1 //===- SIInstrInfo.h - SI Instruction Info Interface ------------*- 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 /// \file 10 /// Interface definition for SIInstrInfo. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef LLVM_LIB_TARGET_AMDGPU_SIINSTRINFO_H 15 #define LLVM_LIB_TARGET_AMDGPU_SIINSTRINFO_H 16 17 #include "AMDGPUMIRFormatter.h" 18 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 19 #include "SIRegisterInfo.h" 20 #include "Utils/AMDGPUBaseInfo.h" 21 #include "llvm/ADT/SetVector.h" 22 #include "llvm/CodeGen/TargetInstrInfo.h" 23 #include "llvm/CodeGen/TargetSchedule.h" 24 25 #define GET_INSTRINFO_HEADER 26 #include "AMDGPUGenInstrInfo.inc" 27 28 namespace llvm { 29 30 class APInt; 31 class GCNSubtarget; 32 class LiveVariables; 33 class MachineDominatorTree; 34 class MachineRegisterInfo; 35 class RegScavenger; 36 class TargetRegisterClass; 37 class ScheduleHazardRecognizer; 38 39 /// Mark the MMO of a uniform load if there are no potentially clobbering stores 40 /// on any path from the start of an entry function to this load. 41 static const MachineMemOperand::Flags MONoClobber = 42 MachineMemOperand::MOTargetFlag1; 43 44 class SIInstrInfo final : public AMDGPUGenInstrInfo { 45 private: 46 const SIRegisterInfo RI; 47 const GCNSubtarget &ST; 48 TargetSchedModel SchedModel; 49 mutable std::unique_ptr<AMDGPUMIRFormatter> Formatter; 50 51 // The inverse predicate should have the negative value. 52 enum BranchPredicate { 53 INVALID_BR = 0, 54 SCC_TRUE = 1, 55 SCC_FALSE = -1, 56 VCCNZ = 2, 57 VCCZ = -2, 58 EXECNZ = -3, 59 EXECZ = 3 60 }; 61 62 using SetVectorType = SmallSetVector<MachineInstr *, 32>; 63 64 static unsigned getBranchOpcode(BranchPredicate Cond); 65 static BranchPredicate getBranchPredicate(unsigned Opcode); 66 67 public: 68 unsigned buildExtractSubReg(MachineBasicBlock::iterator MI, 69 MachineRegisterInfo &MRI, 70 MachineOperand &SuperReg, 71 const TargetRegisterClass *SuperRC, 72 unsigned SubIdx, 73 const TargetRegisterClass *SubRC) const; 74 MachineOperand buildExtractSubRegOrImm(MachineBasicBlock::iterator MI, 75 MachineRegisterInfo &MRI, 76 MachineOperand &SuperReg, 77 const TargetRegisterClass *SuperRC, 78 unsigned SubIdx, 79 const TargetRegisterClass *SubRC) const; 80 private: 81 void swapOperands(MachineInstr &Inst) const; 82 83 std::pair<bool, MachineBasicBlock *> 84 moveScalarAddSub(SetVectorType &Worklist, MachineInstr &Inst, 85 MachineDominatorTree *MDT = nullptr) const; 86 87 void lowerSelect(SetVectorType &Worklist, MachineInstr &Inst, 88 MachineDominatorTree *MDT = nullptr) const; 89 90 void lowerScalarAbs(SetVectorType &Worklist, 91 MachineInstr &Inst) const; 92 93 void lowerScalarXnor(SetVectorType &Worklist, 94 MachineInstr &Inst) const; 95 96 void splitScalarNotBinop(SetVectorType &Worklist, 97 MachineInstr &Inst, 98 unsigned Opcode) const; 99 100 void splitScalarBinOpN2(SetVectorType &Worklist, 101 MachineInstr &Inst, 102 unsigned Opcode) const; 103 104 void splitScalar64BitUnaryOp(SetVectorType &Worklist, 105 MachineInstr &Inst, unsigned Opcode, 106 bool Swap = false) const; 107 108 void splitScalar64BitAddSub(SetVectorType &Worklist, MachineInstr &Inst, 109 MachineDominatorTree *MDT = nullptr) const; 110 111 void splitScalar64BitBinaryOp(SetVectorType &Worklist, MachineInstr &Inst, 112 unsigned Opcode, 113 MachineDominatorTree *MDT = nullptr) const; 114 115 void splitScalar64BitXnor(SetVectorType &Worklist, MachineInstr &Inst, 116 MachineDominatorTree *MDT = nullptr) const; 117 118 void splitScalar64BitBCNT(SetVectorType &Worklist, 119 MachineInstr &Inst) const; 120 void splitScalar64BitBFE(SetVectorType &Worklist, 121 MachineInstr &Inst) const; 122 void movePackToVALU(SetVectorType &Worklist, 123 MachineRegisterInfo &MRI, 124 MachineInstr &Inst) const; 125 126 void addUsersToMoveToVALUWorklist(Register Reg, MachineRegisterInfo &MRI, 127 SetVectorType &Worklist) const; 128 129 void addSCCDefUsersToVALUWorklist(MachineOperand &Op, 130 MachineInstr &SCCDefInst, 131 SetVectorType &Worklist, 132 Register NewCond = Register()) const; 133 void addSCCDefsToVALUWorklist(MachineOperand &Op, 134 SetVectorType &Worklist) const; 135 136 const TargetRegisterClass * 137 getDestEquivalentVGPRClass(const MachineInstr &Inst) const; 138 139 bool checkInstOffsetsDoNotOverlap(const MachineInstr &MIa, 140 const MachineInstr &MIb) const; 141 142 Register findUsedSGPR(const MachineInstr &MI, int OpIndices[3]) const; 143 144 protected: 145 bool swapSourceModifiers(MachineInstr &MI, 146 MachineOperand &Src0, unsigned Src0OpName, 147 MachineOperand &Src1, unsigned Src1OpName) const; 148 149 MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI, 150 unsigned OpIdx0, 151 unsigned OpIdx1) const override; 152 153 public: 154 enum TargetOperandFlags { 155 MO_MASK = 0xf, 156 157 MO_NONE = 0, 158 // MO_GOTPCREL -> symbol@GOTPCREL -> R_AMDGPU_GOTPCREL. 