1 //===-- GCNHazardRecognizers.cpp - GCN Hazard Recognizer Impls ------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements hazard recognizers for scheduling on GCN processors. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "GCNHazardRecognizer.h" 14 #include "GCNSubtarget.h" 15 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 16 #include "llvm/CodeGen/MachineFunction.h" 17 #include "llvm/CodeGen/ScheduleDAG.h" 18 #include "llvm/Support/TargetParser.h" 19 20 using namespace llvm; 21 22 //===----------------------------------------------------------------------===// 23 // Hazard Recoginizer Implementation 24 //===----------------------------------------------------------------------===// 25 26 static bool shouldRunLdsBranchVmemWARHazardFixup(const MachineFunction &MF, 27 const GCNSubtarget &ST); 28 29 GCNHazardRecognizer::GCNHazardRecognizer(const MachineFunction &MF) : 30 IsHazardRecognizerMode(false), 31 CurrCycleInstr(nullptr), 32 MF(MF), 33 ST(MF.getSubtarget<GCNSubtarget>()), 34 TII(*ST.getInstrInfo()), 35 TRI(TII.getRegisterInfo()), 36 ClauseUses(TRI.getNumRegUnits()), 37 ClauseDefs(TRI.getNumRegUnits()) { 38 MaxLookAhead = MF.getRegInfo().isPhysRegUsed(AMDGPU::AGPR0) ? 19 : 5; 39 TSchedModel.init(&ST); 40 RunLdsBranchVmemWARHazardFixup = shouldRunLdsBranchVmemWARHazardFixup(MF, ST); 41 } 42 43 void GCNHazardRecognizer::Reset() { 44 EmittedInstrs.clear(); 45 } 46 47 void GCNHazardRecognizer::EmitInstruction(SUnit *SU) { 48 EmitInstruction(SU->getInstr()); 49 } 50 51 void GCNHazardRecognizer::EmitInstruction(MachineInstr *MI) { 52 CurrCycleInstr = MI; 53 } 54 55 static bool isDivFMas(unsigned Opcode) { 56 return Opcode == AMDGPU::V_DIV_FMAS_F32_e64 || Opcode == AMDGPU::V_DIV_FMAS_F64_e64; 57 } 58 59 static bool isSGetReg(unsigned Opcode) { 60 return Opcode == AMDGPU::S_GETREG_B32; 61 } 62 63 static bool isSSetReg(unsigned Opcode) { 64 switch (Opcode) { 65 case AMDGPU::S_SETREG_B32: 66 case AMDGPU::S_SETREG_B32_mode: 67 case AMDGPU::S_SETREG_IMM32_B32: 68 case AMDGPU::S_SETREG_IMM32_B32_mode: 69 return true; 70 } 71 return false; 72 } 73 74 static bool isRWLane(unsigned Opcode) { 75 return Opcode == AMDGPU::V_READLANE_B32 || Opcode == AMDGPU::V_WRITELANE_B32; 76 } 77 78 static bool isRFE(unsigned Opcode) { 79 return Opcode == AMDGPU::S_RFE_B64; 80 } 81 82 static bool isSMovRel(unsigned Opcode) { 83 switch (Opcode) { 84 case AMDGPU::S_MOVRELS_B32: 85 case AMDGPU::S_MOVRELS_B64: 86 case AMDGPU::S_MOVRELD_B32: 87 case AMDGPU::S_MOVRELD_B64: 88 return true; 89 default: 90 return false; 91 } 92 } 93 94 static bool isDGEMM(unsigned Opcode) { 95 return Opcode == AMDGPU::V_MFMA_F64_4X4X4F64_e64 || 96 Opcode == AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64 || 97 Opcode == AMDGPU::V_MFMA_F64_16X16X4F64_e64 || 98 Opcode == AMDGPU::V_MFMA_F64_16X16X4F64_vgprcd_e64; 99 } 100 101 static bool isXDL(const GCNSubtarget &ST, const MachineInstr &MI) { 102 unsigned Opcode = MI.getOpcode(); 103 104 if (!SIInstrInfo::isMAI(MI) || 105 isDGEMM(Opcode) || 106 Opcode == AMDGPU::V_ACCVGPR_WRITE_B32_e64 || 107 Opcode == AMDGPU::V_ACCVGPR_READ_B32_e64) 108 return false; 109 110 return true; 111 } 112 113 static bool isSendMsgTraceDataOrGDS(const SIInstrInfo &TII, 114 const MachineInstr &MI) { 115 if (TII.isAlwaysGDS(MI.getOpcode())) 116 return true; 117 118 switch (MI.getOpcode()) { 119 case AMDGPU::S_SENDMSG: 120 case AMDGPU::S_SENDMSGHALT: 121 case AMDGPU::S_TTRACEDATA: 122 return true; 123 // These DS opcodes don't support GDS. 124 case AMDGPU::DS_NOP: 125 case AMDGPU::DS_PERMUTE_B32: 126 case AMDGPU::DS_BPERMUTE_B32: 127 return false; 128 default: 129 if (TII.isDS(MI.getOpcode())) { 130 int GDS = AMDGPU::getNamedOperandIdx(MI.getOpcode(), 131 AMDGPU::OpName::gds); 132 if (MI.getOperand(GDS).getImm()) 133 return true; 134 } 135 return false; 136 } 137 } 138 139 static bool isPermlane(const MachineInstr &MI) { 140 unsigned Opcode = MI.getOpcode(); 141 return Opcode == AMDGPU::V_PERMLANE16_B32_e64 || 142 Opcode == AMDGPU::V_PERMLANEX16_B32_e64; 143 } 144 145 static unsigned getHWReg(const SIInstrInfo *TII, const MachineInstr &RegInstr) { 146 const MachineOperand *RegOp = TII->getNamedOperand(RegInstr, 147 AMDGPU::OpName::simm16); 148 return RegOp->getImm() & AMDGPU::Hwreg::ID_MASK_; 149 } 150 151 ScheduleHazardRecognizer::HazardType 152 GCNHazardRecognizer::getHazardType(SUnit *SU, int Stalls) { 153 MachineInstr *MI = SU->getInstr(); 154 // If we are not in "HazardRecognizerMode" and therefore not being run from 155 // the scheduler, track possible stalls from hazards but don't insert noops. 156 auto HazardType = IsHazardRecognizerMode ? NoopHazard : Hazard; 157 158 if (MI->isBundle()) 159 return NoHazard; 160 161 if (SIInstrInfo::isSMRD(*MI) && checkSMRDHazards(MI) > 0) 162 return HazardType; 163 164 if (ST.hasNSAtoVMEMBug() && checkNSAtoVMEMHazard(MI) > 0) 165 return HazardType; 166 167 if (checkFPAtomicToDenormModeHazard(MI) > 0) 168 return HazardType; 169 170 if (ST.hasNoDataDepHazard()) 171 return NoHazard; 172 173 // FIXME: Should flat be considered vmem? 174 if ((SIInstrInfo::isVMEM(*MI) || 175 SIInstrInfo::isFLAT(*MI)) 176 && checkVMEMHazards(MI) > 0) 177 return HazardType; 178 179 if (SIInstrInfo::isVALU(*MI) && checkVALUHazards(MI) > 0) 180 return HazardType; 181 182 if (SIInstrInfo::isDPP(*MI) && checkDPPHazards(MI) > 0) 183 return HazardType; 184 185 if (isDivFMas(MI->getOpcode()) && checkDivFMasHazards(MI) > 0) 186 return HazardType; 187 188 if (isRWLane(MI->getOpcode()) && checkRWLaneHazards(MI) > 0) 189 return HazardType; 190 191 if ((SIInstrInfo::isVALU(*MI) || SIInstrInfo::isVMEM(*MI) || 192 SIInstrInfo::isFLAT(*MI) || SIInstrInfo::isDS(*MI) || 193 SIInstrInfo::isEXP(*MI)) && checkMAIVALUHazards(MI) > 0) 194 return HazardType; 195 196 if (isSGetReg(MI->getOpcode()) && checkGetRegHazards(MI) > 0) 197 return HazardType; 198 199 if (isSSetReg(MI->getOpcode()) && checkSetRegHazards(MI) > 