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 /// Interface definition for SIInstrInfo. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #ifndef LLVM_LIB_TARGET_AMDGPU_SIINSTRINFO_H 16 #define LLVM_LIB_TARGET_AMDGPU_SIINSTRINFO_H 17 18 #include "AMDGPUInstrInfo.h" 19 #include "SIDefines.h" 20 #include "SIRegisterInfo.h" 21 #include "Utils/AMDGPUBaseInfo.h" 22 #include "llvm/ADT/ArrayRef.h" 23 #include "llvm/ADT/SetVector.h" 24 #include "llvm/CodeGen/MachineBasicBlock.h" 25 #include "llvm/CodeGen/MachineFunction.h" 26 #include "llvm/CodeGen/MachineInstr.h" 27 #include "llvm/CodeGen/MachineInstrBuilder.h" 28 #include "llvm/CodeGen/MachineOperand.h" 29 #include "llvm/MC/MCInstrDesc.h" 30 #include "llvm/Support/Compiler.h" 31 #include <cassert> 32 #include <cstdint> 33 34 #define GET_INSTRINFO_HEADER 35 #include "AMDGPUGenInstrInfo.inc" 36 37 namespace llvm { 38 39 class APInt; 40 class MachineDominatorTree; 41 class MachineRegisterInfo; 42 class RegScavenger; 43 class GCNSubtarget; 44 class TargetRegisterClass; 45 46 class SIInstrInfo final : public AMDGPUGenInstrInfo { 47 private: 48 const SIRegisterInfo RI; 49 const GCNSubtarget &ST; 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 bool moveScalarAddSub(SetVectorType &Worklist, MachineInstr &Inst, 84 MachineDominatorTree *MDT = nullptr) const; 85 86 void lowerScalarAbs(SetVectorType &Worklist, 87 MachineInstr &Inst) const; 88 89 void lowerScalarXnor(SetVectorType &Worklist, 90 MachineInstr &Inst) const; 91 92 void splitScalar64BitUnaryOp(SetVectorType &Worklist, 93 MachineInstr &Inst, unsigned Opcode) const; 94 95 void splitScalar64BitAddSub(SetVectorType &Worklist, MachineInstr &Inst, 96 MachineDominatorTree *MDT = nullptr) const; 97 98 void splitScalar64BitBinaryOp(SetVectorType &Worklist, MachineInstr &Inst, 99 unsigned Opcode, 100 MachineDominatorTree *MDT = nullptr) const; 101 102 void splitScalar64BitBCNT(SetVectorType &Worklist, 103 MachineInstr &Inst) const; 104 void splitScalar64BitBFE(SetVectorType &Worklist, 105 MachineInstr &Inst) const; 106 void movePackToVALU(SetVectorType &Worklist, 107 MachineRegisterInfo &MRI, 108 MachineInstr &Inst) const; 109 110 void addUsersToMoveToVALUWorklist(unsigned Reg, MachineRegisterInfo &MRI, 111 SetVectorType &Worklist) const; 112 113 void 114 addSCCDefUsersToVALUWorklist(MachineInstr &SCCDefInst, 115 SetVectorType &Worklist) const; 116 117 const TargetRegisterClass * 118 getDestEquivalentVGPRClass(const MachineInstr &Inst) const; 119 120 bool checkInstOffsetsDoNotOverlap(MachineInstr &MIa, MachineInstr &MIb) const; 121 122 unsigned findUsedSGPR(const MachineInstr &MI, int OpIndices[3]) const; 123 124 protected: 125 bool swapSourceModifiers(MachineInstr &MI, 126 MachineOperand &Src0, unsigned Src0OpName, 127 MachineOperand &Src1, unsigned Src1OpName) const; 128 129 MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI, 130 unsigned OpIdx0, 131 unsigned OpIdx1) const override; 132 133 public: 134 enum TargetOperandFlags { 135 MO_MASK = 0x7, 136 137 MO_NONE = 0, 138 // MO_GOTPCREL -> symbol@GOTPCREL -> R_AMDGPU_GOTPCREL. 139 MO_GOTPCREL = 1, 140 // MO_GOTPCREL32_LO -> symbol@gotpcrel32@lo -> R_AMDGPU_GOTPCREL32_LO. 141 MO_GOTPCREL32 = 2, 142 MO_GOTPCREL32_LO = 2, 143 // MO_GOTPCREL32_HI -> symbol@gotpcrel32@hi -> R_AMDGPU_GOTPCREL32_HI. 144 MO_GOTPCREL32_HI = 3, 145 // MO_REL32_LO -> symbol@rel32@lo -> R_AMDGPU_REL32_LO. 146 MO_REL32 = 4, 147 MO_REL32_LO = 4, 148 // MO_REL32_HI -> symbol@rel32@hi -> R_AMDGPU_REL32_HI. 149 MO_REL32_HI = 5 150 }; 151 152 explicit SIInstrInfo(const GCNSubtarget &ST); 153 154 const SIRegisterInfo &getRegisterInfo() const { 155 return RI; 156 } 157 158 bool isReallyTriviallyReMaterializable(const MachineInstr &MI, 159 AliasAnalysis *AA) const override; 160 161 bool areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, 162 int64_t &Offset1, 163 int64_t &Offset2) const override; 164 165 bool getMemOpBaseRegImmOfs(MachineInstr &LdSt, unsigned &BaseReg, 166 