159 MO_GOTPCREL = 1, 160 // MO_GOTPCREL32_LO -> symbol@gotpcrel32@lo -> R_AMDGPU_GOTPCREL32_LO. 161 MO_GOTPCREL32 = 2, 162 MO_GOTPCREL32_LO = 2, 163 // MO_GOTPCREL32_HI -> symbol@gotpcrel32@hi -> R_AMDGPU_GOTPCREL32_HI. 164 MO_GOTPCREL32_HI = 3, 165 // MO_REL32_LO -> symbol@rel32@lo -> R_AMDGPU_REL32_LO. 166 MO_REL32 = 4, 167 MO_REL32_LO = 4, 168 // MO_REL32_HI -> symbol@rel32@hi -> R_AMDGPU_REL32_HI. 169 MO_REL32_HI = 5, 170 171 MO_FAR_BRANCH_OFFSET = 6, 172 173 MO_ABS32_LO = 8, 174 MO_ABS32_HI = 9, 175 }; 176 177 explicit SIInstrInfo(const GCNSubtarget &ST); 178 179 const SIRegisterInfo &getRegisterInfo() const { 180 return RI; 181 } 182 183 const GCNSubtarget &getSubtarget() const { 184 return ST; 185 } 186 187 bool isReallyTriviallyReMaterializable(const MachineInstr &MI, 188 AAResults *AA) const override; 189 190 bool isIgnorableUse(const MachineOperand &MO) const override; 191 192 bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, 193 int64_t &Offset1, 194 int64_t &Offset2) const override; 195 196 bool getMemOperandsWithOffsetWidth( 197 const MachineInstr &LdSt, 198 SmallVectorImpl<const MachineOperand *> &BaseOps, int64_t &Offset, 199 bool &OffsetIsScalable, unsigned &Width, 200 const TargetRegisterInfo *TRI) const final; 201 202 bool shouldClusterMemOps(ArrayRef<const MachineOperand *> BaseOps1, 203 ArrayRef<const MachineOperand *> BaseOps2, 204 unsigned NumLoads, unsigned NumBytes) const override; 205 206 bool shouldScheduleLoadsNear(SDNode *Load0, SDNode *Load1, int64_t Offset0, 207 int64_t Offset1, unsigned NumLoads) const override; 208 209 void copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, 210 const DebugLoc &DL, MCRegister DestReg, MCRegister SrcReg, 211 bool KillSrc) const override; 212 213 void materializeImmediate(MachineBasicBlock &MBB, 214 MachineBasicBlock::iterator MI, 215 const DebugLoc &DL, 216 unsigned DestReg, 217 int64_t Value) const; 218 219 const TargetRegisterClass *getPreferredSelectRegClass( 220 unsigned Size) const; 221 222 Register insertNE(MachineBasicBlock *MBB, 223 MachineBasicBlock::iterator I, const DebugLoc &DL, 224 Register SrcReg, int Value) const; 225 226 Register insertEQ(MachineBasicBlock *MBB, 227 MachineBasicBlock::iterator I, const DebugLoc &DL, 228 Register SrcReg, int Value) const; 229 230 void storeRegToStackSlot(MachineBasicBlock &MBB, 231 MachineBasicBlock::iterator MI, Register SrcReg, 232 bool isKill, int FrameIndex, 233 const TargetRegisterClass *RC, 234 const TargetRegisterInfo *TRI) const override; 235 236 void loadRegFromStackSlot(MachineBasicBlock &MBB, 237 MachineBasicBlock::iterator MI, Register DestReg, 238 int FrameIndex, const TargetRegisterClass *RC, 239 const TargetRegisterInfo *TRI) const override; 240 241 bool expandPostRAPseudo(MachineInstr &MI) const override; 242 243 // Splits a V_MOV_B64_DPP_PSEUDO opcode into a pair of v_mov_b32_dpp 244 // instructions. Returns a pair of generated instructions. 245 // Can split either post-RA with physical registers or pre-RA with 246 // virtual registers. In latter case IR needs to be in SSA form and 247 // and a REG_SEQUENCE is produced to define original register. 248 std::pair<MachineInstr*, MachineInstr*> 249 expandMovDPP64(MachineInstr &MI) const; 250 251 // Returns an opcode that can be used to move a value to a \p DstRC 252 // register. If there is no hardware instruction that can store to \p 253 // DstRC, then AMDGPU::COPY is returned. 254 unsigned getMovOpcode(const TargetRegisterClass *DstRC) const; 255 256 const MCInstrDesc &getIndirectRegWriteMovRelPseudo(unsigned VecSize, 257 unsigned EltSize, 258 bool IsSGPR) const; 259 260 const MCInstrDesc &getIndirectGPRIDXPseudo(unsigned VecSize, 261 bool IsIndirectSrc) const; 262 LLVM_READONLY 263 int commuteOpcode(unsigned Opc) const; 264 265 LLVM_READONLY 266 inline int commuteOpcode(const MachineInstr &MI) const { 267 return commuteOpcode(MI.getOpcode()); 268 } 269 270 bool findCommutedOpIndices(const MachineInstr &MI, unsigned &SrcOpIdx1, 271 unsigned &SrcOpIdx2) const override; 272 273 bool findCommutedOpIndices(MCInstrDesc Desc, unsigned & SrcOpIdx0, 274 unsigned & SrcOpIdx1) const; 275 276 bool isBranchOffsetInRange(unsigned BranchOpc, 277 int64_t BrOffset) const override; 278 279 MachineBasicBlock *getBranchDestBlock(const MachineInstr &MI) const override; 280 281 void insertIndirectBranch(MachineBasicBlock &MBB, 282 MachineBasicBlock &NewDestBB, 283 MachineBasicBlock &RestoreBB, const DebugLoc &DL, 284 int64_t BrOffset, RegScavenger *RS) const override; 285 286 bool analyzeBranchImpl(MachineBasicBlock &MBB, 287 MachineBasicBlock::iterator I, 288 MachineBasicBlock *&TBB, 289 MachineBasicBlock *&FBB, 290 SmallVectorImpl<MachineOperand> &Cond, 291 bool AllowModify) const; 292 293 bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, 294 MachineBasicBlock *&FBB, 295 SmallVectorImpl<MachineOperand> &Cond, 296 bool AllowModify = false) const override; 297 298 unsigned removeBranch(MachineBasicBlock &MBB, 299 int *BytesRemoved = nullptr) const override; 300 301 unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, 302 MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond, 303 const DebugLoc &DL, 304 int *BytesAdded = nullptr) const override; 305 306 bool reverseBranchCondition( 307 SmallVectorImpl<MachineOperand> &Cond) const override; 308 309 bool canInsertSelect(const MachineBasicBlock &MBB, 310 ArrayRef<MachineOperand> Cond, Register DstReg, 311 Register TrueReg, Register FalseReg, int &CondCycles, 312 int &TrueCycles, int &FalseCycles) const override; 313 314 void insertSelect(MachineBasicBlock &MBB, 315 MachineBasicBlock::iterator I, const DebugLoc &DL, 316 Register DstReg, ArrayRef<MachineOperand> Cond, 317 Register TrueReg, Register FalseReg) const override; 318 319 void insertVectorSelect(MachineBasicBlock &MBB, 320 MachineBasicBlock::iterator I, const DebugLoc &DL, 321 Register DstReg, ArrayRef<MachineOperand> Cond, 322 Register TrueReg, Register FalseReg) const; 323 324 bool analyzeCompare(const MachineInstr &MI, Register &SrcReg, 325 Register &SrcReg2, int64_t &CmpMask, 326 int64_t &CmpValue) const override; 327 328 bool optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg, 329 Register SrcReg2, int64_t CmpMask, int64_t CmpValue, 330 const MachineRegisterInfo *MRI) const override; 331 332 bool 333 areMemAccessesTriviallyDisjoint(const MachineInstr &MIa, 334 const MachineInstr &MIb) const override; 335 336 static bool isFoldableCopy(const MachineInstr &MI); 337 338 void