1 //===-- SIInstrInfo.h - SI Instruction Info Interface -----------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 /// \file 11 /// \brief Interface definition for SIInstrInfo. 12 // 13 //===----------------------------------------------------------------------===// 14 15 16 #ifndef LLVM_LIB_TARGET_AMDGPU_SIINSTRINFO_H 17 #define LLVM_LIB_TARGET_AMDGPU_SIINSTRINFO_H 18 19 #include "AMDGPUInstrInfo.h" 20 #include "SIDefines.h" 21 #include "SIRegisterInfo.h" 22 23 namespace llvm { 24 25 class SIInstrInfo final : public AMDGPUInstrInfo { 26 private: 27 const SIRegisterInfo RI; 28 const SISubtarget &ST; 29 30 // The the inverse predicate should have the negative value. 31 enum BranchPredicate { 32 INVALID_BR = 0, 33 SCC_TRUE = 1, 34 SCC_FALSE = -1, 35 VCCNZ = 2, 36 VCCZ = -2, 37 EXECNZ = -3, 38 EXECZ = 3 39 }; 40 41 static unsigned getBranchOpcode(BranchPredicate Cond); 42 static BranchPredicate getBranchPredicate(unsigned Opcode); 43 44 unsigned buildExtractSubReg(MachineBasicBlock::iterator MI, 45 MachineRegisterInfo &MRI, 46 MachineOperand &SuperReg, 47 const TargetRegisterClass *SuperRC, 48 unsigned SubIdx, 49 const TargetRegisterClass *SubRC) const; 50 MachineOperand buildExtractSubRegOrImm(MachineBasicBlock::iterator MI, 51 MachineRegisterInfo &MRI, 52 MachineOperand &SuperReg, 53 const TargetRegisterClass *SuperRC, 54 unsigned SubIdx, 55 const TargetRegisterClass *SubRC) const; 56 57 void swapOperands(MachineInstr &Inst) const; 58 59 void lowerScalarAbs(SmallVectorImpl<MachineInstr *> &Worklist, 60 MachineInstr &Inst) const; 61 62 void splitScalar64BitUnaryOp(SmallVectorImpl<MachineInstr *> &Worklist, 63 MachineInstr &Inst, unsigned Opcode) const; 64 65 void splitScalar64BitBinaryOp(SmallVectorImpl<MachineInstr *> &Worklist, 66 MachineInstr &Inst, unsigned Opcode) const; 67 68 void splitScalar64BitBCNT(SmallVectorImpl<MachineInstr *> &Worklist, 69 MachineInstr &Inst) const; 70 void splitScalar64BitBFE(SmallVectorImpl<MachineInstr *> &Worklist, 71 MachineInstr &Inst) const; 72 73 void addUsersToMoveToVALUWorklist( 74 unsigned Reg, MachineRegisterInfo &MRI, 75 SmallVectorImpl<MachineInstr *> &Worklist) const; 76 77 void 78 addSCCDefUsersToVALUWorklist(MachineInstr &SCCDefInst, 79 SmallVectorImpl<MachineInstr *> &Worklist) const; 80 81 const TargetRegisterClass * 82 getDestEquivalentVGPRClass(const MachineInstr &Inst) const; 83 84 bool checkInstOffsetsDoNotOverlap(MachineInstr &MIa, MachineInstr &MIb) const; 85 86 unsigned findUsedSGPR(const MachineInstr &MI, int OpIndices[3]) const; 87 88 protected: 89 MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI, 90 unsigned OpIdx0, 91 unsigned OpIdx1) const override; 92 93 public: 94 95 enum TargetOperandFlags { 96 MO_NONE = 0, 97 MO_GOTPCREL = 1 98 }; 99 100 explicit SIInstrInfo(const SISubtarget &); 101 102 const SIRegisterInfo &getRegisterInfo() const { 103 return RI; 104 } 105 106 bool isReallyTriviallyReMaterializable(const MachineInstr &MI, 107 AliasAnalysis *AA) const override; 108 109 bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, 110 int64_t &Offset1, 111 int64_t &Offset2) const override; 112 113 bool getMemOpBaseRegImmOfs(MachineInstr &LdSt, unsigned &BaseReg, 114 int64_t &Offset, 115 const TargetRegisterInfo *TRI) const final; 116 117 bool shouldClusterMemOps(MachineInstr &FirstLdSt, MachineInstr &SecondLdSt, 118 unsigned NumLoads) const final; 119 120 void copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, 121 const DebugLoc &DL, unsigned DestReg, unsigned SrcReg, 122 bool KillSrc) const override; 123 124 unsigned calculateLDSSpillAddress(MachineBasicBlock &MBB, MachineInstr &MI, 125 RegScavenger *RS, unsigned TmpReg, 126 unsigned Offset, unsigned Size) const; 127 128 void storeRegToStackSlot(MachineBasicBlock &MBB, 129 MachineBasicBlock::iterator MI, unsigned SrcReg, 130 bool isKill, int FrameIndex, 131 const TargetRegisterClass *RC, 132 const TargetRegisterInfo *TRI) const override; 133 134 void loadRegFromStackSlot(MachineBasicBlock &MBB, 135 MachineBasicBlock::iterator MI, unsigned DestReg, 136 int FrameIndex, const TargetRegisterClass *RC, 137 const TargetRegisterInfo *TRI) const override; 138 139 bool expandPostRAPseudo(MachineInstr &MI) const override; 140 141 // \brief Returns an opcode that can be used to move a value to a \p DstRC 142 // register. If there is no hardware instruction that can store to \p 143 // DstRC, then AMDGPU::COPY is returned. 144 unsigned getMovOpcode(const TargetRegisterClass *DstRC) const; 145 146 LLVM_READONLY 147 int commuteOpcode(const MachineInstr &MI) const; 148 149 bool findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1, 150 unsigned &SrcOpIdx2) const override; 151 152 bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, 153 MachineBasicBlock *&FBB, 154 SmallVectorImpl<MachineOperand> &Cond, 155 bool AllowModify) const override; 156 157 unsigned RemoveBranch(MachineBasicBlock &MBB) const override; 158 159 unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, 160 MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond, 161 const DebugLoc &DL) const override; 162 163 bool ReverseBranchCondition( 164 SmallVectorImpl<MachineOperand> &Cond) const override; 165 166 bool 167 areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb, 168 AliasAnalysis *AA = nullptr) const override; 169 170 bool FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, unsigned Reg, 171 MachineRegisterInfo *MRI) const final; 172 173 unsigned getMachineCSELookAheadLimit() const override { return 500; } 174 175 MachineInstr *convertToThreeAddress(MachineFunction::iterator &MBB, 176 MachineInstr &MI, 177 LiveVariables *LV) const override; 178 179 bool isSchedulingBoundary(const MachineInstr &MI, 180 const MachineBasicBlock *MBB, 181 const MachineFunction &MF) const override; 182 183 static bool isSALU(const MachineInstr &MI) { 184 return MI.getDesc().TSFlags & SIInstrFlags::SALU; 185 } 186 187 bool isSALU(uint16_t Opcode) const { 188 return get(Opcode).TSFlags & SIInstrFlags::SALU; 189 } 190 191 static bool isVALU(const MachineInstr &MI) { 192 return MI.getDesc().TSFlags & SIInstrFlags::VALU; 193 } 194 195 bool isVALU(uint16_t Opcode) const { 196 return get(Opcode).TSFlags & SIInstrFlags::VALU; 197 } 198 199 static bool isVMEM(const MachineInstr &MI) { 200 return isMUBUF(MI) || isMTBUF(MI) || isMIMG(MI); 201 } 202 203 bool isVMEM(uint16_t Opcode) const { 204 return