0) 200 return HazardType; 201 202 if (isRFE(MI->getOpcode()) && checkRFEHazards(MI) > 0) 203 return HazardType; 204 205 if (ST.hasReadM0MovRelInterpHazard() && 206 (TII.isVINTRP(*MI) || isSMovRel(MI->getOpcode())) && 207 checkReadM0Hazards(MI) > 0) 208 return HazardType; 209 210 if (ST.hasReadM0SendMsgHazard() && isSendMsgTraceDataOrGDS(TII, *MI) && 211 checkReadM0Hazards(MI) > 0) 212 return HazardType; 213 214 if (SIInstrInfo::isMAI(*MI) && checkMAIHazards(MI) > 0) 215 return HazardType; 216 217 if ((SIInstrInfo::isVMEM(*MI) || 218 SIInstrInfo::isFLAT(*MI) || 219 SIInstrInfo::isDS(*MI)) && checkMAILdStHazards(MI) > 0) 220 return HazardType; 221 222 if (MI->isInlineAsm() && checkInlineAsmHazards(MI) > 0) 223 return HazardType; 224 225 return NoHazard; 226 } 227 228 static void insertNoopsInBundle(MachineInstr *MI, const SIInstrInfo &TII, 229 unsigned Quantity) { 230 while (Quantity > 0) { 231 unsigned Arg = std::min(Quantity, 8u); 232 Quantity -= Arg; 233 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), TII.get(AMDGPU::S_NOP)) 234 .addImm(Arg - 1); 235 } 236 } 237 238 void GCNHazardRecognizer::processBundle() { 239 MachineBasicBlock::instr_iterator MI = std::next(CurrCycleInstr->getIterator()); 240 MachineBasicBlock::instr_iterator E = CurrCycleInstr->getParent()->instr_end(); 241 // Check bundled MachineInstr's for hazards. 242 for (; MI != E && MI->isInsideBundle(); ++MI) { 243 CurrCycleInstr = &*MI; 244 unsigned WaitStates = PreEmitNoopsCommon(CurrCycleInstr); 245 246 if (IsHazardRecognizerMode) { 247 fixHazards(CurrCycleInstr); 248 249 insertNoopsInBundle(CurrCycleInstr, TII, WaitStates); 250 } 251 252 // It’s unnecessary to track more than MaxLookAhead instructions. Since we 253 // include the bundled MI directly after, only add a maximum of 254 // (MaxLookAhead - 1) noops to EmittedInstrs. 255 for (unsigned i = 0, e = std::min(WaitStates, MaxLookAhead - 1); i < e; ++i) 256 EmittedInstrs.push_front(nullptr); 257 258 EmittedInstrs.push_front(CurrCycleInstr); 259 EmittedInstrs.resize(MaxLookAhead); 260 } 261 CurrCycleInstr = nullptr; 262 } 263 264 unsigned GCNHazardRecognizer::PreEmitNoops(MachineInstr *MI) { 265 IsHazardRecognizerMode = true; 266 CurrCycleInstr = MI; 267 unsigned W = PreEmitNoopsCommon(MI); 268 fixHazards(MI); 269 CurrCycleInstr = nullptr; 270 return W; 271 } 272 273 unsigned GCNHazardRecognizer::PreEmitNoopsCommon(MachineInstr *MI) { 274 if (MI->isBundle()) 275 return 0; 276 277 int WaitStates = 0; 278 279 if (SIInstrInfo::isSMRD(*MI)) 280 return std::max(WaitStates, checkSMRDHazards(MI)); 281 282 if (ST.hasNSAtoVMEMBug()) 283 WaitStates = std::max(WaitStates, checkNSAtoVMEMHazard(MI)); 284 285 WaitStates = std::max(WaitStates, checkFPAtomicToDenormModeHazard(MI)); 286 287 if (ST.hasNoDataDepHazard()) 288 return WaitStates; 289 290 if (SIInstrInfo::isVMEM(*MI) || SIInstrInfo::isFLAT(*MI)) 291 WaitStates = std::max(WaitStates, checkVMEMHazards(MI)); 292 293 if (SIInstrInfo::isVALU(*MI)) 294 WaitStates = std::max(WaitStates, checkVALUHazards(MI)); 295 296 if (SIInstrInfo::isDPP(*MI)) 297 WaitStates = std::max(WaitStates, checkDPPHazards(MI)); 298 299 if (isDivFMas(MI->getOpcode())) 300 WaitStates = std::max(WaitStates, checkDivFMasHazards(MI)); 301 302 if (isRWLane(MI->getOpcode())) 303 WaitStates = std::max(WaitStates, checkRWLaneHazards(MI)); 304 305 if ((SIInstrInfo::isVALU(*MI) || SIInstrInfo::isVMEM(*MI) || 306 SIInstrInfo::isFLAT(*MI) || SIInstrInfo::isDS(*MI) || 307 SIInstrInfo::isEXP(*MI)) && checkMAIVALUHazards(MI) > 0) 308 WaitStates = std::max(WaitStates, checkMAIVALUHazards(MI)); 309 310 if (MI->isInlineAsm()) 311 return std::max(WaitStates, checkInlineAsmHazards(MI)); 312 313 if (isSGetReg(MI->getOpcode())) 314 return std::max(WaitStates, checkGetRegHazards(MI)); 315 316 if (isSSetReg(MI->getOpcode())) 317 return std::max(WaitStates, checkSetRegHazards(MI)); 318 319 if (isRFE(MI->getOpcode())) 320 return std::max(WaitStates, checkRFEHazards(MI)); 321 322 if (ST.hasReadM0MovRelInterpHazard() && (TII.isVINTRP(*MI) || 323 isSMovRel(MI->getOpcode()))) 324 return std::max(WaitStates, checkReadM0Hazards(MI)); 325 326 if (ST.hasReadM0SendMsgHazard() && isSendMsgTraceDataOrGDS(TII, *MI)) 327 return std::max(WaitStates, checkReadM0Hazards(MI)); 328 329 if (SIInstrInfo::isMAI(*MI)) 330 return std::max(WaitStates, checkMAIHazards(MI)); 331 332 if (SIInstrInfo::isVMEM(*MI) || 333 SIInstrInfo::isFLAT(*MI) || 334 SIInstrInfo::isDS(*MI)) 335 return std::max(WaitStates, checkMAILdStHazards(MI)); 336 337 return WaitStates; 338 } 339 340 void GCNHazardRecognizer::EmitNoop() { 341 EmittedInstrs.push_front(nullptr); 342 } 343 344 void GCNHazardRecognizer::AdvanceCycle() { 345 // When the scheduler detects a stall, it will call AdvanceCycle() without 346 // emitting any instructions. 347 if (!CurrCycleInstr) { 348 EmittedInstrs.push_front(nullptr); 349 return; 350 } 351 352 // Do not track non-instructions which do not affect the wait states. 353 // If included, these instructions can lead to buffer overflow such that 354 // detectable hazards are missed. 355 if (CurrCycleInstr->isMetaInstruction()) { 356 CurrCycleInstr = nullptr; 357 return; 358 } 359 360 if (CurrCycleInstr->isBundle()) { 361 processBundle(); 362 return; 363 } 364 365 unsigned NumWaitStates = TII.getNumWaitStates(*CurrCycleInstr); 366 if (!NumWaitStates) { 367 CurrCycleInstr = nullptr; 368 return; 369 } 370 371 // Keep track of emitted instructions 372 EmittedInstrs.push_front(CurrCycleInstr); 373 374 // Add a nullptr for each additional wait state after the first. Make sure 375 // not to add more than getMaxLookAhead() items to the list, since we 376 // truncate the list to that size right after this loop. 377 for (unsigned i = 1, e = std::min(NumWaitStates, getMaxLookAhead()); 378 i < e; ++i) { 379 EmittedInstrs.push_front(nullptr); 380 } 381 382 // getMaxLookahead() is the largest number of wait states we will ever need 383 // to insert, so there is no point in keeping track of more than that many 384 // wait states. 