int64_t &Offset, 167 const TargetRegisterInfo *TRI) const final; 168 169 bool shouldClusterMemOps(MachineInstr &FirstLdSt, unsigned BaseReg1, 170 MachineInstr &SecondLdSt, unsigned BaseReg2, 171 unsigned NumLoads) const override; 172 173 bool shouldScheduleLoadsNear(SDNode *Load0, SDNode *Load1, int64_t Offset0, 174 int64_t Offset1, unsigned NumLoads) const override; 175 176 void copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, 177 const DebugLoc &DL, unsigned DestReg, unsigned SrcReg, 178 bool KillSrc) const override; 179 180 unsigned calculateLDSSpillAddress(MachineBasicBlock &MBB, MachineInstr &MI, 181 RegScavenger *RS, unsigned TmpReg, 182 unsigned Offset, unsigned Size) const; 183 184 void materializeImmediate(MachineBasicBlock &MBB, 185 MachineBasicBlock::iterator MI, 186 const DebugLoc &DL, 187 unsigned DestReg, 188 int64_t Value) const; 189 190 const TargetRegisterClass *getPreferredSelectRegClass( 191 unsigned Size) const; 192 193 unsigned insertNE(MachineBasicBlock *MBB, 194 MachineBasicBlock::iterator I, const DebugLoc &DL, 195 unsigned SrcReg, int Value) const; 196 197 unsigned insertEQ(MachineBasicBlock *MBB, 198 MachineBasicBlock::iterator I, const DebugLoc &DL, 199 unsigned SrcReg, int Value) const; 200 201 void storeRegToStackSlot(MachineBasicBlock &MBB, 202 MachineBasicBlock::iterator MI, unsigned SrcReg, 203 bool isKill, int FrameIndex, 204 const TargetRegisterClass *RC, 205 const TargetRegisterInfo *TRI) const override; 206 207 void loadRegFromStackSlot(MachineBasicBlock &MBB, 208 MachineBasicBlock::iterator MI, unsigned DestReg, 209 int FrameIndex, const TargetRegisterClass *RC, 210 const TargetRegisterInfo *TRI) const override; 211 212 bool expandPostRAPseudo(MachineInstr &MI) const override; 213 214 // Returns an opcode that can be used to move a value to a \p DstRC 215 // register. If there is no hardware instruction that can store to \p 216 // DstRC, then AMDGPU::COPY is returned. 217 unsigned getMovOpcode(const TargetRegisterClass *DstRC) const; 218 219 LLVM_READONLY 220 int commuteOpcode(unsigned Opc) const; 221 222 LLVM_READONLY 223 inline int commuteOpcode(const MachineInstr &MI) const { 224 return commuteOpcode(MI.getOpcode()); 225 } 226 227 bool findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1, 228 unsigned &SrcOpIdx2) const override; 229 230 bool findCommutedOpIndices(MCInstrDesc Desc, unsigned & SrcOpIdx0, 231 unsigned & SrcOpIdx1) const; 232 233 bool isBranchOffsetInRange(unsigned BranchOpc, 234 int64_t BrOffset) const override; 235 236 MachineBasicBlock *getBranchDestBlock(const MachineInstr &MI) const override; 237 238 unsigned insertIndirectBranch(MachineBasicBlock &MBB, 239 MachineBasicBlock &NewDestBB, 240 const DebugLoc &DL, 241 int64_t BrOffset, 242 RegScavenger *RS = nullptr) const override; 243 244 bool analyzeBranchImpl(MachineBasicBlock &MBB, 245 MachineBasicBlock::iterator I, 246 MachineBasicBlock *&TBB, 247 MachineBasicBlock *&FBB, 248 SmallVectorImpl<MachineOperand> &Cond, 249 bool AllowModify) const; 250 251 bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, 252 MachineBasicBlock *&FBB, 253 SmallVectorImpl<MachineOperand> &Cond, 254 bool AllowModify = false) const override; 255 256 unsigned removeBranch(MachineBasicBlock &MBB, 257 int *BytesRemoved = nullptr) const override; 258 259 unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, 260 MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond, 261 const DebugLoc &DL, 262 int *BytesAdded = nullptr) const override; 263 264 bool reverseBranchCondition( 265 SmallVectorImpl<MachineOperand> &Cond) const override; 266 267 bool canInsertSelect(const MachineBasicBlock &MBB, 268 ArrayRef<MachineOperand> Cond, 269 unsigned TrueReg, unsigned FalseReg, 270 int &CondCycles, 271 int &TrueCycles, int &FalseCycles) const override; 272 273 void insertSelect(MachineBasicBlock &MBB, 274 MachineBasicBlock::iterator I, const DebugLoc &DL, 275 unsigned DstReg, ArrayRef<MachineOperand> Cond, 276 