removeModOperands(MachineInstr &MI) const; 339 340 bool FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, Register Reg, 341 MachineRegisterInfo *MRI) const final; 342 343 unsigned getMachineCSELookAheadLimit() const override { return 500; } 344 345 MachineInstr *convertToThreeAddress(MachineInstr &MI, LiveVariables *LV, 346 LiveIntervals *LIS) const override; 347 348 bool isSchedulingBoundary(const MachineInstr &MI, 349 const MachineBasicBlock *MBB, 350 const MachineFunction &MF) const override; 351 352 static bool isSALU(const MachineInstr &MI) { 353 return MI.getDesc().TSFlags & SIInstrFlags::SALU; 354 } 355 356 bool isSALU(uint16_t Opcode) const { 357 return get(Opcode).TSFlags & SIInstrFlags::SALU; 358 } 359 360 static bool isVALU(const MachineInstr &MI) { 361 return MI.getDesc().TSFlags & SIInstrFlags::VALU; 362 } 363 364 bool isVALU(uint16_t Opcode) const { 365 return get(Opcode).TSFlags & SIInstrFlags::VALU; 366 } 367 368 static bool isVMEM(const MachineInstr &MI) { 369 return isMUBUF(MI) || isMTBUF(MI) || isMIMG(MI); 370 } 371 372 bool isVMEM(uint16_t Opcode) const { 373 return isMUBUF(Opcode) || isMTBUF(Opcode) || isMIMG(Opcode); 374 } 375 376 static bool isSOP1(const MachineInstr &MI) { 377 return MI.getDesc().TSFlags & SIInstrFlags::SOP1; 378 } 379 380 bool isSOP1(uint16_t Opcode) const { 381 return get(Opcode).TSFlags & SIInstrFlags::SOP1; 382 } 383 384 static bool isSOP2(const MachineInstr &MI) { 385 return MI.getDesc().TSFlags & SIInstrFlags::SOP2; 386 } 387 388 bool isSOP2(uint16_t Opcode) const { 389 return get(Opcode).TSFlags & SIInstrFlags::SOP2; 390 } 391 392 static bool isSOPC(const MachineInstr &MI) { 393 return MI.getDesc().TSFlags & SIInstrFlags::SOPC; 394 } 395 396 bool isSOPC(uint16_t Opcode) const { 397 return get(Opcode).TSFlags & SIInstrFlags::SOPC; 398 } 399 400 static bool isSOPK(const MachineInstr &MI) { 401 return MI.getDesc().TSFlags & SIInstrFlags::SOPK; 402 } 403 404 bool isSOPK(uint16_t Opcode) const { 405 return get(Opcode).TSFlags & SIInstrFlags::SOPK; 406 } 407 408 static bool isSOPP(const MachineInstr &MI) { 409 return MI.getDesc().TSFlags & SIInstrFlags::SOPP; 410 } 411 412 bool isSOPP(uint16_t Opcode) const { 413 return get(Opcode).TSFlags & SIInstrFlags::SOPP; 414 } 415 416 static bool isPacked(const MachineInstr &MI) { 417 return MI.getDesc().TSFlags & SIInstrFlags::IsPacked; 418 } 419 420 bool isPacked(uint16_t Opcode) const { 421 return get(Opcode).TSFlags & SIInstrFlags::IsPacked; 422 } 423 424 static bool isVOP1(const MachineInstr &MI) { 425 return MI.getDesc().TSFlags & SIInstrFlags::VOP1; 426 } 427 428 bool isVOP1(uint16_t Opcode) const { 429 return get(Opcode).TSFlags & SIInstrFlags::VOP1; 430 } 431 432 static bool isVOP2(const MachineInstr &MI) { 433 return MI.getDesc().TSFlags & SIInstrFlags::VOP2; 434 } 435 436 bool isVOP2(uint16_t Opcode) const { 437 return get(Opcode).TSFlags & SIInstrFlags::VOP2; 438 } 439 440 static bool isVOP3(const MachineInstr &MI) { 441 return MI.getDesc().TSFlags & SIInstrFlags::VOP3; 442 } 443 444 bool isVOP3(uint16_t Opcode) const { 445 return get(Opcode).TSFlags & SIInstrFlags::VOP3; 446 } 447 448 static bool isSDWA(const MachineInstr &MI) { 449 return MI.getDesc().TSFlags & SIInstrFlags::SDWA; 450 } 451 452 bool isSDWA(uint16_t Opcode) const { 453 return get(Opcode).TSFlags & SIInstrFlags::SDWA; 454 } 455 456 static bool isVOPC(const MachineInstr &MI) { 457 return MI.getDesc().TSFlags & SIInstrFlags::VOPC; 458 } 459 460 bool isVOPC(uint16_t Opcode) const { 461 return get(Opcode).TSFlags & SIInstrFlags::VOPC; 462 } 463 464 static bool isMUBUF(const MachineInstr &MI) { 465 return MI.getDesc().TSFlags & SIInstrFlags::MUBUF; 466 } 467 468 bool isMUBUF(uint16_t Opcode) const { 469 return get(Opcode).TSFlags & SIInstrFlags::MUBUF; 470 } 471 472 static bool isMTBUF(const MachineInstr &MI) { 473 return MI.getDesc().TSFlags & SIInstrFlags::MTBUF; 474 } 475 476 bool isMTBUF(uint16_t Opcode) const { 477 return get(Opcode).TSFlags & SIInstrFlags::MTBUF; 478 } 479 480 static bool isSMRD(const MachineInstr &MI) { 481 return MI.getDesc().TSFlags & SIInstrFlags::SMRD; 482 } 483 484 bool isSMRD(uint16_t Opcode) const { 485 return get(Opcode).TSFlags & SIInstrFlags::SMRD; 486 } 487 488 bool isBufferSMRD(const MachineInstr &MI) const; 489 490 static bool isDS(const MachineInstr &MI) { 491 return MI.getDesc().TSFlags & SIInstrFlags::DS; 492 } 493 494 bool isDS(uint16_t Opcode) const { 495 return get(Opcode).TSFlags & SIInstrFlags::DS; 496 } 497 498 bool isAlwaysGDS(uint16_t Opcode) const; 499 500 static bool isMIMG(const MachineInstr &MI) { 501 return MI.getDesc().TSFlags & SIInstrFlags::MIMG; 502 } 503 504 bool isMIMG(uint16_t Opcode) const { 505 return get(Opcode).TSFlags & SIInstrFlags::MIMG; 506 } 507 508 static bool isGather4(const MachineInstr &MI) { 509 return MI.getDesc().TSFlags & SIInstrFlags::Gather4; 510 } 511 512 bool isGather4(uint16_t Opcode) const { 513 return get(Opcode).TSFlags & SIInstrFlags::Gather4; 514 } 515 516 static bool isFLAT(const MachineInstr &MI) { 517 return MI.getDesc().TSFlags & SIInstrFlags::FLAT; 518 } 519 520 // Is a FLAT encoded instruction which accesses a specific segment, 521 // i.e. global_* or scratch_*. 522 static bool isSegmentSpecificFLAT(const MachineInstr &MI) { 523 auto Flags = MI.getDesc().TSFlags; 524 return Flags & (SIInstrFlags::FlatGlobal | SIInstrFlags::FlatScratch); 525 } 526 527 bool isSegmentSpecificFLAT(uint16_t Opcode) const { 528 auto Flags = get(Opcode).TSFlags; 529 return Flags & (SIInstrFlags::FlatGlobal | SIInstrFlags::FlatScratch); 530 } 531 532 static bool isFLATGlobal(const MachineInstr &MI) { 533 return MI.getDesc().TSFlags & SIInstrFlags::FlatGlobal; 534 } 535 536 bool isFLATGlobal(uint16_t Opcode) const { 537 return get(Opcode).TSFlags & SIInstrFlags::FlatGlobal; 538 } 539 540 static bool isFLATScratch(const MachineInstr &MI) { 541 return MI.getDesc().TSFlags & SIInstrFlags::FlatScratch; 542 } 543 544 bool isFLATScratch(uint16_t Opcode) const { 545 return get(Opcode).TSFlags & SIInstrFlags::FlatScratch; 546 } 547 548 // Any FLAT encoded instruction, including global_* and scratch_*. 