385 EmittedInstrs.resize(getMaxLookAhead()); 386 387 CurrCycleInstr = nullptr; 388 } 389 390 void GCNHazardRecognizer::RecedeCycle() { 391 llvm_unreachable("hazard recognizer does not support bottom-up scheduling."); 392 } 393 394 //===----------------------------------------------------------------------===// 395 // Helper Functions 396 //===----------------------------------------------------------------------===// 397 398 typedef function_ref<bool(const MachineInstr &, int WaitStates)> IsExpiredFn; 399 400 // Returns a minimum wait states since \p I walking all predecessors. 401 // Only scans until \p IsExpired does not return true. 402 // Can only be run in a hazard recognizer mode. 403 static int getWaitStatesSince(GCNHazardRecognizer::IsHazardFn IsHazard, 404 const MachineBasicBlock *MBB, 405 MachineBasicBlock::const_reverse_instr_iterator I, 406 int WaitStates, IsExpiredFn IsExpired, 407 DenseSet<const MachineBasicBlock *> &Visited) { 408 for (auto E = MBB->instr_rend(); I != E; ++I) { 409 // Don't add WaitStates for parent BUNDLE instructions. 410 if (I->isBundle()) 411 continue; 412 413 if (IsHazard(*I)) 414 return WaitStates; 415 416 if (I->isInlineAsm() || I->isMetaInstruction()) 417 continue; 418 419 WaitStates += SIInstrInfo::getNumWaitStates(*I); 420 421 if (IsExpired(*I, WaitStates)) 422 return std::numeric_limits<int>::max(); 423 } 424 425 int MinWaitStates = std::numeric_limits<int>::max(); 426 for (MachineBasicBlock *Pred : MBB->predecessors()) { 427 if (!Visited.insert(Pred).second) 428 continue; 429 430 int W = getWaitStatesSince(IsHazard, Pred, Pred->instr_rbegin(), 431 WaitStates, IsExpired, Visited); 432 433 MinWaitStates = std::min(MinWaitStates, W); 434 } 435 436 return MinWaitStates; 437 } 438 439 static int getWaitStatesSince(GCNHazardRecognizer::IsHazardFn IsHazard, 440 const MachineInstr *MI, IsExpiredFn IsExpired) { 441 DenseSet<const MachineBasicBlock *> Visited; 442 return getWaitStatesSince(IsHazard, MI->getParent(), 443 std::next(MI->getReverseIterator()), 444 0, IsExpired, Visited); 445 } 446 447 int GCNHazardRecognizer::getWaitStatesSince(IsHazardFn IsHazard, int Limit) { 448 if (IsHazardRecognizerMode) { 449 auto IsExpiredFn = [Limit](const MachineInstr &, int WaitStates) { 450 return WaitStates >= Limit; 451 }; 452 return ::getWaitStatesSince(IsHazard, CurrCycleInstr, IsExpiredFn); 453 } 454 455 int WaitStates = 0; 456 for (MachineInstr *MI : EmittedInstrs) { 457 if (MI) { 458 if (IsHazard(*MI)) 459 return WaitStates; 460 461 if (MI->isInlineAsm()) 462 continue; 463 } 464 ++WaitStates; 465 466 if (WaitStates >= Limit) 467 break; 468 } 469 return std::numeric_limits<int>::max(); 470 } 471 472 int GCNHazardRecognizer::getWaitStatesSinceDef(unsigned Reg, 473 IsHazardFn IsHazardDef, 474 int Limit) { 475 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 476 477 auto IsHazardFn = [IsHazardDef, TRI, Reg](const MachineInstr &MI) { 478 return IsHazardDef(MI) && MI.modifiesRegister(Reg, TRI); 479 }; 480 481 return getWaitStatesSince(IsHazardFn, Limit); 482 } 483 484 int GCNHazardRecognizer::getWaitStatesSinceSetReg(IsHazardFn IsHazard, 485 int Limit) { 486 auto IsHazardFn = [IsHazard](const MachineInstr &MI) { 487 return isSSetReg(MI.getOpcode()) && IsHazard(MI); 488 }; 489 490 return getWaitStatesSince(IsHazardFn, Limit); 491 } 492 493 //===----------------------------------------------------------------------===// 494 // No-op Hazard Detection 495 //===----------------------------------------------------------------------===// 496 497 static void addRegUnits(const SIRegisterInfo &TRI, BitVector &BV, 498 MCRegister Reg) { 499 for (MCRegUnitIterator RUI(Reg, &TRI); RUI.isValid(); ++RUI) 500 BV.set(*RUI); 501 } 502 503 static void addRegsToSet(const SIRegisterInfo &TRI, 504 iterator_range<MachineInstr::const_mop_iterator> Ops, 505 BitVector &Set) { 506 for (const MachineOperand &Op : Ops) { 507 if (Op.isReg()) 508 addRegUnits(TRI, Set, Op.getReg().asMCReg()); 509 } 510 } 511 512 void GCNHazardRecognizer::addClauseInst(const MachineInstr &MI) { 513 // XXX: Do we need to worry about implicit operands 514 addRegsToSet(TRI, MI.defs(), ClauseDefs); 515 addRegsToSet(TRI, MI.uses(), ClauseUses); 516 } 517 518 static bool breaksSMEMSoftClause(MachineInstr *MI) { 519 return !SIInstrInfo::isSMRD(*MI); 520 } 521 522 static bool breaksVMEMSoftClause(MachineInstr *MI) { 523 return !SIInstrInfo::isVMEM(*MI) && !SIInstrInfo::isFLAT(*MI); 524 } 525 526 int GCNHazardRecognizer::checkSoftClauseHazards(MachineInstr *MEM) { 527 // SMEM soft clause are only present on VI+, and only matter if xnack is 528 // enabled. 529 if (!ST.isXNACKEnabled()) 530 return 0; 531 532 bool IsSMRD = TII.isSMRD(*MEM); 533 534 resetClause(); 535 536 // A soft-clause is any group of consecutive SMEM instructions. The 537 // instructions in this group may return out of order and/or may be 538 // replayed (i.e. the same instruction issued more than once). 539 // 540 // In order to handle these situations correctly we need to make sure that 541 // when a clause has more than one instruction, no instruction in the clause 542 // writes to a register that is read by another instruction in the clause 543 // (including itself). If we encounter this situaion, we need to break the 544 // clause by inserting a non SMEM instruction. 545 546 for (MachineInstr *MI : EmittedInstrs) { 547 // When we hit a non-SMEM instruction then we have passed the start of the 548 // clause and we can stop. 549 if (!MI) 550 break; 551 552 if (IsSMRD ? breaksSMEMSoftClause(MI) : breaksVMEMSoftClause(MI)) 553 break; 554 555 addClauseInst(*MI); 556 } 557 558 if (ClauseDefs.none()) 559 return 0; 560 561 // We need to make sure not to put loads and stores in the same clause if they 562 // use the same address. For now, just start a new clause whenever we see a 563 // store. 