isMUBUF(Opcode) || isMTBUF(Opcode) || isMIMG(Opcode); 205 } 206 207 static bool isSOP1(const MachineInstr &MI) { 208 return MI.getDesc().TSFlags & SIInstrFlags::SOP1; 209 } 210 211 bool isSOP1(uint16_t Opcode) const { 212 return get(Opcode).TSFlags & SIInstrFlags::SOP1; 213 } 214 215 static bool isSOP2(const MachineInstr &MI) { 216 return MI.getDesc().TSFlags & SIInstrFlags::SOP2; 217 } 218 219 bool isSOP2(uint16_t Opcode) const { 220 return get(Opcode).TSFlags & SIInstrFlags::SOP2; 221 } 222 223 static bool isSOPC(const MachineInstr &MI) { 224 return MI.getDesc().TSFlags & SIInstrFlags::SOPC; 225 } 226 227 bool isSOPC(uint16_t Opcode) const { 228 return get(Opcode).TSFlags & SIInstrFlags::SOPC; 229 } 230 231 static bool isSOPK(const MachineInstr &MI) { 232 return MI.getDesc().TSFlags & SIInstrFlags::SOPK; 233 } 234 235 bool isSOPK(uint16_t Opcode) const { 236 return get(Opcode).TSFlags & SIInstrFlags::SOPK; 237 } 238 239 static bool isSOPP(const MachineInstr &MI) { 240 return MI.getDesc().TSFlags & SIInstrFlags::SOPP; 241 } 242 243 bool isSOPP(uint16_t Opcode) const { 244 return get(Opcode).TSFlags & SIInstrFlags::SOPP; 245 } 246 247 static bool isVOP1(const MachineInstr &MI) { 248 return MI.getDesc().TSFlags & SIInstrFlags::VOP1; 249 } 250 251 bool isVOP1(uint16_t Opcode) const { 252 return get(Opcode).TSFlags & SIInstrFlags::VOP1; 253 } 254 255 static bool isVOP2(const MachineInstr &MI) { 256 return MI.getDesc().TSFlags & SIInstrFlags::VOP2; 257 } 258 259 bool isVOP2(uint16_t Opcode) const { 260 return get(Opcode).TSFlags & SIInstrFlags::VOP2; 261 } 262 263 static bool isVOP3(const MachineInstr &MI) { 264 return MI.getDesc().TSFlags & SIInstrFlags::VOP3; 265 } 266 267 bool isVOP3(uint16_t Opcode) const { 268 return get(Opcode).TSFlags & SIInstrFlags::VOP3; 269 } 270 271 static bool isVOPC(const MachineInstr &MI) { 272 return MI.getDesc().TSFlags & SIInstrFlags::VOPC; 273 } 274 275 bool isVOPC(uint16_t Opcode) const { 276 return get(Opcode).TSFlags & SIInstrFlags::VOPC; 277 } 278 279 static bool isMUBUF(const MachineInstr &MI) { 280 return MI.getDesc().TSFlags & SIInstrFlags::MUBUF; 281 } 282 283 bool isMUBUF(uint16_t Opcode) const { 284 return get(Opcode).TSFlags & SIInstrFlags::MUBUF; 285 } 286 287 static bool isMTBUF(const MachineInstr &MI) { 288 return MI.getDesc().TSFlags & SIInstrFlags::MTBUF; 289 } 290 291 bool isMTBUF(uint16_t Opcode) const { 292 return get(Opcode).TSFlags & SIInstrFlags::MTBUF; 293 } 294 295 static bool isSMRD(const MachineInstr &MI) { 296 return MI.getDesc().TSFlags & SIInstrFlags::SMRD; 297 } 298 299 bool isSMRD(uint16_t Opcode) const { 300 return get(Opcode).TSFlags & SIInstrFlags::SMRD; 301 } 302 303 static bool isDS(const MachineInstr &MI) { 304 return MI.getDesc().TSFlags & SIInstrFlags::DS; 305 } 306 307 bool isDS(uint16_t Opcode) const { 308 return get(Opcode).TSFlags & SIInstrFlags::DS; 309 } 310 311 static bool isMIMG(const MachineInstr &MI) { 312 return MI.getDesc().TSFlags & SIInstrFlags::MIMG; 313 } 314 315 bool isMIMG(uint16_t Opcode) const { 316 return get(Opcode).TSFlags & SIInstrFlags::MIMG; 317 } 318 319 static bool isGather4(const MachineInstr &MI) { 320 return MI.getDesc().TSFlags & SIInstrFlags::Gather4; 321 } 322 323 