unsigned TrueReg, unsigned FalseReg) const override; 277 278 void insertVectorSelect(MachineBasicBlock &MBB, 279 MachineBasicBlock::iterator I, const DebugLoc &DL, 280 unsigned DstReg, ArrayRef<MachineOperand> Cond, 281 unsigned TrueReg, unsigned FalseReg) const; 282 283 unsigned getAddressSpaceForPseudoSourceKind( 284 unsigned Kind) const override; 285 286 bool 287 areMemAccessesTriviallyDisjoint(MachineInstr &MIa, MachineInstr &MIb, 288 AliasAnalysis *AA = nullptr) const override; 289 290 bool isFoldableCopy(const MachineInstr &MI) const; 291 292 bool FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, unsigned Reg, 293 MachineRegisterInfo *MRI) const final; 294 295 unsigned getMachineCSELookAheadLimit() const override { return 500; } 296 297 MachineInstr *convertToThreeAddress(MachineFunction::iterator &MBB, 298 MachineInstr &MI, 299 LiveVariables *LV) const override; 300 301 bool isSchedulingBoundary(const MachineInstr &MI, 302 const MachineBasicBlock *MBB, 303 const MachineFunction &MF) const override; 304 305 static bool isSALU(const MachineInstr &MI) { 306 return MI.getDesc().TSFlags & SIInstrFlags::SALU; 307 } 308 309 bool isSALU(uint16_t Opcode) const { 310 return get(Opcode).TSFlags & SIInstrFlags::SALU; 311 } 312 313 static bool isVALU(const MachineInstr &MI) { 314 return MI.getDesc().TSFlags & SIInstrFlags::VALU; 315 } 316 317 bool isVALU(uint16_t Opcode) const { 318 return get(Opcode).TSFlags & SIInstrFlags::VALU; 319 } 320 321 static bool isVMEM(const MachineInstr &MI) { 322 return isMUBUF(MI) || isMTBUF(MI) || isMIMG(MI); 323 } 324 325 bool isVMEM(uint16_t Opcode) const { 326 return isMUBUF(Opcode) || isMTBUF(Opcode) || isMIMG(Opcode); 327 } 328 329 static bool isSOP1(const MachineInstr &MI) { 330 return MI.getDesc().TSFlags & SIInstrFlags::SOP1; 331 } 332 333 bool isSOP1(uint16_t Opcode) const { 334 return get(Opcode).TSFlags & SIInstrFlags::SOP1; 335 } 336 337 static bool isSOP2(const MachineInstr &MI) { 338 return MI.getDesc().TSFlags & SIInstrFlags::SOP2; 339 } 340 341 bool isSOP2(uint16_t Opcode) const { 342 return get(Opcode).TSFlags & SIInstrFlags::SOP2; 343 } 344 345 static bool isSOPC(const MachineInstr &MI) { 346 return MI.getDesc().TSFlags & SIInstrFlags::SOPC; 347 } 348 349 bool isSOPC(uint16_t Opcode) const { 350 return get(Opcode).TSFlags & SIInstrFlags::SOPC; 351 } 352 353 static bool isSOPK(const MachineInstr &MI) { 354 return MI.getDesc().TSFlags & SIInstrFlags::SOPK; 355 } 356 357 bool isSOPK(uint16_t Opcode) const { 358 return get(Opcode).TSFlags & SIInstrFlags::SOPK; 359 } 360 361 static bool isSOPP(const MachineInstr &MI) { 362 return MI.getDesc().TSFlags & SIInstrFlags::SOPP; 363 } 364 365 bool isSOPP(uint16_t Opcode) const { 366 return get(Opcode).TSFlags & SIInstrFlags::SOPP; 367 } 368 369 static bool isVOP1(const MachineInstr &MI) { 370 return MI.getDesc().TSFlags & SIInstrFlags::VOP1; 371 } 372 373 bool isVOP1(uint16_t Opcode) const { 374 return get(Opcode).TSFlags & SIInstrFlags::VOP1; 375 } 376 377 static bool isVOP2(const MachineInstr &MI) { 378 return MI.getDesc().TSFlags & SIInstrFlags::VOP2; 379 } 380 381 bool isVOP2(uint16_t Opcode) const { 382 return get(Opcode).TSFlags & SIInstrFlags::VOP2; 383 } 384 385 static bool isVOP3(const MachineInstr &MI) { 386 return MI.getDesc().TSFlags & SIInstrFlags::VOP3; 387 } 388 389 bool isVOP3(uint16_t Opcode) const { 390 return get(Opcode).TSFlags & SIInstrFlags::VOP3; 391 } 392 393 static bool isSDWA(const MachineInstr &MI) { 394 return MI.getDesc().TSFlags & SIInstrFlags::SDWA; 395 } 396 397 bool isSDWA(uint16_t Opcode) const { 398 return get(Opcode).TSFlags & SIInstrFlags::SDWA; 399 } 400 401 static bool isVOPC(const MachineInstr &MI) { 402 return MI.getDesc().TSFlags & SIInstrFlags::VOPC; 403 } 404 405 bool isVOPC(uint16_t Opcode) const { 406 return get(Opcode).TSFlags & SIInstrFlags::VOPC; 407 } 408 409 static bool isMUBUF(const MachineInstr &MI) { 410 return MI.getDesc().TSFlags & SIInstrFlags::MUBUF; 