549 bool isFLAT(uint16_t Opcode) const { 550 return get(Opcode).TSFlags & SIInstrFlags::FLAT; 551 } 552 553 static bool isEXP(const MachineInstr &MI) { 554 return MI.getDesc().TSFlags & SIInstrFlags::EXP; 555 } 556 557 static bool isDualSourceBlendEXP(const MachineInstr &MI) { 558 if (!isEXP(MI)) 559 return false; 560 unsigned Target = MI.getOperand(0).getImm(); 561 return Target == AMDGPU::Exp::ET_DUAL_SRC_BLEND0 || 562 Target == AMDGPU::Exp::ET_DUAL_SRC_BLEND1; 563 } 564 565 bool isEXP(uint16_t Opcode) const { 566 return get(Opcode).TSFlags & SIInstrFlags::EXP; 567 } 568 569 static bool isAtomicNoRet(const MachineInstr &MI) { 570 return MI.getDesc().TSFlags & SIInstrFlags::IsAtomicNoRet; 571 } 572 573 bool isAtomicNoRet(uint16_t Opcode) const { 574 return get(Opcode).TSFlags & SIInstrFlags::IsAtomicNoRet; 575 } 576 577 static bool isAtomicRet(const MachineInstr &MI) { 578 return MI.getDesc().TSFlags & SIInstrFlags::IsAtomicRet; 579 } 580 581 bool isAtomicRet(uint16_t Opcode) const { 582 return get(Opcode).TSFlags & SIInstrFlags::IsAtomicRet; 583 } 584 585 static bool isAtomic(const MachineInstr &MI) { 586 return MI.getDesc().TSFlags & (SIInstrFlags::IsAtomicRet | 587 SIInstrFlags::IsAtomicNoRet); 588 } 589 590 bool isAtomic(uint16_t Opcode) const { 591 return get(Opcode).TSFlags & (SIInstrFlags::IsAtomicRet | 592 SIInstrFlags::IsAtomicNoRet); 593 } 594 595 static bool isWQM(const MachineInstr &MI) { 596 return MI.getDesc().TSFlags & SIInstrFlags::WQM; 597 } 598 599 bool isWQM(uint16_t Opcode) const { 600 return get(Opcode).TSFlags & SIInstrFlags::WQM; 601 } 602 603 static bool isDisableWQM(const MachineInstr &MI) { 604 return MI.getDesc().TSFlags & SIInstrFlags::DisableWQM; 605 } 606 607 bool isDisableWQM(uint16_t Opcode) const { 608 return get(Opcode).TSFlags & SIInstrFlags::DisableWQM; 609 } 610 611 static bool isVGPRSpill(const MachineInstr &MI) { 612 return MI.getDesc().TSFlags & SIInstrFlags::VGPRSpill; 613 } 614 615 bool isVGPRSpill(uint16_t Opcode) const { 616 return get(Opcode).TSFlags & SIInstrFlags::VGPRSpill; 617 } 618 619 static bool isSGPRSpill(const MachineInstr &MI) { 620 return MI.getDesc().TSFlags & SIInstrFlags::SGPRSpill; 621 } 622 623 bool isSGPRSpill(uint16_t Opcode) const { 624 return get(Opcode).TSFlags & SIInstrFlags::SGPRSpill; 625 } 626 627 static bool isDPP(const MachineInstr &MI) { 628 return MI.getDesc().TSFlags & SIInstrFlags::DPP; 629 } 630 631 bool isDPP(uint16_t Opcode) const { 632 return get(Opcode).TSFlags & SIInstrFlags::DPP; 633 } 634 635 static bool isTRANS(const MachineInstr &MI) { 636 return MI.getDesc().TSFlags & SIInstrFlags::TRANS; 637 } 638 639 bool isTRANS(uint16_t Opcode) const { 640 return get(Opcode).TSFlags & SIInstrFlags::TRANS; 641 } 642 643 static bool isVOP3P(const MachineInstr &MI) { 644 return MI.getDesc().TSFlags & SIInstrFlags::VOP3P; 645 } 646 647 bool isVOP3P(uint16_t Opcode) const { 648 return get(Opcode).TSFlags & SIInstrFlags::VOP3P; 649 } 650 651 static bool isVINTRP(const MachineInstr &MI) { 652 return MI.getDesc().TSFlags & SIInstrFlags::VINTRP; 653 } 654 655 bool isVINTRP(uint16_t Opcode) const { 656 return get(Opcode).TSFlags & SIInstrFlags::VINTRP; 657 } 658 659 static bool isMAI(const MachineInstr &MI) { 660 return MI.getDesc().TSFlags & SIInstrFlags::IsMAI; 661 } 662 663 bool isMAI(uint16_t Opcode) const { 664 return get(Opcode).TSFlags & SIInstrFlags::IsMAI; 665 } 666 667 static bool isMFMA(const MachineInstr &MI) { 668 return isMAI(MI) && MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 && 669 MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64; 670 } 671 672 static bool isDOT(const MachineInstr &MI) { 673 return MI.getDesc().TSFlags & SIInstrFlags::IsDOT; 674 } 675 676 bool isDOT(uint16_t Opcode) const { 677 return get(Opcode).TSFlags & SIInstrFlags::IsDOT; 678 } 679 680 static bool isLDSDIR(const MachineInstr &MI) { 681 return MI.getDesc().TSFlags & SIInstrFlags::LDSDIR; 682 } 683 684 bool isLDSDIR(uint16_t Opcode) const { 685 return get(Opcode).TSFlags & SIInstrFlags::LDSDIR; 686 } 687 688 static bool isVINTERP(const MachineInstr &MI) { 689 return MI.getDesc().TSFlags & SIInstrFlags::VINTERP; 690 } 691 692 bool isVINTERP(uint16_t Opcode) const { 693 return get(Opcode).TSFlags & SIInstrFlags::VINTERP; 694 } 695 696 static bool isScalarUnit(const MachineInstr &MI) { 697 return MI.getDesc().TSFlags & (SIInstrFlags::SALU | SIInstrFlags::SMRD); 698 } 699 700 static bool usesVM_CNT(const MachineInstr &MI) { 701 return MI.getDesc().TSFlags & SIInstrFlags::VM_CNT; 702 } 703 704 static bool usesLGKM_CNT(const MachineInstr &MI) { 705 return MI.getDesc().TSFlags & SIInstrFlags::LGKM_CNT; 706 } 707 708 static bool sopkIsZext(const MachineInstr &MI) { 709 return MI.getDesc().TSFlags & SIInstrFlags::SOPK_ZEXT; 710 } 711 712 bool sopkIsZext(uint16_t Opcode) const { 713 return get(Opcode).TSFlags & SIInstrFlags::SOPK_ZEXT; 714 } 715 716 /// \returns true if this is an s_store_dword* instruction. This is more 717 /// specific than than isSMEM && mayStore. 