bool isGather4(uint16_t Opcode) const { 324 return get(Opcode).TSFlags & SIInstrFlags::Gather4; 325 } 326 327 static bool isFLAT(const MachineInstr &MI) { 328 return MI.getDesc().TSFlags & SIInstrFlags::FLAT; 329 } 330 331 bool isFLAT(uint16_t Opcode) const { 332 return get(Opcode).TSFlags & SIInstrFlags::FLAT; 333 } 334 335 static bool isWQM(const MachineInstr &MI) { 336 return MI.getDesc().TSFlags & SIInstrFlags::WQM; 337 } 338 339 bool isWQM(uint16_t Opcode) const { 340 return get(Opcode).TSFlags & SIInstrFlags::WQM; 341 } 342 343 static bool isVGPRSpill(const MachineInstr &MI) { 344 return MI.getDesc().TSFlags & SIInstrFlags::VGPRSpill; 345 } 346 347 bool isVGPRSpill(uint16_t Opcode) const { 348 return get(Opcode).TSFlags & SIInstrFlags::VGPRSpill; 349 } 350 351 static bool isDPP(const MachineInstr &MI) { 352 return MI.getDesc().TSFlags & SIInstrFlags::DPP; 353 } 354 355 bool isDPP(uint16_t Opcode) const { 356 return get(Opcode).TSFlags & SIInstrFlags::DPP; 357 } 358 359 static bool isScalarUnit(const MachineInstr &MI) { 360 return MI.getDesc().TSFlags & (SIInstrFlags::SALU | SIInstrFlags::SMRD); 361 } 362 363 static bool usesVM_CNT(const MachineInstr &MI) { 364 return MI.getDesc().TSFlags & SIInstrFlags::VM_CNT; 365 } 366 367 bool isVGPRCopy(const MachineInstr &MI) const { 368 assert(MI.isCopy()); 369 unsigned Dest = MI.getOperand(0).getReg(); 370 const MachineFunction &MF = *MI.getParent()->getParent(); 371 const MachineRegisterInfo &MRI = MF.getRegInfo(); 372 return !RI.isSGPRReg(MRI, Dest); 373 } 374 375 bool isInlineConstant(const APInt &Imm) const; 376 bool isInlineConstant(const MachineOperand &MO, unsigned OpSize) const; 377 bool isLiteralConstant(const MachineOperand &MO, unsigned OpSize) const; 378 379 bool isImmOperandLegal(const MachineInstr &MI, unsigned OpNo, 380 const MachineOperand &MO) const; 381 382 /// \brief Return true if this 64-bit VALU instruction has a 32-bit encoding. 383 /// This function will return false if you pass it a 32-bit instruction. 384 bool hasVALU32BitEncoding(unsigned Opcode) const; 385 386 /// \brief Returns true if this operand uses the constant bus. 387 bool usesConstantBus(const MachineRegisterInfo &MRI, 388 const MachineOperand &MO, 389 unsigned OpSize) const; 390 391 /// \brief Return true if this instruction has any modifiers. 392 /// e.g. src[012]_mod, omod, clamp. 393 bool hasModifiers(unsigned Opcode) const; 394 395 bool hasModifiersSet(const MachineInstr &MI, 396 unsigned OpName) const; 397 398 bool verifyInstruction(const MachineInstr &MI, 399 StringRef &ErrInfo) const override; 400 401 static unsigned getVALUOp(const MachineInstr &MI); 402 403 bool isSALUOpSupportedOnVALU(const MachineInstr &MI) const; 404 405 /// \brief Return the correct register class for \p OpNo. For target-specific 406 /// instructions, this will return the register class that has been defined 407 /// in tablegen. For generic instructions, like REG_SEQUENCE it will return 408 /// the register class of its machine operand. 409 /// to infer the correct register class base on the other operands. 410 const TargetRegisterClass *getOpRegClass(const MachineInstr &MI, 411 unsigned OpNo) const; 412 413 /// \brief Return the size in bytes of the operand OpNo on the given 414 // instruction opcode. 