411 } 412 413 bool isMUBUF(uint16_t Opcode) const { 414 return get(Opcode).TSFlags & SIInstrFlags::MUBUF; 415 } 416 417 static bool isMTBUF(const MachineInstr &MI) { 418 return MI.getDesc().TSFlags & SIInstrFlags::MTBUF; 419 } 420 421 bool isMTBUF(uint16_t Opcode) const { 422 return get(Opcode).TSFlags & SIInstrFlags::MTBUF; 423 } 424 425 static bool isSMRD(const MachineInstr &MI) { 426 return MI.getDesc().TSFlags & SIInstrFlags::SMRD; 427 } 428 429 bool isSMRD(uint16_t Opcode) const { 430 return get(Opcode).TSFlags & SIInstrFlags::SMRD; 431 } 432 433 bool isBufferSMRD(const MachineInstr &MI) const; 434 435 static bool isDS(const MachineInstr &MI) { 436 return MI.getDesc().TSFlags & SIInstrFlags::DS; 437 } 438 439 bool isDS(uint16_t Opcode) const { 440 return get(Opcode).TSFlags & SIInstrFlags::DS; 441 } 442 443 static bool isMIMG(const MachineInstr &MI) { 444 return MI.getDesc().TSFlags & SIInstrFlags::MIMG; 445 } 446 447 bool isMIMG(uint16_t Opcode) const { 448 return get(Opcode).TSFlags & SIInstrFlags::MIMG; 449 } 450 451 static bool isGather4(const MachineInstr &MI) { 452 return MI.getDesc().TSFlags & SIInstrFlags::Gather4; 453 } 454 455 bool isGather4(uint16_t Opcode) const { 456 return get(Opcode).TSFlags & SIInstrFlags::Gather4; 457 } 458 459 static bool isFLAT(const MachineInstr &MI) { 460 return MI.getDesc().TSFlags & SIInstrFlags::FLAT; 461 } 462 463 // Is a FLAT encoded instruction which accesses a specific segment, 464 // i.e. global_* or scratch_*. 465 static bool isSegmentSpecificFLAT(const MachineInstr &MI) { 466 auto Flags = MI.getDesc().TSFlags; 467 return (Flags & SIInstrFlags::FLAT) && !(Flags & SIInstrFlags::LGKM_CNT); 468 } 469 470 // Any FLAT encoded instruction, including global_* and scratch_*. 471 bool isFLAT(uint16_t Opcode) const { 472 return get(Opcode).TSFlags & SIInstrFlags::FLAT; 473 } 474 475 static bool isEXP(const MachineInstr &MI) { 476 return MI.getDesc().TSFlags & SIInstrFlags::EXP; 477 } 478 479 bool isEXP(uint16_t Opcode) const { 480 return get(Opcode).TSFlags & SIInstrFlags::EXP; 481 } 482 483 static bool isWQM(const MachineInstr &MI) { 484 return MI.getDesc().TSFlags & SIInstrFlags::WQM; 485 } 486 487 bool isWQM(uint16_t Opcode) const { 488 return get(Opcode).TSFlags & SIInstrFlags::WQM; 489 } 490 491 static bool isDisableWQM(const MachineInstr &MI) { 492 return MI.getDesc().TSFlags & SIInstrFlags::DisableWQM; 493 } 494 495 bool isDisableWQM(uint16_t Opcode) const { 496 return get(Opcode).TSFlags & SIInstrFlags::DisableWQM; 497 } 498 499 static bool isVGPRSpill(const MachineInstr &MI) { 500 return MI.getDesc().TSFlags & SIInstrFlags::VGPRSpill; 501 } 502 503 bool isVGPRSpill(uint16_t Opcode) const { 504 return get(Opcode).TSFlags & SIInstrFlags::VGPRSpill; 505 } 506 507 static bool isSGPRSpill(const MachineInstr &MI) { 508 return MI.getDesc().TSFlags & SIInstrFlags::SGPRSpill; 509 } 510 511 bool isSGPRSpill(uint16_t Opcode) const { 512 return get(Opcode).TSFlags & SIInstrFlags::SGPRSpill; 513 } 514 515 static bool isDPP(const MachineInstr &MI) { 516 return MI.getDesc().TSFlags & SIInstrFlags::DPP; 517 } 518 519 bool isDPP(uint16_t Opcode) const { 520 return get(Opcode).TSFlags & SIInstrFlags::DPP; 521 } 522 523 static bool isVOP3P(const MachineInstr &MI) { 524 return MI.getDesc().TSFlags & SIInstrFlags::VOP3P; 525 } 526 527 bool isVOP3P(uint16_t Opcode) const { 528 return get(Opcode).TSFlags & SIInstrFlags::VOP3P; 529 } 530 531 static bool isVINTRP(const MachineInstr &MI) { 532 return MI.getDesc().TSFlags & SIInstrFlags::VINTRP; 533 } 534 535 bool isVINTRP(uint16_t Opcode) const { 536 return get(Opcode).TSFlags & SIInstrFlags::VINTRP; 537 } 538 539 static bool isScalarUnit(const MachineInstr &MI) { 540 return MI.getDesc().TSFlags & (SIInstrFlags::SALU | SIInstrFlags::SMRD); 541 } 542 543 static bool usesVM_CNT(const MachineInstr &MI) { 544 return MI.getDesc().TSFlags & SIInstrFlags::VM_CNT; 545 } 546 547 static bool usesLGKM_CNT(const MachineInstr &MI) { 548 return MI.getDesc().TSFlags & SIInstrFlags::LGKM_CNT; 549 } 550 551 static bool sopkIsZext(const MachineInstr &MI) { 552 return MI.getDesc().TSFlags & SIInstrFlags::SOPK_ZEXT; 553 } 554 555 bool sopkIsZext(uint16_t Opcode) const { 556 return get(Opcode).TSFlags & SIInstrFlags::SOPK_ZEXT; 557 } 558 559 /// \returns true if this is an s_store_dword* instruction. This is more 560 /// specific than than isSMEM && mayStore. 