564 if (MEM->mayStore()) 565 return 1; 566 567 addClauseInst(*MEM); 568 569 // If the set of defs and uses intersect then we cannot add this instruction 570 // to the clause, so we have a hazard. 571 return ClauseDefs.anyCommon(ClauseUses) ? 1 : 0; 572 } 573 574 int GCNHazardRecognizer::checkSMRDHazards(MachineInstr *SMRD) { 575 int WaitStatesNeeded = 0; 576 577 WaitStatesNeeded = checkSoftClauseHazards(SMRD); 578 579 // This SMRD hazard only affects SI. 580 if (!ST.hasSMRDReadVALUDefHazard()) 581 return WaitStatesNeeded; 582 583 // A read of an SGPR by SMRD instruction requires 4 wait states when the 584 // SGPR was written by a VALU instruction. 585 int SmrdSgprWaitStates = 4; 586 auto IsHazardDefFn = [this](const MachineInstr &MI) { 587 return TII.isVALU(MI); 588 }; 589 auto IsBufferHazardDefFn = [this](const MachineInstr &MI) { 590 return TII.isSALU(MI); 591 }; 592 593 bool IsBufferSMRD = TII.isBufferSMRD(*SMRD); 594 595 for (const MachineOperand &Use : SMRD->uses()) { 596 if (!Use.isReg()) 597 continue; 598 int WaitStatesNeededForUse = 599 SmrdSgprWaitStates - getWaitStatesSinceDef(Use.getReg(), IsHazardDefFn, 600 SmrdSgprWaitStates); 601 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 602 603 // This fixes what appears to be undocumented hardware behavior in SI where 604 // s_mov writing a descriptor and s_buffer_load_dword reading the descriptor 605 // needs some number of nops in between. We don't know how many we need, but 606 // let's use 4. This wasn't discovered before probably because the only 607 // case when this happens is when we expand a 64-bit pointer into a full 608 // descriptor and use s_buffer_load_dword instead of s_load_dword, which was 609 // probably never encountered in the closed-source land. 610 if (IsBufferSMRD) { 611 int WaitStatesNeededForUse = 612 SmrdSgprWaitStates - getWaitStatesSinceDef(Use.getReg(), 613 IsBufferHazardDefFn, 614 SmrdSgprWaitStates); 615 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 616 } 617 } 618 619 return WaitStatesNeeded; 620 } 621 622 int GCNHazardRecognizer::checkVMEMHazards(MachineInstr* VMEM) { 623 if (!ST.hasVMEMReadSGPRVALUDefHazard()) 624 return 0; 625 626 int WaitStatesNeeded = checkSoftClauseHazards(VMEM); 627 628 // A read of an SGPR by a VMEM instruction requires 5 wait states when the 629 // SGPR was written by a VALU Instruction. 630 const int VmemSgprWaitStates = 5; 631 auto IsHazardDefFn = [this](const MachineInstr &MI) { 632 return TII.isVALU(MI); 633 }; 634 for (const MachineOperand &Use : VMEM->uses()) { 635 if (!Use.isReg() || TRI.isVectorRegister(MF.getRegInfo(), Use.getReg())) 636 continue; 637 638 int WaitStatesNeededForUse = 639 VmemSgprWaitStates - getWaitStatesSinceDef(Use.getReg(), IsHazardDefFn, 640 VmemSgprWaitStates); 641 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 642 } 643 return WaitStatesNeeded; 644 } 645 646 int GCNHazardRecognizer::checkDPPHazards(MachineInstr *DPP) { 647 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 648 const SIInstrInfo *TII = ST.getInstrInfo(); 649 650 // Check for DPP VGPR read after VALU VGPR write and EXEC write. 651 int DppVgprWaitStates = 2; 652 int DppExecWaitStates = 5; 653 int WaitStatesNeeded = 0; 654 auto IsHazardDefFn = [TII](const MachineInstr &MI) { 655 return TII->isVALU(MI); 656 }; 657 658 for (const MachineOperand &Use : DPP->uses()) { 659 if (!Use.isReg() || !TRI->isVGPR(MF.getRegInfo(), Use.getReg())) 660 continue; 661 int WaitStatesNeededForUse = 662 DppVgprWaitStates - getWaitStatesSinceDef( 663 Use.getReg(), 664 [](const MachineInstr &) { return true; }, 665 DppVgprWaitStates); 666 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 667 } 668 669 WaitStatesNeeded = std::max( 670 WaitStatesNeeded, 671 DppExecWaitStates - getWaitStatesSinceDef(AMDGPU::EXEC, IsHazardDefFn, 672 DppExecWaitStates)); 673 674 return WaitStatesNeeded; 675 } 676 677 int GCNHazardRecognizer::checkDivFMasHazards(MachineInstr *DivFMas) { 678 const SIInstrInfo *TII = ST.getInstrInfo(); 679 680 // v_div_fmas requires 4 wait states after a write to vcc from a VALU 681 // instruction. 682 const int DivFMasWaitStates = 4; 683 auto IsHazardDefFn = [TII](const MachineInstr &MI) { 684 return TII->isVALU(MI); 685 }; 686 int WaitStatesNeeded = getWaitStatesSinceDef(AMDGPU::VCC, IsHazardDefFn, 687 DivFMasWaitStates); 688 689 return DivFMasWaitStates - WaitStatesNeeded; 690 } 691 692 int GCNHazardRecognizer::checkGetRegHazards(MachineInstr *GetRegInstr) { 693 const SIInstrInfo *TII = ST.getInstrInfo(); 694 unsigned GetRegHWReg = getHWReg(TII, *GetRegInstr); 695 696 const int GetRegWaitStates = 2; 697 auto IsHazardFn = [TII, GetRegHWReg](const MachineInstr &MI) { 698 return GetRegHWReg == getHWReg(TII, MI); 699 }; 700 int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazardFn, GetRegWaitStates); 701 702 return GetRegWaitStates - WaitStatesNeeded; 703 } 704 705 int GCNHazardRecognizer::checkSetRegHazards(MachineInstr *SetRegInstr) { 706 const SIInstrInfo *TII = ST.getInstrInfo(); 707 unsigned HWReg = getHWReg(TII, *SetRegInstr); 708 709 const int SetRegWaitStates = ST.getSetRegWaitStates(); 710 auto IsHazardFn = [TII, HWReg](const MachineInstr &MI) { 711 return HWReg == getHWReg(TII, MI); 712 }; 713 int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazardFn, SetRegWaitStates); 714 return SetRegWaitStates - WaitStatesNeeded; 715 } 716 717 int GCNHazardRecognizer::createsVALUHazard(const MachineInstr &MI) { 718 if (!MI.mayStore()) 719 return -1; 720 721 const SIInstrInfo *TII = ST.getInstrInfo(); 722 unsigned Opcode = MI.getOpcode(); 723 const MCInstrDesc &Desc = MI.getDesc(); 724 725 int VDataIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vdata); 726 int VDataRCID = -1; 727 if (VDataIdx != -1) 728 VDataRCID = Desc.OpInfo[VDataIdx].RegClass; 729 730 if (TII->isMUBUF(MI) || TII->isMTBUF(MI)) { 731 // There is no hazard if the instruction does not use vector regs 732 // (like wbinvl1) 733 if (VDataIdx == -1) 734 return -1; 735 // For MUBUF/MTBUF instructions this hazard only exists if the 736 // instruction is not using a register in the soffset field. 