718 static bool isScalarStore(const MachineInstr &MI) { 719 return MI.getDesc().TSFlags & SIInstrFlags::SCALAR_STORE; 720 } 721 722 bool isScalarStore(uint16_t Opcode) const { 723 return get(Opcode).TSFlags & SIInstrFlags::SCALAR_STORE; 724 } 725 726 static bool isFixedSize(const MachineInstr &MI) { 727 return MI.getDesc().TSFlags & SIInstrFlags::FIXED_SIZE; 728 } 729 730 bool isFixedSize(uint16_t Opcode) const { 731 return get(Opcode).TSFlags & SIInstrFlags::FIXED_SIZE; 732 } 733 734 static bool hasFPClamp(const MachineInstr &MI) { 735 return MI.getDesc().TSFlags & SIInstrFlags::FPClamp; 736 } 737 738 bool hasFPClamp(uint16_t Opcode) const { 739 return get(Opcode).TSFlags & SIInstrFlags::FPClamp; 740 } 741 742 static bool hasIntClamp(const MachineInstr &MI) { 743 return MI.getDesc().TSFlags & SIInstrFlags::IntClamp; 744 } 745 746 uint64_t getClampMask(const MachineInstr &MI) const { 747 const uint64_t ClampFlags = SIInstrFlags::FPClamp | 748 SIInstrFlags::IntClamp | 749 SIInstrFlags::ClampLo | 750 SIInstrFlags::ClampHi; 751 return MI.getDesc().TSFlags & ClampFlags; 752 } 753 754 static bool usesFPDPRounding(const MachineInstr &MI) { 755 return MI.getDesc().TSFlags & SIInstrFlags::FPDPRounding; 756 } 757 758 bool usesFPDPRounding(uint16_t Opcode) const { 759 return get(Opcode).TSFlags & SIInstrFlags::FPDPRounding; 760 } 761 762 static bool isFPAtomic(const MachineInstr &MI) { 763 return MI.getDesc().TSFlags & SIInstrFlags::FPAtomic; 764 } 765 766 bool isFPAtomic(uint16_t Opcode) const { 767 return get(Opcode).TSFlags & SIInstrFlags::FPAtomic; 768 } 769 770 bool isVGPRCopy(const MachineInstr &MI) const { 771 assert(MI.isCopy()); 772 Register Dest = MI.getOperand(0).getReg(); 773 const MachineFunction &MF = *MI.getParent()->getParent(); 774 const MachineRegisterInfo &MRI = MF.getRegInfo(); 775 return !RI.isSGPRReg(MRI, Dest); 776 } 777 778 bool hasVGPRUses(const MachineInstr &MI) const { 779 const MachineFunction &MF = *MI.getParent()->getParent(); 780 const MachineRegisterInfo &MRI = MF.getRegInfo(); 781 return llvm::any_of(MI.explicit_uses(), 782 [&MRI, this](const MachineOperand &MO) { 783 return MO.isReg() && RI.isVGPR(MRI, MO.getReg());}); 784 } 785 786 /// Return true if the instruction modifies the mode register.q 787 static bool modifiesModeRegister(const MachineInstr &MI); 788 789 /// Whether we must prevent this instruction from executing with EXEC = 0. 790 bool hasUnwantedEffectsWhenEXECEmpty(const MachineInstr &MI) const; 791 792 /// Returns true if the instruction could potentially depend on the value of 793 /// exec. If false, exec dependencies may safely be ignored. 794 bool mayReadEXEC(const MachineRegisterInfo &MRI, const MachineInstr &MI) const; 795 796 bool isInlineConstant(const APInt &Imm) const; 797 798 bool isInlineConstant(const APFloat &Imm) const { 799 return isInlineConstant(Imm.bitcastToAPInt()); 800 } 801 802 bool isInlineConstant(const MachineOperand &MO, uint8_t OperandType) const; 803 804 bool isInlineConstant(const MachineOperand &MO, 805 const MCOperandInfo &OpInfo) const { 806 return isInlineConstant(MO, OpInfo.OperandType); 807 } 808 809 /// \p returns true if \p UseMO is substituted with \p DefMO in \p MI it would 810 /// be an inline immediate. 811 bool isInlineConstant(const MachineInstr &MI, 812 const MachineOperand &UseMO, 813 const MachineOperand &DefMO) const { 814 assert(UseMO.getParent() == &MI); 815 int OpIdx = MI.getOperandNo(&UseMO); 816 if (!MI.getDesc().OpInfo || OpIdx >= MI.getDesc().NumOperands) { 817 return false; 818 } 819 820 return isInlineConstant(DefMO, MI.getDesc().OpInfo[OpIdx]); 821 } 822 823 /// \p returns true if the operand \p OpIdx in \p MI is a valid inline 824 /// immediate. 825 bool isInlineConstant(const MachineInstr &MI, unsigned OpIdx) const { 826 const MachineOperand &MO = MI.getOperand(OpIdx); 827 return isInlineConstant(MO, MI.getDesc().OpInfo[OpIdx].OperandType); 828 } 829 830 bool isInlineConstant(const MachineInstr &MI, unsigned OpIdx, 831 const MachineOperand &MO) const { 832 if (!MI.getDesc().OpInfo || OpIdx >= MI.getDesc().NumOperands) 833 return false; 834 835 if (MI.isCopy()) { 836 unsigned Size = getOpSize(MI, OpIdx); 837 assert(Size == 8 || Size == 4); 838 839 uint8_t OpType = (Size == 8) ? 840 AMDGPU::OPERAND_REG_IMM_INT64 : AMDGPU::OPERAND_REG_IMM_INT32; 841 return isInlineConstant(MO, OpType); 842 } 843 844 return isInlineConstant(MO, MI.getDesc().OpInfo[OpIdx].OperandType); 845 } 846 847 bool isInlineConstant(const MachineOperand &MO) const { 848 const MachineInstr *Parent = MO.getParent(); 849 return isInlineConstant(*Parent, Parent->getOperandNo(&MO)); 850 } 851 852 bool isLiteralConstant(const MachineOperand &MO, 853 const MCOperandInfo &OpInfo) const { 854 return MO.isImm() && !isInlineConstant(MO, OpInfo.OperandType); 855 } 856 857 bool isLiteralConstant(const MachineInstr &MI, int OpIdx) const { 858 const MachineOperand &MO = MI.getOperand(OpIdx); 859 return MO.isImm() && !isInlineConstant(MI, OpIdx); 860 } 861 862 // Returns true if this operand could potentially require a 32-bit literal 863 // operand, but not necessarily. A FrameIndex for example could resolve to an 864 // inline immediate value that will not require an additional 4-bytes; this 865 // assumes that it will. 866 bool isLiteralConstantLike(const MachineOperand &MO, 867 const MCOperandInfo &OpInfo) const; 868 869 bool isImmOperandLegal(const MachineInstr &MI, unsigned OpNo, 870 const MachineOperand &MO) const; 871 872 /// Return true if this 64-bit VALU instruction has a 32-bit encoding. 873 /// This function will return false if you pass it a 32-bit instruction. 874 bool hasVALU32BitEncoding(unsigned Opcode) const; 875 876 /// Returns true if this operand uses the constant bus. 877 bool usesConstantBus(const MachineRegisterInfo &MRI, 878 const MachineOperand &MO, 879 const MCOperandInfo &OpInfo) const; 880 881 /// Return true if this instruction has any modifiers. 