561 static bool isScalarStore(const MachineInstr &MI) { 562 return MI.getDesc().TSFlags & SIInstrFlags::SCALAR_STORE; 563 } 564 565 bool isScalarStore(uint16_t Opcode) const { 566 return get(Opcode).TSFlags & SIInstrFlags::SCALAR_STORE; 567 } 568 569 static bool isFixedSize(const MachineInstr &MI) { 570 return MI.getDesc().TSFlags & SIInstrFlags::FIXED_SIZE; 571 } 572 573 bool isFixedSize(uint16_t Opcode) const { 574 return get(Opcode).TSFlags & SIInstrFlags::FIXED_SIZE; 575 } 576 577 static bool hasFPClamp(const MachineInstr &MI) { 578 return MI.getDesc().TSFlags & SIInstrFlags::FPClamp; 579 } 580 581 bool hasFPClamp(uint16_t Opcode) const { 582 return get(Opcode).TSFlags & SIInstrFlags::FPClamp; 583 } 584 585 static bool hasIntClamp(const MachineInstr &MI) { 586 return MI.getDesc().TSFlags & SIInstrFlags::IntClamp; 587 } 588 589 uint64_t getClampMask(const MachineInstr &MI) const { 590 const uint64_t ClampFlags = SIInstrFlags::FPClamp | 591 SIInstrFlags::IntClamp | 592 SIInstrFlags::ClampLo | 593 SIInstrFlags::ClampHi; 594 return MI.getDesc().TSFlags & ClampFlags; 595 } 596 597 bool isVGPRCopy(const MachineInstr &MI) const { 598 assert(MI.isCopy()); 599 unsigned Dest = MI.getOperand(0).getReg(); 600 const MachineFunction &MF = *MI.getParent()->getParent(); 601 const MachineRegisterInfo &MRI = MF.getRegInfo(); 602 return !RI.isSGPRReg(MRI, Dest); 603 } 604 605 /// Whether we must prevent this instruction from executing with EXEC = 0. 606 bool hasUnwantedEffectsWhenEXECEmpty(const MachineInstr &MI) const; 607 608 bool isInlineConstant(const APInt &Imm) const; 609 610 bool isInlineConstant(const MachineOperand &MO, uint8_t OperandType) const; 611 612 bool isInlineConstant(const MachineOperand &MO, 613 const MCOperandInfo &OpInfo) const { 614 return isInlineConstant(MO, OpInfo.OperandType); 615 } 616 617 /// \p returns true if \p UseMO is substituted with \p DefMO in \p MI it would 618 /// be an inline immediate. 619 bool isInlineConstant(const MachineInstr &MI, 620 const MachineOperand &UseMO, 621 const MachineOperand &DefMO) const { 622 assert(UseMO.getParent() == &MI); 623 int OpIdx = MI.getOperandNo(&UseMO); 624 if (!MI.getDesc().OpInfo || OpIdx >= MI.getDesc().NumOperands) { 625 return false; 626 } 627 628 return isInlineConstant(DefMO, MI.getDesc().OpInfo[OpIdx]); 629 } 630 631 /// \p returns true if the operand \p OpIdx in \p MI is a valid inline 632 /// immediate. 633 bool isInlineConstant(const MachineInstr &MI, unsigned OpIdx) const { 634 const MachineOperand &MO = MI.getOperand(OpIdx); 635 return isInlineConstant(MO, MI.getDesc().OpInfo[OpIdx].OperandType); 636 } 637 638 bool isInlineConstant(const MachineInstr &MI, unsigned OpIdx, 639 const MachineOperand &MO) const { 640 if (!MI.getDesc().OpInfo || OpIdx >= MI.getDesc().NumOperands) 641 return false; 642 643 if (MI.isCopy()) { 644 unsigned Size = getOpSize(MI, OpIdx); 645 assert(Size == 8 || Size == 4); 646 647 uint8_t OpType = (Size == 8) ? 