415 unsigned getOpSize(uint16_t Opcode, unsigned OpNo) const { 416 const MCOperandInfo &OpInfo = get(Opcode).OpInfo[OpNo]; 417 418 if (OpInfo.RegClass == -1) { 419 // If this is an immediate operand, this must be a 32-bit literal. 420 assert(OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE); 421 return 4; 422 } 423 424 return RI.getRegClass(OpInfo.RegClass)->getSize(); 425 } 426 427 /// \brief This form should usually be preferred since it handles operands 428 /// with unknown register classes. 429 unsigned getOpSize(const MachineInstr &MI, unsigned OpNo) const { 430 return getOpRegClass(MI, OpNo)->getSize(); 431 } 432 433 /// \returns true if it is legal for the operand at index \p OpNo 434 /// to read a VGPR. 435 bool canReadVGPR(const MachineInstr &MI, unsigned OpNo) const; 436 437 /// \brief Legalize the \p OpIndex operand of this instruction by inserting 438 /// a MOV. For example: 439 /// ADD_I32_e32 VGPR0, 15 440 /// to 441 /// MOV VGPR1, 15 442 /// ADD_I32_e32 VGPR0, VGPR1 443 /// 444 /// If the operand being legalized is a register, then a COPY will be used 445 /// instead of MOV. 446 void legalizeOpWithMove(MachineInstr &MI, unsigned OpIdx) const; 447 448 /// \brief Check if \p MO is a legal operand if it was the \p OpIdx Operand 449 /// for \p MI. 450 bool isOperandLegal(const MachineInstr &MI, unsigned OpIdx, 451 const MachineOperand *MO = nullptr) const; 452 453 /// \brief Check if \p MO would be a valid operand for the given operand 454 /// definition \p OpInfo. Note this does not attempt to validate constant bus 455 /// restrictions (e.g. literal constant usage). 456 bool isLegalVSrcOperand(const MachineRegisterInfo &MRI, 457 const MCOperandInfo &OpInfo, 458 const MachineOperand &MO) const; 459 460 /// \brief Check if \p MO (a register operand) is a legal register for the 461 /// given operand description. 462 bool isLegalRegOperand(const MachineRegisterInfo &MRI, 463 const MCOperandInfo &OpInfo, 464 const MachineOperand &MO) const; 465 466 /// \brief Legalize operands in \p MI by either commuting it or inserting a 467 /// copy of src1. 468 void legalizeOperandsVOP2(MachineRegisterInfo &MRI, MachineInstr &MI) const; 469 470 /// \brief Fix operands in \p MI to satisfy constant bus requirements. 471 void legalizeOperandsVOP3(MachineRegisterInfo &MRI, MachineInstr &MI) const; 472 473 /// Copy a value from a VGPR (\p SrcReg) to SGPR. This function can only 474 /// be used when it is know that the value in SrcReg is same across all 475 /// threads in the wave. 476 /// \returns The SGPR register that \p SrcReg was copied to. 477 unsigned readlaneVGPRToSGPR(unsigned SrcReg, MachineInstr &UseMI, 478 MachineRegisterInfo &MRI) const; 479 480 void legalizeOperandsSMRD(MachineRegisterInfo &MRI, MachineInstr &MI) const; 481 482 /// \brief Legalize all operands in this instruction. This function may 483 /// create new instruction and insert them before \p MI. 484 void legalizeOperands(MachineInstr &MI) const; 485 486 /// \brief Replace this instruction's opcode with the equivalent VALU 487 /// opcode. This function will also move the users of \p MI to the 488 /// VALU if necessary. 