// unused register. 1594 /// 1595 /// For example, AVX instructions may copy part of a register operand into 1596 /// the unused high bits of the destination register. 1597 /// 1598 /// vcvtsi2sdq %rax, undef %xmm0, %xmm14 1599 /// 1600 /// In the code above, vcvtsi2sdq copies %xmm0[127:64] into %xmm14 creating a 1601 /// false dependence on any previous write to %xmm0. 1602 /// 1603 /// This hook works similarly to getPartialRegUpdateClearance, except that it 1604 /// does not take an operand index. Instead sets \p OpNum to the index of the 1605 /// unused register. getUndefRegClearance(const MachineInstr & MI,unsigned & OpNum,const TargetRegisterInfo * TRI)1606 virtual unsigned getUndefRegClearance(const MachineInstr &MI, unsigned &OpNum, 1607 const TargetRegisterInfo *TRI) const { 1608 // The default implementation returns 0 for no undef register dependency. 1609 return 0; 1610 } 1611 1612 /// Insert a dependency-breaking instruction 1613 /// before MI to eliminate an unwanted dependency on OpNum. 1614 /// 1615 /// If it wasn't possible to avoid a def in the last N instructions before MI 1616 /// (see getPartialRegUpdateClearance), this hook will be called to break the 1617 /// unwanted dependency. 1618 /// 1619 /// On x86, an xorps instruction can be used as a dependency breaker: 1620 /// 1621 /// addps %xmm1, %xmm0 1622 /// movaps %xmm0, (%rax) 1623 /// xorps %xmm0, %xmm0 1624 /// cvtsi2ss %rbx, %xmm0 1625 /// 1626 /// An <imp-kill> operand should be added to MI if an instruction was 1627 /// inserted. This ties the instructions together in the post-ra scheduler. 1628 /// breakPartialRegDependency(MachineInstr & MI,unsigned OpNum,const TargetRegisterInfo * TRI)1629 virtual void breakPartialRegDependency(MachineInstr &MI, unsigned OpNum, 1630 const TargetRegisterInfo *TRI) const {} 1631 1632 /// Create machine specific model for scheduling. 1633 virtual DFAPacketizer * CreateTargetScheduleState(const TargetSubtargetInfo &)1634 CreateTargetScheduleState(const TargetSubtargetInfo &) const { 1635 return nullptr; 1636 } 1637 1638 /// Sometimes, it is possible for the target 1639 /// to tell, even without aliasing information, that two MIs access different 1640 /// memory addresses. This function returns true if two MIs access different 1641 /// memory addresses and false otherwise. 1642 /// 1643 /// Assumes any physical registers used to compute addresses have the same 1644 /// value for both instructions. (This is the most useful assumption for 1645 /// post-RA scheduling.) 1646 /// 1647 /// See also MachineInstr::mayAlias, which is implemented on top of this 1648 /// function. 1649 virtual bool areMemAccessesTriviallyDisjoint(const MachineInstr & MIa,const MachineInstr & MIb)1650 areMemAccessesTriviallyDisjoint(const MachineInstr &MIa, 1651 const MachineInstr &MIb) const { 1652 assert(MIa.mayLoadOrStore() && 1653 "MIa must load from or modify a memory location"); 1654 assert(MIb.mayLoadOrStore() && 1655 "MIb must load from or modify a memory location"); 1656 return false; 1657 } 1658 1659 /// Return the value to use for the MachineCSE's LookAheadLimit, 1660 /// which is a heuristic used for CSE'ing phys reg defs. getMachineCSELookAheadLimit()1661 virtual unsigned getMachineCSELookAheadLimit() const { 1662 // The default lookahead is small to prevent unprofitable quadratic 1663 // behavior. 1664 return 5; 1665 } 1666 1667 /// Return an array that contains the ids of the target indices (used for the 1668 /// TargetIndex machine operand) and their names. 1669 /// 1670 /// MIR Serialization is able to serialize only the target indices that are 1671 /// defined by this method. 1672 virtual ArrayRef<std::pair<int, const char *>> getSerializableTargetIndices()1673 getSerializableTargetIndices() const { 1674 return None; 1675 } 1676 1677 /// Decompose the machine operand's target flags into two values - the direct 1678 /// target flag value and any of bit flags that are applied. 1679 virtual std::pair<unsigned, unsigned> decomposeMachineOperandsTargetFlags(unsigned)1680 decomposeMachineOperandsTargetFlags(unsigned /*TF*/) const { 1681 return std::make_pair(0u, 0u); 1682 } 1683 1684 /// Return an array that contains the direct target flag values and their 1685 /// names. 