648 AMDGPU::OPERAND_REG_IMM_INT64 : AMDGPU::OPERAND_REG_IMM_INT32; 649 return isInlineConstant(MO, OpType); 650 } 651 652 return isInlineConstant(MO, MI.getDesc().OpInfo[OpIdx].OperandType); 653 } 654 655 bool isInlineConstant(const MachineOperand &MO) const { 656 const MachineInstr *Parent = MO.getParent(); 657 return isInlineConstant(*Parent, Parent->getOperandNo(&MO)); 658 } 659 660 bool isLiteralConstant(const MachineOperand &MO, 661 const MCOperandInfo &OpInfo) const { 662 return MO.isImm() && !isInlineConstant(MO, OpInfo.OperandType); 663 } 664 665 bool isLiteralConstant(const MachineInstr &MI, int OpIdx) const { 666 const MachineOperand &MO = MI.getOperand(OpIdx); 667 return MO.isImm() && !isInlineConstant(MI, OpIdx); 668 } 669 670 // Returns true if this operand could potentially require a 32-bit literal 671 // operand, but not necessarily. A FrameIndex for example could resolve to an 672 // inline immediate value that will not require an additional 4-bytes; this 673 // assumes that it will. 674 bool isLiteralConstantLike(const MachineOperand &MO, 675 const MCOperandInfo &OpInfo) const; 676 677 bool isImmOperandLegal(const MachineInstr &MI, unsigned OpNo, 678 const MachineOperand &MO) const; 679 680 /// Return true if this 64-bit VALU instruction has a 32-bit encoding. 681 /// This function will return false if you pass it a 32-bit instruction. 682 bool hasVALU32BitEncoding(unsigned Opcode) const; 683 684 /// Returns true if this operand uses the constant bus. 685 bool usesConstantBus(const MachineRegisterInfo &MRI, 686 const MachineOperand &MO, 687 const MCOperandInfo &OpInfo) const; 688 689 /// Return true if this instruction has any modifiers. 690 /// e.g. src[012]_mod, omod, clamp. 691 bool hasModifiers(unsigned Opcode) const; 692 693 bool hasModifiersSet(const MachineInstr &MI, 694 unsigned OpName) const; 695 bool hasAnyModifiersSet(const MachineInstr &MI) const; 696 697 bool canShrink(const MachineInstr &MI, 698 const MachineRegisterInfo &MRI) const; 699 700 MachineInstr *buildShrunkInst(MachineInstr &MI, 701 unsigned NewOpcode) const; 702 703 bool verifyInstruction(const MachineInstr &MI, 704 StringRef &ErrInfo) const override; 705 706 unsigned getVALUOp(const MachineInstr &MI) const; 707 708 /// Return the correct register class for \p OpNo. For target-specific 709 /// instructions, this will return the register class that has been defined 710 /// in tablegen. For generic instructions, like REG_SEQUENCE it will return 711 /// the register class of its machine operand. 712 /// to infer the correct register class base on the other operands. 713 const TargetRegisterClass *getOpRegClass(const MachineInstr &MI, 714 unsigned OpNo) const; 715 716 /// Return the size in bytes of the operand OpNo on the given 717 // instruction opcode. 718 unsigned getOpSize(uint16_t Opcode, unsigned OpNo) const { 719 const MCOperandInfo &OpInfo = get(Opcode).OpInfo[OpNo]; 720 721 if (OpInfo.RegClass == -1) { 722 // If this is an immediate operand, this must be a 32-bit literal. 723 assert(OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE); 724 return 4; 725 } 726 727 return RI.getRegSizeInBits(*RI.getRegClass(OpInfo.RegClass)) / 8; 728 } 729 730 /// This form should usually be preferred since it handles operands 731 /// with unknown register classes. 732 unsigned getOpSize(const MachineInstr &MI, unsigned OpNo) const { 733 const MachineOperand &MO = MI.getOperand(OpNo); 734 if (MO.isReg()) { 735 if (unsigned SubReg = MO.getSubReg()) { 736 assert(RI.getRegSizeInBits(*RI.getSubClassWithSubReg( 737 MI.getParent()->getParent()->getRegInfo(). 738 getRegClass(MO.getReg()), SubReg)) >= 32 && 739 "Sub-dword subregs are not supported"); 740 return RI.getSubRegIndexLaneMask(SubReg).getNumLanes() * 4; 741 } 742 } 743 return RI.getRegSizeInBits(*getOpRegClass(MI, OpNo)) / 8; 744 } 745 746 /// \returns true if it is legal for the operand at index \p OpNo 747 /// to read a VGPR. 748 bool canReadVGPR(const MachineInstr &MI, unsigned OpNo) const; 749 750 /// Legalize the \p OpIndex operand of this instruction by inserting 751 /// a MOV. For example: 752 /// ADD_I32_e32 VGPR0, 15 753 /// to 754 /// MOV VGPR1, 15 755 /// ADD_I32_e32 VGPR0, VGPR1 756 /// 757 /// If the operand being legalized is a register, then a COPY will be used 758 /// instead of MOV. 759 void legalizeOpWithMove(MachineInstr &MI, unsigned OpIdx) const; 760 761 /// Check if \p MO is a legal operand if it was the \p OpIdx Operand 762 /// for \p MI. 763 bool isOperandLegal(const MachineInstr &MI, unsigned OpIdx, 764 const MachineOperand *MO = nullptr) const; 765 766 /// Check if \p MO would be a valid operand for the given operand 767 /// definition \p OpInfo. Note this does not attempt to validate constant bus 768 /// restrictions (e.g. literal constant usage). 