489 void moveToVALU(MachineInstr &MI) const; 490 491 void insertWaitStates(MachineBasicBlock &MBB,MachineBasicBlock::iterator MI, 492 int Count) const; 493 494 void insertNoop(MachineBasicBlock &MBB, 495 MachineBasicBlock::iterator MI) const override; 496 497 /// \brief Return the number of wait states that result from executing this 498 /// instruction. 499 unsigned getNumWaitStates(const MachineInstr &MI) const; 500 501 /// \brief Returns the operand named \p Op. If \p MI does not have an 502 /// operand named \c Op, this function returns nullptr. 503 LLVM_READONLY 504 MachineOperand *getNamedOperand(MachineInstr &MI, unsigned OperandName) const; 505 506 LLVM_READONLY 507 const MachineOperand *getNamedOperand(const MachineInstr &MI, 508 unsigned OpName) const { 509 return getNamedOperand(const_cast<MachineInstr &>(MI), OpName); 510 } 511 512 /// Get required immediate operand 513 int64_t getNamedImmOperand(const MachineInstr &MI, unsigned OpName) const { 514 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OpName); 515 return MI.getOperand(Idx).getImm(); 516 } 517 518 uint64_t getDefaultRsrcDataFormat() const; 519 uint64_t getScratchRsrcWords23() const; 520 521 bool isLowLatencyInstruction(const MachineInstr &MI) const; 522 bool isHighLatencyInstruction(const MachineInstr &MI) const; 523 524 /// \brief Return the descriptor of the target-specific machine instruction 525 /// that corresponds to the specified pseudo or native opcode. 526 const MCInstrDesc &getMCOpcodeFromPseudo(unsigned Opcode) const { 527 return get(pseudoToMCOpcode(Opcode)); 528 } 529 530 unsigned getInstSizeInBytes(const MachineInstr &MI) const; 531 532 ArrayRef<std::pair<int, const char *>> 533 getSerializableTargetIndices() const override; 534 535 ScheduleHazardRecognizer * 536 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, 537 const ScheduleDAG *DAG) const override; 538 539 ScheduleHazardRecognizer * 540 CreateTargetPostRAHazardRecognizer(const MachineFunction &MF) const override; 541 }; 542 543 namespace AMDGPU { 544 LLVM_READONLY 545 int getVOPe64(uint16_t Opcode); 546 547 LLVM_READONLY 548 int getVOPe32(uint16_t Opcode); 549 550 LLVM_READONLY 551 int getCommuteRev(uint16_t Opcode); 552 553 LLVM_READONLY 554 int getCommuteOrig(uint16_t Opcode); 555 556 LLVM_READONLY 557 int getAddr64Inst(uint16_t Opcode); 558 559 LLVM_READONLY 560 int getAtomicRetOp(uint16_t Opcode); 561 562 LLVM_READONLY 563 int getAtomicNoRetOp(uint16_t Opcode); 564 565 const uint64_t RSRC_DATA_FORMAT = 0xf00000000000LL; 566 const uint64_t RSRC_ELEMENT_SIZE_SHIFT = (32 + 19); 567 const uint64_t RSRC_INDEX_STRIDE_SHIFT = (32 + 21); 568 const uint64_t RSRC_TID_ENABLE = UINT64_C(1) << (32 + 23); 569 } // End namespace AMDGPU 570 571 namespace SI { 572 namespace KernelInputOffsets { 573 574 /// Offsets in bytes from the start of the input buffer 575 enum Offsets { 576 NGROUPS_X = 0, 577 NGROUPS_Y = 4, 578 NGROUPS_Z = 8, 579 GLOBAL_SIZE_X = 12, 580 GLOBAL_SIZE_Y = 16, 581 GLOBAL_SIZE_Z = 20, 582 LOCAL_SIZE_X = 24, 583 LOCAL_SIZE_Y = 28, 584 LOCAL_SIZE_Z = 32 585 }; 586 587 } // End namespace KernelInputOffsets 588 } // End namespace SI 589 590 } // End namespace llvm 591 592 #endif 593