737 const MachineOperand *SOffset = 738 TII->getNamedOperand(MI, AMDGPU::OpName::soffset); 739 // If we have no soffset operand, then assume this field has been 740 // hardcoded to zero. 741 if (AMDGPU::getRegBitWidth(VDataRCID) > 64 && 742 (!SOffset || !SOffset->isReg())) 743 return VDataIdx; 744 } 745 746 // MIMG instructions create a hazard if they don't use a 256-bit T# and 747 // the store size is greater than 8 bytes and they have more than two bits 748 // of their dmask set. 749 // All our MIMG definitions use a 256-bit T#, so we can skip checking for them. 750 if (TII->isMIMG(MI)) { 751 int SRsrcIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::srsrc); 752 assert(SRsrcIdx != -1 && 753 AMDGPU::getRegBitWidth(Desc.OpInfo[SRsrcIdx].RegClass) == 256); 754 (void)SRsrcIdx; 755 } 756 757 if (TII->isFLAT(MI)) { 758 int DataIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vdata); 759 if (AMDGPU::getRegBitWidth(Desc.OpInfo[DataIdx].RegClass) > 64) 760 return DataIdx; 761 } 762 763 return -1; 764 } 765 766 int 767 GCNHazardRecognizer::checkVALUHazardsHelper(const MachineOperand &Def, 768 const MachineRegisterInfo &MRI) { 769 // Helper to check for the hazard where VMEM instructions that store more than 770 // 8 bytes can have there store data over written by the next instruction. 771 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 772 773 const int VALUWaitStates = 1; 774 int WaitStatesNeeded = 0; 775 776 if (!TRI->isVectorRegister(MRI, Def.getReg())) 777 return WaitStatesNeeded; 778 Register Reg = Def.getReg(); 779 auto IsHazardFn = [this, Reg, TRI](const MachineInstr &MI) { 780 int DataIdx = createsVALUHazard(MI); 781 return DataIdx >= 0 && 782 TRI->regsOverlap(MI.getOperand(DataIdx).getReg(), Reg); 783 }; 784 int WaitStatesNeededForDef = 785 VALUWaitStates - getWaitStatesSince(IsHazardFn, VALUWaitStates); 786 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForDef); 787 788 return WaitStatesNeeded; 789 } 790 791 int GCNHazardRecognizer::checkVALUHazards(MachineInstr *VALU) { 792 // This checks for the hazard where VMEM instructions that store more than 793 // 8 bytes can have there store data over written by the next instruction. 794 if (!ST.has12DWordStoreHazard()) 795 return 0; 796 797 const MachineRegisterInfo &MRI = MF.getRegInfo(); 798 int WaitStatesNeeded = 0; 799 800 for (const MachineOperand &Def : VALU->defs()) { 801 WaitStatesNeeded = std::max(WaitStatesNeeded, checkVALUHazardsHelper(Def, MRI)); 802 } 803 804 return WaitStatesNeeded; 805 } 806 807 int GCNHazardRecognizer::checkInlineAsmHazards(MachineInstr *IA) { 808 // This checks for hazards associated with inline asm statements. 809 // Since inline asms can contain just about anything, we use this 810 // to call/leverage other check*Hazard routines. Note that 811 // this function doesn't attempt to address all possible inline asm 812 // hazards (good luck), but is a collection of what has been 813 // problematic thus far. 814 815 // see checkVALUHazards() 816 if (!ST.has12DWordStoreHazard()) 817 return 0; 818 819 const MachineRegisterInfo &MRI = MF.getRegInfo(); 820 int WaitStatesNeeded = 0; 821 822 for (unsigned I = InlineAsm::MIOp_FirstOperand, E = IA->getNumOperands(); 823 I != E; ++I) { 824 const MachineOperand &Op = IA->getOperand(I); 825 if (Op.isReg() && Op.isDef()) { 826 WaitStatesNeeded = std::max(WaitStatesNeeded, checkVALUHazardsHelper(Op, MRI)); 827 } 828 } 829 830 return WaitStatesNeeded; 831 } 832 833 int GCNHazardRecognizer::checkRWLaneHazards(MachineInstr *RWLane) { 834 const SIInstrInfo *TII = ST.getInstrInfo(); 835 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 836 const MachineRegisterInfo &MRI = MF.getRegInfo(); 837 838 const MachineOperand *LaneSelectOp = 839 TII->getNamedOperand(*RWLane, AMDGPU::OpName::src1); 840 841 if (!LaneSelectOp->isReg() || !TRI->isSGPRReg(MRI, LaneSelectOp->getReg())) 842 return 0; 843 844 Register LaneSelectReg = LaneSelectOp->getReg(); 845 auto IsHazardFn = [TII](const MachineInstr &MI) { return TII->isVALU(MI); }; 846 847 const int RWLaneWaitStates = 4; 848 int WaitStatesSince = getWaitStatesSinceDef(LaneSelectReg, IsHazardFn, 849 RWLaneWaitStates); 850 return RWLaneWaitStates - WaitStatesSince; 851 } 852 853 int GCNHazardRecognizer::checkRFEHazards(MachineInstr *RFE) { 854 if (!ST.hasRFEHazards()) 855 return 0; 856 857 const SIInstrInfo *TII = ST.getInstrInfo(); 858 859 const int RFEWaitStates = 1; 860 861 auto IsHazardFn = [TII](const MachineInstr &MI) { 862 return getHWReg(TII, MI) == AMDGPU::Hwreg::ID_TRAPSTS; 863 }; 864 int WaitStatesNeeded = getWaitStatesSinceSetReg(IsHazardFn, RFEWaitStates); 865 return RFEWaitStates - WaitStatesNeeded; 866 } 867 868 int GCNHazardRecognizer::checkReadM0Hazards(MachineInstr *MI) { 869 const SIInstrInfo *TII = ST.getInstrInfo(); 870 const int SMovRelWaitStates = 1; 871 auto IsHazardFn = [TII](const MachineInstr &MI) { return TII->isSALU(MI); }; 872 return SMovRelWaitStates - getWaitStatesSinceDef(AMDGPU::M0, IsHazardFn, 873 SMovRelWaitStates); 874 } 875 876 void GCNHazardRecognizer::fixHazards(MachineInstr *MI) { 877 fixVMEMtoScalarWriteHazards(MI); 878 fixVcmpxPermlaneHazards(MI); 879 fixSMEMtoVectorWriteHazards(MI); 880 fixVcmpxExecWARHazard(MI); 881 fixLdsBranchVmemWARHazard(MI); 882 } 883 884 bool GCNHazardRecognizer::fixVcmpxPermlaneHazards(MachineInstr *MI) { 885 if (!ST.hasVcmpxPermlaneHazard() || !isPermlane(*MI)) 886 return false; 887 888 const SIInstrInfo *TII = ST.getInstrInfo(); 889 auto IsHazardFn = [TII](const MachineInstr &MI) { return TII->isVOPC(MI); }; 890 891 auto IsExpiredFn = [](const MachineInstr &MI, int) { 892 unsigned Opc = MI.getOpcode(); 893 return SIInstrInfo::isVALU(MI) && Opc != AMDGPU::V_NOP_e32 && 894 Opc != AMDGPU::V_NOP_e64 && Opc != AMDGPU::V_NOP_sdwa; 895 }; 896 897 if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) == 898 std::numeric_limits<int>::max()) 899 return false; 900 901 // V_NOP will be discarded by SQ. 902 // Use V_MOB_B32 v?, v?. Register must be alive so use src0 of V_PERMLANE* 903 // which is always a VGPR and available. 