769 bool isLegalVSrcOperand(const MachineRegisterInfo &MRI, 770 const MCOperandInfo &OpInfo, 771 const MachineOperand &MO) const; 772 773 /// Check if \p MO (a register operand) is a legal register for the 774 /// given operand description. 775 bool isLegalRegOperand(const MachineRegisterInfo &MRI, 776 const MCOperandInfo &OpInfo, 777 const MachineOperand &MO) const; 778 779 /// Legalize operands in \p MI by either commuting it or inserting a 780 /// copy of src1. 781 void legalizeOperandsVOP2(MachineRegisterInfo &MRI, MachineInstr &MI) const; 782 783 /// Fix operands in \p MI to satisfy constant bus requirements. 784 void legalizeOperandsVOP3(MachineRegisterInfo &MRI, MachineInstr &MI) const; 785 786 /// Copy a value from a VGPR (\p SrcReg) to SGPR. This function can only 787 /// be used when it is know that the value in SrcReg is same across all 788 /// threads in the wave. 789 /// \returns The SGPR register that \p SrcReg was copied to. 790 unsigned readlaneVGPRToSGPR(unsigned SrcReg, MachineInstr &UseMI, 791 MachineRegisterInfo &MRI) const; 792 793 void legalizeOperandsSMRD(MachineRegisterInfo &MRI, MachineInstr &MI) const; 794 795 void legalizeGenericOperand(MachineBasicBlock &InsertMBB, 796 MachineBasicBlock::iterator I, 797 const TargetRegisterClass *DstRC, 798 MachineOperand &Op, MachineRegisterInfo &MRI, 799 const DebugLoc &DL) const; 800 801 /// Legalize all operands in this instruction. This function may create new 802 /// instructions and control-flow around \p MI. If present, \p MDT is 803 /// updated. 804 void legalizeOperands(MachineInstr &MI, 805 MachineDominatorTree *MDT = nullptr) const; 806 807 /// Replace this instruction's opcode with the equivalent VALU 808 /// opcode. This function will also move the users of \p MI to the 809 /// VALU if necessary. If present, \p MDT is updated. 810 void moveToVALU(MachineInstr &MI, MachineDominatorTree *MDT = nullptr) const; 811 812 void insertWaitStates(MachineBasicBlock &MBB,MachineBasicBlock::iterator MI, 813 int Count) const; 814 815 void insertNoop(MachineBasicBlock &MBB, 816 MachineBasicBlock::iterator MI) const override; 817 818 void insertReturn(MachineBasicBlock &MBB) const; 819 /// Return the number of wait states that result from executing this 820 /// instruction. 821 unsigned getNumWaitStates(const MachineInstr &MI) const; 822 823 /// Returns the operand named \p Op. If \p MI does not have an 824 /// operand named \c Op, this function returns nullptr. 825 LLVM_READONLY 826 MachineOperand *getNamedOperand(MachineInstr &MI, unsigned OperandName) const; 827 828 LLVM_READONLY 829 const MachineOperand *getNamedOperand(const MachineInstr &MI, 830 unsigned OpName) const { 831 return getNamedOperand(const_cast<MachineInstr &>(MI), OpName); 832 } 833 834 /// Get required immediate operand 835 int64_t getNamedImmOperand(const MachineInstr &MI, unsigned OpName) const { 836 int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OpName); 837 return MI.getOperand(Idx).getImm(); 838 } 839 840 uint64_t getDefaultRsrcDataFormat() const; 841 uint64_t getScratchRsrcWords23() const; 842 843 bool isLowLatencyInstruction(const MachineInstr &MI) const; 844 bool isHighLatencyInstruction(const MachineInstr &MI) const; 845 846 /// Return the descriptor of the target-specific machine instruction 847 /// that corresponds to the specified pseudo or native opcode. 