882 /// e.g. src[012]_mod, omod, clamp. 883 bool hasModifiers(unsigned Opcode) const; 884 885 bool hasModifiersSet(const MachineInstr &MI, 886 unsigned OpName) const; 887 bool hasAnyModifiersSet(const MachineInstr &MI) const; 888 889 bool canShrink(const MachineInstr &MI, 890 const MachineRegisterInfo &MRI) const; 891 892 MachineInstr *buildShrunkInst(MachineInstr &MI, 893 unsigned NewOpcode) const; 894 895 bool verifyInstruction(const MachineInstr &MI, 896 StringRef &ErrInfo) const override; 897 898 unsigned getVALUOp(const MachineInstr &MI) const; 899 900 /// Return the correct register class for \p OpNo. For target-specific 901 /// instructions, this will return the register class that has been defined 902 /// in tablegen. For generic instructions, like REG_SEQUENCE it will return 903 /// the register class of its machine operand. 904 /// to infer the correct register class base on the other operands. 905 const TargetRegisterClass *getOpRegClass(const MachineInstr &MI, 906 unsigned OpNo) const; 907 908 /// Return the size in bytes of the operand OpNo on the given 909 // instruction opcode. 910 unsigned getOpSize(uint16_t Opcode, unsigned OpNo) const { 911 const MCOperandInfo &OpInfo = get(Opcode).OpInfo[OpNo]; 912 913 if (OpInfo.RegClass == -1) { 914 // If this is an immediate operand, this must be a 32-bit literal. 915 assert(OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE); 916 return 4; 917 } 918 919 return RI.getRegSizeInBits(*RI.getRegClass(OpInfo.RegClass)) / 8; 920 } 921 922 /// This form should usually be preferred since it handles operands 923 /// with unknown register classes. 924 unsigned getOpSize(const MachineInstr &MI, unsigned OpNo) const { 925 const MachineOperand &MO = MI.getOperand(OpNo); 926 if (MO.isReg()) { 927 if (unsigned SubReg = MO.getSubReg()) { 928 return RI.getSubRegIdxSize(SubReg) / 8; 929 } 930 } 931 return RI.getRegSizeInBits(*getOpRegClass(MI, OpNo)) / 8; 932 } 933 934 /// Legalize the \p OpIndex operand of this instruction by inserting 935 /// a MOV. For example: 936 /// ADD_I32_e32 VGPR0, 15 937 /// to 938 /// MOV VGPR1, 15 939 /// ADD_I32_e32 VGPR0, VGPR1 940 /// 941 /// If the operand being legalized is a register, then a COPY will be used 942 /// instead of MOV. 943 void legalizeOpWithMove(MachineInstr &MI, unsigned OpIdx) const; 944 945 /// Check if \p MO is a legal operand if it was the \p OpIdx Operand 946 /// for \p MI. 947 bool isOperandLegal(const MachineInstr &MI, unsigned OpIdx, 948 const MachineOperand *MO = nullptr) const; 949 950 /// Check if \p MO would be a valid operand for the given operand 951 /// definition \p OpInfo. Note this does not attempt to validate constant bus 952 /// restrictions (e.g. literal constant usage). 953 bool isLegalVSrcOperand(const MachineRegisterInfo &MRI, 954 const MCOperandInfo &OpInfo, 955 const MachineOperand &MO) const; 956 957 /// Check if \p MO (a register operand) is a legal register for the 958 /// given operand description. 959 bool isLegalRegOperand(const MachineRegisterInfo &MRI, 960 const MCOperandInfo &OpInfo, 961 const MachineOperand &MO) const; 962 963 /// Legalize operands in \p MI by either commuting it or inserting a 964 /// copy of src1. 965 void legalizeOperandsVOP2(MachineRegisterInfo &MRI, MachineInstr &MI) const; 966 967 /// Fix operands in \p MI to satisfy constant bus requirements. 968 void legalizeOperandsVOP3(MachineRegisterInfo &MRI, MachineInstr &MI) const; 969 970 /// Copy a value from a VGPR (\p SrcReg) to SGPR. This function can only 971 /// be used when it is know that the value in SrcReg is same across all 972 /// threads in the wave. 973 /// \returns The SGPR register that \p SrcReg was copied to. 974 Register readlaneVGPRToSGPR(Register SrcReg, MachineInstr &UseMI, 975 MachineRegisterInfo &MRI) const; 976 977 void legalizeOperandsSMRD(MachineRegisterInfo &MRI, MachineInstr &MI) const; 978 void legalizeOperandsFLAT(MachineRegisterInfo &MRI, MachineInstr &MI) const; 979 980 void legalizeGenericOperand(MachineBasicBlock &InsertMBB, 981 MachineBasicBlock::iterator I, 982 const TargetRegisterClass *DstRC, 983 MachineOperand &Op, MachineRegisterInfo &MRI, 984 const DebugLoc &DL) const; 985 986 /// Legalize all operands in this instruction. This function may create new 987 /// instructions and control-flow around \p MI. If present, \p MDT is 988 /// updated. 989 /// \returns A new basic block that contains \p MI if new blocks were created. 990 MachineBasicBlock * 991 legalizeOperands(MachineInstr &MI, MachineDominatorTree *MDT = nullptr) const; 992 993 /// Change SADDR form of a FLAT \p Inst to its VADDR form if saddr operand 994 /// was moved to VGPR. \returns true if succeeded. 995 bool moveFlatAddrToVGPR(MachineInstr &Inst) const; 996 997 /// Replace this instruction's opcode with the equivalent VALU 998 /// opcode. This function will also move the users of \p MI to the 999 /// VALU if necessary. If present, \p MDT is updated. 1000 MachineBasicBlock *moveToVALU(MachineInstr &MI, 1001 MachineDominatorTree *MDT = nullptr) const; 1002 1003 void insertNoop(MachineBasicBlock &MBB, 1004 MachineBasicBlock::iterator MI) const override; 1005 1006 void insertNoops(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, 1007 unsigned Quantity) const override; 1008 1009 void insertReturn(MachineBasicBlock &MBB) const; 1010 /// Return the number of wait states that result from executing this 1011 /// instruction. 1012 static unsigned getNumWaitStates(const MachineInstr &MI); 1013 1014 /// Returns the operand named \p Op. If \p MI does not have an 1015 /// operand named \c Op, this function returns nullptr. 