904 auto *Src0 = TII->getNamedOperand(*MI, AMDGPU::OpName::src0); 905 Register Reg = Src0->getReg(); 906 bool IsUndef = Src0->isUndef(); 907 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), 908 TII->get(AMDGPU::V_MOV_B32_e32)) 909 .addReg(Reg, RegState::Define | (IsUndef ? RegState::Dead : 0)) 910 .addReg(Reg, IsUndef ? RegState::Undef : RegState::Kill); 911 912 return true; 913 } 914 915 bool GCNHazardRecognizer::fixVMEMtoScalarWriteHazards(MachineInstr *MI) { 916 if (!ST.hasVMEMtoScalarWriteHazard()) 917 return false; 918 919 if (!SIInstrInfo::isSALU(*MI) && !SIInstrInfo::isSMRD(*MI)) 920 return false; 921 922 if (MI->getNumDefs() == 0) 923 return false; 924 925 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 926 927 auto IsHazardFn = [TRI, MI](const MachineInstr &I) { 928 if (!SIInstrInfo::isVMEM(I) && !SIInstrInfo::isDS(I) && 929 !SIInstrInfo::isFLAT(I)) 930 return false; 931 932 for (const MachineOperand &Def : MI->defs()) { 933 const MachineOperand *Op = 934 I.findRegisterUseOperand(Def.getReg(), false, TRI); 935 if (!Op) 936 continue; 937 return true; 938 } 939 return false; 940 }; 941 942 auto IsExpiredFn = [](const MachineInstr &MI, int) { 943 return SIInstrInfo::isVALU(MI) || 944 (MI.getOpcode() == AMDGPU::S_WAITCNT && 945 !MI.getOperand(0).getImm()) || 946 (MI.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR && 947 MI.getOperand(0).getImm() == 0xffe3); 948 }; 949 950 if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) == 951 std::numeric_limits<int>::max()) 952 return false; 953 954 const SIInstrInfo *TII = ST.getInstrInfo(); 955 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), 956 TII->get(AMDGPU::S_WAITCNT_DEPCTR)) 957 .addImm(0xffe3); 958 return true; 959 } 960 961 bool GCNHazardRecognizer::fixSMEMtoVectorWriteHazards(MachineInstr *MI) { 962 if (!ST.hasSMEMtoVectorWriteHazard()) 963 return false; 964 965 if (!SIInstrInfo::isVALU(*MI)) 966 return false; 967 968 unsigned SDSTName; 969 switch (MI->getOpcode()) { 970 case AMDGPU::V_READLANE_B32: 971 case AMDGPU::V_READFIRSTLANE_B32: 972 SDSTName = AMDGPU::OpName::vdst; 973 break; 974 default: 975 SDSTName = AMDGPU::OpName::sdst; 976 break; 977 } 978 979 const SIInstrInfo *TII = ST.getInstrInfo(); 980 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 981 const AMDGPU::IsaVersion IV = AMDGPU::getIsaVersion(ST.getCPU()); 982 const MachineOperand *SDST = TII->getNamedOperand(*MI, SDSTName); 983 if (!SDST) { 984 for (const auto &MO : MI->implicit_operands()) { 985 if (MO.isDef() && TRI->isSGPRClass(TRI->getPhysRegClass(MO.getReg()))) { 986 SDST = &MO; 987 break; 988 } 989 } 990 } 991 992 if (!SDST) 993 return false; 994 995 const Register SDSTReg = SDST->getReg(); 996 auto IsHazardFn = [SDSTReg, TRI](const MachineInstr &I) { 997 return SIInstrInfo::isSMRD(I) && I.readsRegister(SDSTReg, TRI); 998 }; 999 1000 auto IsExpiredFn = [TII, IV](const MachineInstr &MI, int) { 1001 if (TII->isSALU(MI)) { 1002 switch (MI.getOpcode()) { 1003 case AMDGPU::S_SETVSKIP: 1004 case AMDGPU::S_VERSION: 1005 case AMDGPU::S_WAITCNT_VSCNT: 1006 case AMDGPU::S_WAITCNT_VMCNT: 1007 case AMDGPU::S_WAITCNT_EXPCNT: 1008 // These instructions cannot not mitigate the hazard. 1009 return false; 1010 case AMDGPU::S_WAITCNT_LGKMCNT: 1011 // Reducing lgkmcnt count to 0 always mitigates the hazard. 1012 return (MI.getOperand(1).getImm() == 0) && 1013 (MI.getOperand(0).getReg() == AMDGPU::SGPR_NULL); 1014 case AMDGPU::S_WAITCNT: { 1015 const int64_t Imm = MI.getOperand(0).getImm(); 1016 AMDGPU::Waitcnt Decoded = AMDGPU::decodeWaitcnt(IV, Imm); 1017 return (Decoded.LgkmCnt == 0); 1018 } 1019 default: 1020 // SOPP instructions cannot mitigate the hazard. 1021 if (TII->isSOPP(MI)) 1022 return false; 1023 // At this point the SALU can be assumed to mitigate the hazard 1024 // because either: 1025 // (a) it is independent of the at risk SMEM (breaking chain), 1026 // or 1027 // (b) it is dependent on the SMEM, in which case an appropriate 1028 // s_waitcnt lgkmcnt _must_ exist between it and the at risk 1029 // SMEM instruction. 1030 return true; 1031 } 1032 } 1033 return false; 1034 }; 1035 1036 if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) == 1037 std::numeric_limits<int>::max()) 1038 return false; 1039 1040 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), 1041 TII->get(AMDGPU::S_MOV_B32), AMDGPU::SGPR_NULL) 1042 .addImm(0); 1043 return true; 1044 } 1045 1046 bool GCNHazardRecognizer::fixVcmpxExecWARHazard(MachineInstr *MI) { 1047 if (!ST.hasVcmpxExecWARHazard() || !SIInstrInfo::isVALU(*MI)) 1048 return false; 1049 1050 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 1051 if (!MI->modifiesRegister(AMDGPU::EXEC, TRI)) 1052 return false; 1053 1054 auto IsHazardFn = [TRI](const MachineInstr &I) { 1055 if (SIInstrInfo::isVALU(I)) 1056 return false; 1057 return I.readsRegister(AMDGPU::EXEC, TRI); 1058 }; 1059 1060 const SIInstrInfo *TII = ST.getInstrInfo(); 1061 auto IsExpiredFn = [TII, TRI](const MachineInstr &MI, int) { 1062 if (SIInstrInfo::isVALU(MI)) { 1063 if (TII->getNamedOperand(MI, AMDGPU::OpName::sdst)) 1064 return true; 1065 for (auto MO : MI.implicit_operands()) 1066 if (MO.isDef() && TRI->isSGPRClass(TRI->getPhysRegClass(MO.getReg()))) 1067 return true; 1068 } 1069 if (MI.getOpcode() == AMDGPU::S_WAITCNT_DEPCTR && 1070 (MI.getOperand(0).getImm() & 0xfffe) == 0xfffe) 1071 return true; 1072 return false; 1073 }; 1074 1075 if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) == 1076 std::numeric_limits<int>::max()) 1077 return false; 1078 1079 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), 1080 TII->get(AMDGPU::S_WAITCNT_DEPCTR)) 1081 .addImm(0xfffe); 1082 return true; 1083 } 1084 1085 static bool shouldRunLdsBranchVmemWARHazardFixup(const MachineFunction &MF, 1086 const GCNSubtarget &ST) { 1087 if (!ST.hasLdsBranchVmemWARHazard()) 1088 return false; 1089 1090 // Check if the necessary condition for the hazard is met: both LDS and VMEM 1091 // instructions need to appear in the same function. 