1686 /// 1687 /// MIR Serialization is able to serialize only the target flags that are 1688 /// defined by this method. 1689 virtual ArrayRef<std::pair<unsigned, const char *>> getSerializableDirectMachineOperandTargetFlags()1690 getSerializableDirectMachineOperandTargetFlags() const { 1691 return None; 1692 } 1693 1694 /// Return an array that contains the bitmask target flag values and their 1695 /// names. 1696 /// 1697 /// MIR Serialization is able to serialize only the target flags that are 1698 /// defined by this method. 1699 virtual ArrayRef<std::pair<unsigned, const char *>> getSerializableBitmaskMachineOperandTargetFlags()1700 getSerializableBitmaskMachineOperandTargetFlags() const { 1701 return None; 1702 } 1703 1704 /// Return an array that contains the MMO target flag values and their 1705 /// names. 1706 /// 1707 /// MIR Serialization is able to serialize only the MMO target flags that are 1708 /// defined by this method. 1709 virtual ArrayRef<std::pair<MachineMemOperand::Flags, const char *>> getSerializableMachineMemOperandTargetFlags()1710 getSerializableMachineMemOperandTargetFlags() const { 1711 return None; 1712 } 1713 1714 /// Determines whether \p Inst is a tail call instruction. Override this 1715 /// method on targets that do not properly set MCID::Return and MCID::Call on 1716 /// tail call instructions." isTailCall(const MachineInstr & Inst)1717 virtual bool isTailCall(const MachineInstr &Inst) const { 1718 return Inst.isReturn() && Inst.isCall(); 1719 } 1720 1721 /// True if the instruction is bound to the top of its basic block and no 1722 /// other instructions shall be inserted before it. This can be implemented 1723 /// to prevent register allocator to insert spills before such instructions. isBasicBlockPrologue(const MachineInstr & MI)1724 virtual bool isBasicBlockPrologue(const MachineInstr &MI) const { 1725 return false; 1726 } 1727 1728 /// During PHI eleimination lets target to make necessary checks and 1729 /// insert the copy to the PHI destination register in a target specific 1730 /// manner. createPHIDestinationCopy(MachineBasicBlock & MBB,MachineBasicBlock::iterator InsPt,const DebugLoc & DL,Register Src,Register Dst)1731 virtual MachineInstr *createPHIDestinationCopy( 1732 MachineBasicBlock &MBB, MachineBasicBlock::iterator InsPt, 1733 const DebugLoc &DL, Register Src, Register Dst) const { 1734 return BuildMI(MBB, InsPt, DL, get(TargetOpcode::COPY), Dst) 1735 .addReg(Src); 1736 } 1737 1738 /// During PHI eleimination lets target to make necessary checks and 1739 /// insert the copy to the PHI destination register in a target specific 1740 /// manner. createPHISourceCopy(MachineBasicBlock & MBB,MachineBasicBlock::iterator InsPt,const DebugLoc & DL,Register Src,unsigned SrcSubReg,Register Dst)1741 virtual MachineInstr *createPHISourceCopy(MachineBasicBlock &MBB, 1742 MachineBasicBlock::iterator InsPt, 1743 const DebugLoc &DL, Register Src, 1744 unsigned SrcSubReg, 1745 Register Dst) const { 1746 return BuildMI(MBB, InsPt, DL, get(TargetOpcode::COPY), Dst) 1747 .addReg(Src, 0, SrcSubReg); 1748 } 1749 1750 /// Returns a \p outliner::OutlinedFunction struct containing target-specific 1751 /// information for a set of outlining candidates. getOutliningCandidateInfo(std::vector<outliner::Candidate> & RepeatedSequenceLocs)1752 virtual outliner::OutlinedFunction getOutliningCandidateInfo( 1753 std::vector<outliner::Candidate> &RepeatedSequenceLocs) const { 1754 llvm_unreachable( 1755 "Target didn't implement TargetInstrInfo::getOutliningCandidateInfo!"); 1756 } 1757 1758 /// Returns how or if \p MI should be outlined. 