1016 LLVM_READONLY 1017 MachineOperand *getNamedOperand(MachineInstr &MI, unsigned OperandName) const; 1018 1019 LLVM_READONLY 1020 const MachineOperand *getNamedOperand(const MachineInstr &MI, 1021 unsigned OpName) const { 1022 return getNamedOperand(const_cast<MachineInstr &>(MI), OpName); 1023 } 1024 1025 /// Get required immediate operand 1026 int64_t getNamedImmOperand(const MachineInstr &MI, unsigned OpName) const { 1027 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OpName); 1028 return MI.getOperand(Idx).getImm(); 1029 } 1030 1031 uint64_t getDefaultRsrcDataFormat() const; 1032 uint64_t getScratchRsrcWords23() const; 1033 1034 bool isLowLatencyInstruction(const MachineInstr &MI) const; 1035 bool isHighLatencyDef(int Opc) const override; 1036 1037 /// Return the descriptor of the target-specific machine instruction 1038 /// that corresponds to the specified pseudo or native opcode. 1039 const MCInstrDesc &getMCOpcodeFromPseudo(unsigned Opcode) const { 1040 return get(pseudoToMCOpcode(Opcode)); 1041 } 1042 1043 unsigned isStackAccess(const MachineInstr &MI, int &FrameIndex) const; 1044 unsigned isSGPRStackAccess(const MachineInstr &MI, int &FrameIndex) const; 1045 1046 unsigned isLoadFromStackSlot(const MachineInstr &MI, 1047 int &FrameIndex) const override; 1048 unsigned isStoreToStackSlot(const MachineInstr &MI, 1049 int &FrameIndex) const override; 1050 1051 unsigned getInstBundleSize(const MachineInstr &MI) const; 1052 unsigned getInstSizeInBytes(const MachineInstr &MI) const override; 1053 1054 bool mayAccessFlatAddressSpace(const MachineInstr &MI) const; 1055 1056 bool isNonUniformBranchInstr(MachineInstr &Instr) const; 1057 1058 void convertNonUniformIfRegion(MachineBasicBlock *IfEntry, 1059 MachineBasicBlock *IfEnd) const; 1060 1061 void convertNonUniformLoopRegion(MachineBasicBlock *LoopEntry, 1062 MachineBasicBlock *LoopEnd) const; 1063 1064 std::pair<unsigned, unsigned> 1065 decomposeMachineOperandsTargetFlags(unsigned TF) const override; 1066 1067 ArrayRef<std::pair<int, const char *>> 1068 getSerializableTargetIndices() const override; 1069 1070 ArrayRef<std::pair<unsigned, const char *>> 1071 getSerializableDirectMachineOperandTargetFlags() const override; 1072 1073 ArrayRef<std::pair<MachineMemOperand::Flags, const char *>> 1074 getSerializableMachineMemOperandTargetFlags() const override; 1075 1076 ScheduleHazardRecognizer * 1077 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, 1078 const ScheduleDAG *DAG) const override; 1079 1080 ScheduleHazardRecognizer * 1081 CreateTargetPostRAHazardRecognizer(const MachineFunction &MF) const override; 1082 1083 ScheduleHazardRecognizer * 1084 CreateTargetMIHazardRecognizer(const InstrItineraryData *II, 1085 const ScheduleDAGMI *DAG) const override; 1086 1087 bool isBasicBlockPrologue(const MachineInstr &MI) const override; 1088 1089 MachineInstr *createPHIDestinationCopy(MachineBasicBlock &MBB, 1090 MachineBasicBlock::iterator InsPt, 1091 const DebugLoc &DL, Register Src, 1092 Register Dst) const override; 1093 1094 MachineInstr *createPHISourceCopy(MachineBasicBlock &MBB, 1095 MachineBasicBlock::iterator InsPt, 1096 const DebugLoc &DL, Register Src, 1097 unsigned SrcSubReg, 1098 Register Dst) const override; 1099 1100 bool isWave32() const; 1101 1102 /// Return a partially built integer add instruction without carry. 1103 /// Caller must add source operands. 1104 /// For pre-GFX9 it will generate unused carry destination operand. 1105 /// TODO: After GFX9 it should return a no-carry operation. 1106 MachineInstrBuilder getAddNoCarry(MachineBasicBlock &MBB, 1107 MachineBasicBlock::iterator I, 1108 const DebugLoc &DL, 1109 Register DestReg) const; 1110 1111 MachineInstrBuilder getAddNoCarry(MachineBasicBlock &MBB, 1112 MachineBasicBlock::iterator I, 1113 const DebugLoc &DL, 1114 Register DestReg, 1115 RegScavenger &RS) const; 1116 1117 static bool isKillTerminator(unsigned Opcode); 1118 const MCInstrDesc &getKillTerminatorFromPseudo(unsigned Opcode) const; 1119 1120 static bool isLegalMUBUFImmOffset(unsigned Imm) { 1121 return isUInt<12>(Imm); 1122 } 1123 1124 /// Returns if \p Offset is legal for the subtarget as the offset to a FLAT 1125 /// encoded instruction. If \p Signed, this is for an instruction that 1126 /// interprets the offset as signed. 1127 bool isLegalFLATOffset(int64_t Offset, unsigned AddrSpace, 1128 uint64_t FlatVariant) const; 1129 1130 /// Split \p COffsetVal into {immediate offset field, remainder offset} 1131 /// values. 1132 std::pair<int64_t, int64_t> splitFlatOffset(int64_t COffsetVal, 1133 unsigned AddrSpace, 1134 uint64_t FlatVariant) const; 1135 1136 /// \brief Return a target-specific opcode if Opcode is a pseudo instruction. 1137 /// Return -1 if the target-specific opcode for the pseudo instruction does 1138 /// not exist. If Opcode is not a pseudo instruction, this is identity. 1139 int pseudoToMCOpcode(int Opcode) const; 1140 1141 /// \brief Check if this instruction should only be used by assembler. 1142 /// Return true if this opcode should not be used by codegen. 1143 bool isAsmOnlyOpcode(int MCOp) const; 1144 1145 const TargetRegisterClass *getRegClass(const MCInstrDesc &TID, unsigned OpNum, 1146 const TargetRegisterInfo *TRI, 1147 const MachineFunction &MF) 1148 const override; 1149 1150 void fixImplicitOperands(MachineInstr &MI) const; 1151 1152 MachineInstr *foldMemoryOperandImpl(MachineFunction &MF, MachineInstr &MI, 1153 ArrayRef<unsigned> Ops, 1154 MachineBasicBlock::iterator InsertPt, 1155 int FrameIndex, 1156 LiveIntervals *LIS = nullptr, 1157 VirtRegMap *VRM = nullptr) const override; 1158 1159 unsigned getInstrLatency(const InstrItineraryData *ItinData, 1160 const MachineInstr &MI, 1161 unsigned *PredCost = nullptr) const override; 1162 1163 const MIRFormatter *getMIRFormatter() const override { 1164 if (!Formatter.get()) 1165 Formatter = std::make_unique<AMDGPUMIRFormatter>(); 1166 return Formatter.get(); 1167 } 1168 1169 static unsigned getDSShaderTypeValue(const MachineFunction &MF); 1170 1171 const TargetSchedModel &getSchedModel() const { return SchedModel; } 1172 1173 // Enforce operand's \p OpName even alignment if required by target. 