1092 bool HasLds = false; 1093 bool HasVmem = false; 1094 for (auto &MBB : MF) { 1095 for (auto &MI : MBB) { 1096 HasLds |= SIInstrInfo::isDS(MI); 1097 HasVmem |= 1098 SIInstrInfo::isVMEM(MI) || SIInstrInfo::isSegmentSpecificFLAT(MI); 1099 if (HasLds && HasVmem) 1100 return true; 1101 } 1102 } 1103 return false; 1104 } 1105 1106 bool GCNHazardRecognizer::fixLdsBranchVmemWARHazard(MachineInstr *MI) { 1107 if (!RunLdsBranchVmemWARHazardFixup) 1108 return false; 1109 1110 assert(ST.hasLdsBranchVmemWARHazard()); 1111 1112 auto IsHazardInst = [](const MachineInstr &MI) { 1113 if (SIInstrInfo::isDS(MI)) 1114 return 1; 1115 if (SIInstrInfo::isVMEM(MI) || SIInstrInfo::isSegmentSpecificFLAT(MI)) 1116 return 2; 1117 return 0; 1118 }; 1119 1120 auto InstType = IsHazardInst(*MI); 1121 if (!InstType) 1122 return false; 1123 1124 auto IsExpiredFn = [&IsHazardInst](const MachineInstr &I, int) { 1125 return IsHazardInst(I) || (I.getOpcode() == AMDGPU::S_WAITCNT_VSCNT && 1126 I.getOperand(0).getReg() == AMDGPU::SGPR_NULL && 1127 !I.getOperand(1).getImm()); 1128 }; 1129 1130 auto IsHazardFn = [InstType, &IsHazardInst](const MachineInstr &I) { 1131 if (!I.isBranch()) 1132 return false; 1133 1134 auto IsHazardFn = [InstType, IsHazardInst](const MachineInstr &I) { 1135 auto InstType2 = IsHazardInst(I); 1136 return InstType2 && InstType != InstType2; 1137 }; 1138 1139 auto IsExpiredFn = [InstType, &IsHazardInst](const MachineInstr &I, int) { 1140 auto InstType2 = IsHazardInst(I); 1141 if (InstType == InstType2) 1142 return true; 1143 1144 return I.getOpcode() == AMDGPU::S_WAITCNT_VSCNT && 1145 I.getOperand(0).getReg() == AMDGPU::SGPR_NULL && 1146 !I.getOperand(1).getImm(); 1147 }; 1148 1149 return ::getWaitStatesSince(IsHazardFn, &I, IsExpiredFn) != 1150 std::numeric_limits<int>::max(); 1151 }; 1152 1153 if (::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn) == 1154 std::numeric_limits<int>::max()) 1155 return false; 1156 1157 const SIInstrInfo *TII = ST.getInstrInfo(); 1158 BuildMI(*MI->getParent(), MI, MI->getDebugLoc(), 1159 TII->get(AMDGPU::S_WAITCNT_VSCNT)) 1160 .addReg(AMDGPU::SGPR_NULL, RegState::Undef) 1161 .addImm(0); 1162 1163 return true; 1164 } 1165 1166 int GCNHazardRecognizer::checkNSAtoVMEMHazard(MachineInstr *MI) { 1167 int NSAtoVMEMWaitStates = 1; 1168 1169 if (!ST.hasNSAtoVMEMBug()) 1170 return 0; 1171 1172 if (!SIInstrInfo::isMUBUF(*MI) && !SIInstrInfo::isMTBUF(*MI)) 1173 return 0; 1174 1175 const SIInstrInfo *TII = ST.getInstrInfo(); 1176 const auto *Offset = TII->getNamedOperand(*MI, AMDGPU::OpName::offset); 1177 if (!Offset || (Offset->getImm() & 6) == 0) 1178 return 0; 1179 1180 auto IsHazardFn = [TII](const MachineInstr &I) { 1181 if (!SIInstrInfo::isMIMG(I)) 1182 return false; 1183 const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(I.getOpcode()); 1184 return Info->MIMGEncoding == AMDGPU::MIMGEncGfx10NSA && 1185 TII->getInstSizeInBytes(I) >= 16; 1186 }; 1187 1188 return NSAtoVMEMWaitStates - getWaitStatesSince(IsHazardFn, 1); 1189 } 1190 1191 int GCNHazardRecognizer::checkFPAtomicToDenormModeHazard(MachineInstr *MI) { 1192 int FPAtomicToDenormModeWaitStates = 3; 1193 1194 if (MI->getOpcode() != AMDGPU::S_DENORM_MODE) 1195 return 0; 1196 1197 auto IsHazardFn = [](const MachineInstr &I) { 1198 if (!SIInstrInfo::isVMEM(I) && !SIInstrInfo::isFLAT(I)) 1199 return false; 1200 return SIInstrInfo::isFPAtomic(I); 1201 }; 1202 1203 auto IsExpiredFn = [](const MachineInstr &MI, int WaitStates) { 1204 if (WaitStates >= 3 || SIInstrInfo::isVALU(MI)) 1205 return true; 1206 1207 switch (MI.getOpcode()) { 1208 case AMDGPU::S_WAITCNT: 1209 case AMDGPU::S_WAITCNT_VSCNT: 1210 case AMDGPU::S_WAITCNT_VMCNT: 1211 case AMDGPU::S_WAITCNT_EXPCNT: 1212 case AMDGPU::S_WAITCNT_LGKMCNT: 1213 case AMDGPU::S_WAIT_IDLE: 1214 return true; 1215 default: 1216 break; 1217 } 1218 1219 return false; 1220 }; 1221 1222 return FPAtomicToDenormModeWaitStates - 1223 ::getWaitStatesSince(IsHazardFn, MI, IsExpiredFn); 1224 } 1225 1226 int GCNHazardRecognizer::checkMAIHazards(MachineInstr *MI) { 1227 assert(SIInstrInfo::isMAI(*MI)); 1228 1229 return ST.hasGFX90AInsts() ? checkMAIHazards90A(MI) : checkMAIHazards908(MI); 1230 } 1231 1232 int GCNHazardRecognizer::checkMAIHazards908(MachineInstr *MI) { 1233 int WaitStatesNeeded = 0; 1234 unsigned Opc = MI->getOpcode(); 1235 1236 auto IsVALUFn = [](const MachineInstr &MI) { 1237 return SIInstrInfo::isVALU(MI); 1238 }; 1239 1240 if (Opc != AMDGPU::V_ACCVGPR_READ_B32_e64) { // MFMA or v_accvgpr_write 1241 const int LegacyVALUWritesVGPRWaitStates = 2; 1242 const int VALUWritesExecWaitStates = 4; 1243 const int MaxWaitStates = 4; 1244 1245 int WaitStatesNeededForUse = VALUWritesExecWaitStates - 1246 getWaitStatesSinceDef(AMDGPU::EXEC, IsVALUFn, MaxWaitStates); 1247 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1248 1249 if (WaitStatesNeeded < MaxWaitStates) { 1250 for (const MachineOperand &Use : MI->explicit_uses()) { 1251 const int MaxWaitStates = 2; 1252 1253 if (!Use.isReg() || !TRI.isVGPR(MF.getRegInfo(), Use.getReg())) 1254 continue; 1255 1256 int WaitStatesNeededForUse = LegacyVALUWritesVGPRWaitStates - 1257 getWaitStatesSinceDef(Use.getReg(), IsVALUFn, MaxWaitStates); 1258 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1259 1260 if (WaitStatesNeeded == MaxWaitStates) 1261 break; 1262 } 1263 } 1264 } 1265 1266 auto IsMFMAFn = [](const MachineInstr &MI) { 1267 return SIInstrInfo::isMAI(MI) && 