1759 virtual outliner::InstrType getOutliningType(MachineBasicBlock::iterator & MIT,unsigned Flags)1760 getOutliningType(MachineBasicBlock::iterator &MIT, unsigned Flags) const { 1761 llvm_unreachable( 1762 "Target didn't implement TargetInstrInfo::getOutliningType!"); 1763 } 1764 1765 /// Optional target hook that returns true if \p MBB is safe to outline from, 1766 /// and returns any target-specific information in \p Flags. isMBBSafeToOutlineFrom(MachineBasicBlock & MBB,unsigned & Flags)1767 virtual bool isMBBSafeToOutlineFrom(MachineBasicBlock &MBB, 1768 unsigned &Flags) const { 1769 return true; 1770 } 1771 1772 /// Insert a custom frame for outlined functions. buildOutlinedFrame(MachineBasicBlock & MBB,MachineFunction & MF,const outliner::OutlinedFunction & OF)1773 virtual void buildOutlinedFrame(MachineBasicBlock &MBB, MachineFunction &MF, 1774 const outliner::OutlinedFunction &OF) const { 1775 llvm_unreachable( 1776 "Target didn't implement TargetInstrInfo::buildOutlinedFrame!"); 1777 } 1778 1779 /// Insert a call to an outlined function into the program. 1780 /// Returns an iterator to the spot where we inserted the call. This must be 1781 /// implemented by the target. 1782 virtual MachineBasicBlock::iterator insertOutlinedCall(Module & M,MachineBasicBlock & MBB,MachineBasicBlock::iterator & It,MachineFunction & MF,const outliner::Candidate & C)1783 insertOutlinedCall(Module &M, MachineBasicBlock &MBB, 1784 MachineBasicBlock::iterator &It, MachineFunction &MF, 1785 const outliner::Candidate &C) const { 1786 llvm_unreachable( 1787 "Target didn't implement TargetInstrInfo::insertOutlinedCall!"); 1788 } 1789 1790 /// Return true if the function can safely be outlined from. 1791 /// A function \p MF is considered safe for outlining if an outlined function 1792 /// produced from instructions in F will produce a program which produces the 1793 /// same output for any set of given inputs. isFunctionSafeToOutlineFrom(MachineFunction & MF,bool OutlineFromLinkOnceODRs)1794 virtual bool isFunctionSafeToOutlineFrom(MachineFunction &MF, 1795 bool OutlineFromLinkOnceODRs) const { 1796 llvm_unreachable("Target didn't implement " 1797 "TargetInstrInfo::isFunctionSafeToOutlineFrom!"); 1798 } 1799 1800 /// Return true if the function should be outlined from by default. shouldOutlineFromFunctionByDefault(MachineFunction & MF)1801 virtual bool shouldOutlineFromFunctionByDefault(MachineFunction &MF) const { 1802 return false; 1803 } 1804 1805 /// Produce the expression describing the \p MI loading a value into 1806 /// the physical register \p Reg. This hook should only be used with 1807 /// \p MIs belonging to VReg-less functions. 1808 virtual Optional<ParamLoadedValue> describeLoadedValue(const MachineInstr &MI, 1809 Register Reg) const; 1810 1811 /// Return MIR formatter to format/parse MIR operands. Target can override 1812 /// this virtual function and return target specific MIR formatter. getMIRFormatter()1813 virtual const MIRFormatter *getMIRFormatter() const { 1814 if (!Formatter.get()) 1815 Formatter = std::make_unique<MIRFormatter>(); 1816 return Formatter.get(); 1817 } 1818 1819 private: 1820 mutable std::unique_ptr<MIRFormatter> Formatter; 1821 unsigned CallFrameSetupOpcode, CallFrameDestroyOpcode; 1822 unsigned CatchRetOpcode; 1823 unsigned ReturnOpcode; 1824 }; 1825 1826 /// Provide DenseMapInfo for TargetInstrInfo::RegSubRegPair. 1827 template <> struct DenseMapInfo<TargetInstrInfo::RegSubRegPair> { 1828 using RegInfo = DenseMapInfo<unsigned>; 1829 1830 static inline TargetInstrInfo::RegSubRegPair getEmptyKey() { 1831 return TargetInstrInfo::RegSubRegPair(RegInfo::getEmptyKey(), 1832 RegInfo::getEmptyKey()); 1833 } 1834 1835 static inline TargetInstrInfo::RegSubRegPair getTombstoneKey() { 1836 return TargetInstrInfo::RegSubRegPair(RegInfo::getTombstoneKey(), 1837 RegInfo::getTombstoneKey()); 1838 } 1839 1840 /// Reuse getHashValue implementation from 1841 /// std::pair<unsigned, unsigned>. 1842 static unsigned getHashValue(const TargetInstrInfo::RegSubRegPair &Val) { 1843 std::pair<unsigned, unsigned> PairVal = std::make_pair(Val.Reg, Val.SubReg); 1844 return DenseMapInfo<std::pair<unsigned, unsigned>>::getHashValue(PairVal); 1845 } 1846 1847 static bool isEqual(const TargetInstrInfo::RegSubRegPair &LHS, 1848 const TargetInstrInfo::RegSubRegPair &RHS) { 1849 return RegInfo::isEqual(LHS.Reg, RHS.Reg) && 1850 RegInfo::isEqual(LHS.SubReg, RHS.SubReg); 1851 } 1852 }; 1853 1854 } // end namespace llvm 1855 1856 #endif // LLVM_TARGET_TARGETINSTRINFO_H 1857