848 const MCInstrDesc &getMCOpcodeFromPseudo(unsigned Opcode) const { 849 return get(pseudoToMCOpcode(Opcode)); 850 } 851 852 unsigned isStackAccess(const MachineInstr &MI, int &FrameIndex) const; 853 unsigned isSGPRStackAccess(const MachineInstr &MI, int &FrameIndex) const; 854 855 unsigned isLoadFromStackSlot(const MachineInstr &MI, 856 int &FrameIndex) const override; 857 unsigned isStoreToStackSlot(const MachineInstr &MI, 858 int &FrameIndex) const override; 859 860 unsigned getInstBundleSize(const MachineInstr &MI) const; 861 unsigned getInstSizeInBytes(const MachineInstr &MI) const override; 862 863 bool mayAccessFlatAddressSpace(const MachineInstr &MI) const; 864 865 bool isNonUniformBranchInstr(MachineInstr &Instr) const; 866 867 void convertNonUniformIfRegion(MachineBasicBlock *IfEntry, 868 MachineBasicBlock *IfEnd) const; 869 870 void convertNonUniformLoopRegion(MachineBasicBlock *LoopEntry, 871 MachineBasicBlock *LoopEnd) const; 872 873 std::pair<unsigned, unsigned> 874 decomposeMachineOperandsTargetFlags(unsigned TF) const override; 875 876 ArrayRef<std::pair<int, const char *>> 877 getSerializableTargetIndices() const override; 878 879 ArrayRef<std::pair<unsigned, const char *>> 880 getSerializableDirectMachineOperandTargetFlags() const override; 881 882 ScheduleHazardRecognizer * 883 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, 884 const ScheduleDAG *DAG) const override; 885 886 ScheduleHazardRecognizer * 887 CreateTargetPostRAHazardRecognizer(const MachineFunction &MF) const override; 888 889 bool isBasicBlockPrologue(const MachineInstr &MI) const override; 890 891 /// Return a partially built integer add instruction without carry. 892 /// Caller must add source operands. 893 /// For pre-GFX9 it will generate unused carry destination operand. 894 /// TODO: After GFX9 it should return a no-carry operation. 895 MachineInstrBuilder getAddNoCarry(MachineBasicBlock &MBB, 896 MachineBasicBlock::iterator I, 897 const DebugLoc &DL, 898 unsigned DestReg) const; 899 900 static bool isKillTerminator(unsigned Opcode); 901 const MCInstrDesc &getKillTerminatorFromPseudo(unsigned Opcode) const; 902 903 static bool isLegalMUBUFImmOffset(unsigned Imm) { 904 return isUInt<12>(Imm); 905 } 906 907 /// \brief Return a target-specific opcode if Opcode is a pseudo instruction. 908 /// Return -1 if the target-specific opcode for the pseudo instruction does 909 /// not exist. If Opcode is not a pseudo instruction, this is identity. 910 int pseudoToMCOpcode(int Opcode) const; 911 912 }; 913 914 namespace AMDGPU { 915 916 LLVM_READONLY 917 int getVOPe64(uint16_t Opcode); 918 919 LLVM_READONLY 920 int getVOPe32(uint16_t Opcode); 921 922 LLVM_READONLY 923 int getSDWAOp(uint16_t Opcode); 924 925 LLVM_READONLY 926 int getBasicFromSDWAOp(uint16_t Opcode); 927 928 LLVM_READONLY 929 int getCommuteRev(uint16_t Opcode); 930 931 LLVM_READONLY 932 int getCommuteOrig(uint16_t Opcode); 933 934 LLVM_READONLY 935 int getAddr64Inst(uint16_t Opcode); 936 937 /// Check if \p Opcode is an Addr64 opcode. 938 /// 939 /// \returns \p Opcode if it is an Addr64 opcode, otherwise -1. 940 LLVM_READONLY 941 int getIfAddr64Inst(uint16_t Opcode); 942 943 LLVM_READONLY 944 int getMUBUFNoLdsInst(uint16_t Opcode); 945 946 LLVM_READONLY 947 int getAtomicRetOp(uint16_t Opcode); 948 949 LLVM_READONLY 950 int getAtomicNoRetOp(uint16_t Opcode); 951 952 LLVM_READONLY 953 int getSOPKOp(uint16_t Opcode); 954 955 const uint64_t RSRC_DATA_FORMAT = 0xf00000000000LL; 956 const uint64_t RSRC_ELEMENT_SIZE_SHIFT = (32 + 19); 957 const uint64_t RSRC_INDEX_STRIDE_SHIFT = (32 + 21); 958 const uint64_t RSRC_TID_ENABLE = UINT64_C(1) << (32 + 23); 959 960 // For MachineOperands. 961 enum TargetFlags { 962 TF_LONG_BRANCH_FORWARD = 1 << 0, 963 TF_LONG_BRANCH_BACKWARD = 1 << 1 964 }; 965 966 } // end namespace AMDGPU 967 968 namespace SI { 969 namespace KernelInputOffsets { 970 971 /// Offsets in bytes from the start of the input buffer 972 enum Offsets { 973 NGROUPS_X = 0, 974 NGROUPS_Y = 4, 975 NGROUPS_Z = 8, 976 GLOBAL_SIZE_X = 12, 977 GLOBAL_SIZE_Y = 16, 978 GLOBAL_SIZE_Z = 20, 979 LOCAL_SIZE_X = 24, 980 LOCAL_SIZE_Y = 28, 981 LOCAL_SIZE_Z = 32 982 }; 983 984 } // end namespace KernelInputOffsets 985 } // end namespace SI 986 987 } // end namespace llvm 988 989 #endif // LLVM_LIB_TARGET_AMDGPU_SIINSTRINFO_H 990