1174 // This is used if an operand is a 32 bit register but needs to be aligned 1175 // regardless. 1176 void enforceOperandRCAlignment(MachineInstr &MI, unsigned OpName) const; 1177 }; 1178 1179 /// \brief Returns true if a reg:subreg pair P has a TRC class 1180 inline bool isOfRegClass(const TargetInstrInfo::RegSubRegPair &P, 1181 const TargetRegisterClass &TRC, 1182 MachineRegisterInfo &MRI) { 1183 auto *RC = MRI.getRegClass(P.Reg); 1184 if (!P.SubReg) 1185 return RC == &TRC; 1186 auto *TRI = MRI.getTargetRegisterInfo(); 1187 return RC == TRI->getMatchingSuperRegClass(RC, &TRC, P.SubReg); 1188 } 1189 1190 /// \brief Create RegSubRegPair from a register MachineOperand 1191 inline 1192 TargetInstrInfo::RegSubRegPair getRegSubRegPair(const MachineOperand &O) { 1193 assert(O.isReg()); 1194 return TargetInstrInfo::RegSubRegPair(O.getReg(), O.getSubReg()); 1195 } 1196 1197 /// \brief Return the SubReg component from REG_SEQUENCE 1198 TargetInstrInfo::RegSubRegPair getRegSequenceSubReg(MachineInstr &MI, 1199 unsigned SubReg); 1200 1201 /// \brief Return the defining instruction for a given reg:subreg pair 1202 /// skipping copy like instructions and subreg-manipulation pseudos. 1203 /// Following another subreg of a reg:subreg isn't supported. 1204 MachineInstr *getVRegSubRegDef(const TargetInstrInfo::RegSubRegPair &P, 1205 MachineRegisterInfo &MRI); 1206 1207 /// \brief Return false if EXEC is not changed between the def of \p VReg at \p 1208 /// DefMI and the use at \p UseMI. Should be run on SSA. Currently does not 1209 /// attempt to track between blocks. 1210 bool execMayBeModifiedBeforeUse(const MachineRegisterInfo &MRI, 1211 Register VReg, 1212 const MachineInstr &DefMI, 1213 const MachineInstr &UseMI); 1214 1215 /// \brief Return false if EXEC is not changed between the def of \p VReg at \p 1216 /// DefMI and all its uses. Should be run on SSA. Currently does not attempt to 1217 /// track between blocks. 1218 bool execMayBeModifiedBeforeAnyUse(const MachineRegisterInfo &MRI, 1219 Register VReg, 1220 const MachineInstr &DefMI); 1221 1222 namespace AMDGPU { 1223 1224 LLVM_READONLY 1225 int getVOPe64(uint16_t Opcode); 1226 1227 LLVM_READONLY 1228 int getVOPe32(uint16_t Opcode); 1229 1230 LLVM_READONLY 1231 int getSDWAOp(uint16_t Opcode); 1232 1233 LLVM_READONLY 1234 int getDPPOp32(uint16_t Opcode); 1235 1236 LLVM_READONLY 1237 int getBasicFromSDWAOp(uint16_t Opcode); 1238 1239 LLVM_READONLY 1240 int getCommuteRev(uint16_t Opcode); 1241 1242 LLVM_READONLY 1243 int getCommuteOrig(uint16_t Opcode); 1244 1245 LLVM_READONLY 1246 int getAddr64Inst(uint16_t Opcode); 1247 1248 /// Check if \p Opcode is an Addr64 opcode. 1249 /// 1250 /// \returns \p Opcode if it is an Addr64 opcode, otherwise -1. 1251 LLVM_READONLY 1252 int getIfAddr64Inst(uint16_t Opcode); 1253 1254 LLVM_READONLY 1255 int getAtomicNoRetOp(uint16_t Opcode); 1256 1257 LLVM_READONLY 1258 int getSOPKOp(uint16_t Opcode); 1259 1260 /// \returns SADDR form of a FLAT Global instruction given an \p Opcode 1261 /// of a VADDR form. 1262 LLVM_READONLY 1263 int getGlobalSaddrOp(uint16_t Opcode); 1264 1265 /// \returns VADDR form of a FLAT Global instruction given an \p Opcode 1266 /// of a SADDR form. 1267 LLVM_READONLY 1268 int getGlobalVaddrOp(uint16_t Opcode); 1269 1270 LLVM_READONLY 1271 int getVCMPXNoSDstOp(uint16_t Opcode); 1272 1273 /// \returns ST form with only immediate offset of a FLAT Scratch instruction 1274 /// given an \p Opcode of an SS (SADDR) form. 1275 LLVM_READONLY 1276 int getFlatScratchInstSTfromSS(uint16_t Opcode); 1277 1278 /// \returns SV (VADDR) form of a FLAT Scratch instruction given an \p Opcode 1279 /// of an SVS (SADDR + VADDR) form. 1280 LLVM_READONLY 1281 int getFlatScratchInstSVfromSVS(uint16_t Opcode); 1282 1283 /// \returns SS (SADDR) form of a FLAT Scratch instruction given an \p Opcode 1284 /// of an SV (VADDR) form. 1285 LLVM_READONLY 1286 int getFlatScratchInstSSfromSV(uint16_t Opcode); 1287 1288 /// \returns SV (VADDR) form of a FLAT Scratch instruction given an \p Opcode 1289 /// of an SS (SADDR) form. 1290 LLVM_READONLY 1291 int getFlatScratchInstSVfromSS(uint16_t Opcode); 1292 1293 /// \returns earlyclobber version of a MAC MFMA is exists. 1294 LLVM_READONLY 1295 int getMFMAEarlyClobberOp(uint16_t Opcode); 1296 1297 /// \returns v_cmpx version of a v_cmp instruction. 1298 LLVM_READONLY 1299 int getVCMPXOpFromVCMP(uint16_t Opcode); 1300 1301 const uint64_t RSRC_DATA_FORMAT = 0xf00000000000LL; 1302 const uint64_t RSRC_ELEMENT_SIZE_SHIFT = (32 + 19); 1303 const uint64_t RSRC_INDEX_STRIDE_SHIFT = (32 + 21); 1304 const uint64_t RSRC_TID_ENABLE = UINT64_C(1) << (32 + 23); 1305 1306 } // end namespace AMDGPU 1307 1308 namespace SI { 1309 namespace KernelInputOffsets { 1310 1311 /// Offsets in bytes from the start of the input buffer 1312 enum Offsets { 1313 NGROUPS_X = 0, 1314 NGROUPS_Y = 4, 1315 NGROUPS_Z = 8, 1316 GLOBAL_SIZE_X = 12, 1317 GLOBAL_SIZE_Y = 16, 1318 GLOBAL_SIZE_Z = 20, 1319 LOCAL_SIZE_X = 24, 1320 LOCAL_SIZE_Y = 28, 1321 LOCAL_SIZE_Z = 32 1322 }; 1323 1324 } // end namespace KernelInputOffsets 1325 } // end namespace SI 1326 1327 } // end namespace llvm 1328 1329 #endif // LLVM_LIB_TARGET_AMDGPU_SIINSTRINFO_H 1330