1268 MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 && 1269 MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64; 1270 }; 1271 1272 for (const MachineOperand &Op : MI->explicit_operands()) { 1273 if (!Op.isReg() || !TRI.isAGPR(MF.getRegInfo(), Op.getReg())) 1274 continue; 1275 1276 if (Op.isDef() && Opc != AMDGPU::V_ACCVGPR_WRITE_B32_e64) 1277 continue; 1278 1279 const int MFMAWritesAGPROverlappedSrcABWaitStates = 4; 1280 const int MFMAWritesAGPROverlappedSrcCWaitStates = 2; 1281 const int MFMA4x4WritesAGPRAccVgprReadWaitStates = 4; 1282 const int MFMA16x16WritesAGPRAccVgprReadWaitStates = 10; 1283 const int MFMA32x32WritesAGPRAccVgprReadWaitStates = 18; 1284 const int MFMA4x4WritesAGPRAccVgprWriteWaitStates = 1; 1285 const int MFMA16x16WritesAGPRAccVgprWriteWaitStates = 7; 1286 const int MFMA32x32WritesAGPRAccVgprWriteWaitStates = 15; 1287 const int MaxWaitStates = 18; 1288 Register Reg = Op.getReg(); 1289 unsigned HazardDefLatency = 0; 1290 1291 auto IsOverlappedMFMAFn = [Reg, &IsMFMAFn, &HazardDefLatency, 1292 this](const MachineInstr &MI) { 1293 if (!IsMFMAFn(MI)) 1294 return false; 1295 Register DstReg = MI.getOperand(0).getReg(); 1296 if (DstReg == Reg) 1297 return false; 1298 HazardDefLatency = 1299 std::max(HazardDefLatency, TSchedModel.computeInstrLatency(&MI)); 1300 return TRI.regsOverlap(DstReg, Reg); 1301 }; 1302 1303 int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsOverlappedMFMAFn, 1304 MaxWaitStates); 1305 int NeedWaitStates = MFMAWritesAGPROverlappedSrcABWaitStates; 1306 int SrcCIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2); 1307 int OpNo = MI->getOperandNo(&Op); 1308 if (OpNo == SrcCIdx) { 1309 NeedWaitStates = MFMAWritesAGPROverlappedSrcCWaitStates; 1310 } else if (Opc == AMDGPU::V_ACCVGPR_READ_B32_e64) { 1311 switch (HazardDefLatency) { 1312 case 2: NeedWaitStates = MFMA4x4WritesAGPRAccVgprReadWaitStates; 1313 break; 1314 case 8: NeedWaitStates = MFMA16x16WritesAGPRAccVgprReadWaitStates; 1315 break; 1316 case 16: LLVM_FALLTHROUGH; 1317 default: NeedWaitStates = MFMA32x32WritesAGPRAccVgprReadWaitStates; 1318 break; 1319 } 1320 } else if (Opc == AMDGPU::V_ACCVGPR_WRITE_B32_e64) { 1321 switch (HazardDefLatency) { 1322 case 2: NeedWaitStates = MFMA4x4WritesAGPRAccVgprWriteWaitStates; 1323 break; 1324 case 8: NeedWaitStates = MFMA16x16WritesAGPRAccVgprWriteWaitStates; 1325 break; 1326 case 16: LLVM_FALLTHROUGH; 1327 default: NeedWaitStates = MFMA32x32WritesAGPRAccVgprWriteWaitStates; 1328 break; 1329 } 1330 } 1331 1332 int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef; 1333 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1334 1335 if (WaitStatesNeeded == MaxWaitStates) 1336 return WaitStatesNeeded; // Early exit. 1337 1338 auto IsAccVgprWriteFn = [Reg, this](const MachineInstr &MI) { 1339 if (MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64) 1340 return false; 1341 Register DstReg = MI.getOperand(0).getReg(); 1342 return TRI.regsOverlap(Reg, DstReg); 1343 }; 1344 1345 const int AccVGPRWriteMFMAReadSrcCWaitStates = 1; 1346 const int AccVGPRWriteMFMAReadSrcABWaitStates = 3; 1347 const int AccVGPRWriteAccVgprReadWaitStates = 3; 1348 NeedWaitStates = AccVGPRWriteMFMAReadSrcABWaitStates; 1349 if (OpNo == SrcCIdx) 1350 NeedWaitStates = AccVGPRWriteMFMAReadSrcCWaitStates; 1351 else if (Opc == AMDGPU::V_ACCVGPR_READ_B32_e64) 1352 NeedWaitStates = AccVGPRWriteAccVgprReadWaitStates; 1353 1354 WaitStatesNeededForUse = NeedWaitStates - 1355 getWaitStatesSinceDef(Reg, IsAccVgprWriteFn, MaxWaitStates); 1356 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1357 1358 if (WaitStatesNeeded == MaxWaitStates) 1359 return WaitStatesNeeded; // Early exit. 1360 } 1361 1362 if (Opc == AMDGPU::V_ACCVGPR_WRITE_B32_e64) { 1363 const int MFMA4x4ReadSrcCAccVgprWriteWaitStates = 0; 1364 const int MFMA16x16ReadSrcCAccVgprWriteWaitStates = 5; 1365 const int MFMA32x32ReadSrcCAccVgprWriteWaitStates = 13; 1366 const int MaxWaitStates = 13; 1367 Register DstReg = MI->getOperand(0).getReg(); 1368 unsigned HazardDefLatency = 0; 1369 1370 auto IsSrcCMFMAFn = [DstReg, &IsMFMAFn, &HazardDefLatency, 1371 this](const MachineInstr &MI) { 1372 if (!IsMFMAFn(MI)) 1373 return false; 1374 Register Reg = TII.getNamedOperand(MI, AMDGPU::OpName::src2)->getReg(); 1375 HazardDefLatency = 1376 std::max(HazardDefLatency, TSchedModel.computeInstrLatency(&MI)); 1377 return TRI.regsOverlap(Reg, DstReg); 1378 }; 1379 1380 int WaitStatesSince = getWaitStatesSince(IsSrcCMFMAFn, MaxWaitStates); 1381 int NeedWaitStates; 1382 switch (HazardDefLatency) { 1383 case 2: NeedWaitStates = MFMA4x4ReadSrcCAccVgprWriteWaitStates; 1384 break; 1385 case 8: NeedWaitStates = MFMA16x16ReadSrcCAccVgprWriteWaitStates; 1386 break; 1387 case 16: LLVM_FALLTHROUGH; 1388 default: NeedWaitStates = MFMA32x32ReadSrcCAccVgprWriteWaitStates; 1389 break; 1390 } 1391 1392 int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSince; 1393 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1394 } 1395 1396 return WaitStatesNeeded; 1397 } 1398 1399 int GCNHazardRecognizer::checkMAIHazards90A(MachineInstr *MI) { 1400 int WaitStatesNeeded = 0; 1401 unsigned Opc = MI->getOpcode(); 1402 1403 auto IsMFMAFn = [](const MachineInstr &MI) { 1404 return SIInstrInfo::isMAI(MI) && 1405 MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 && 1406 MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64; 1407 }; 1408 1409 auto IsLegacyVALUFn = [&IsMFMAFn](const MachineInstr &MI) { 1410 return SIInstrInfo::isVALU(MI) && !IsMFMAFn(MI); 1411 }; 1412 1413 auto IsLegacyVALUNotDotFn = [&IsMFMAFn](const MachineInstr &MI) { 1414 return SIInstrInfo::isVALU(MI) && !IsMFMAFn(MI) && !SIInstrInfo::isDOT(MI); 1415 }; 1416 1417 if (!IsMFMAFn(*MI)) 1418 return WaitStatesNeeded; 1419 1420 const int VALUWritesExecWaitStates = 4; 1421 int WaitStatesNeededForUse = VALUWritesExecWaitStates - 1422 getWaitStatesSinceDef(AMDGPU::EXEC, IsLegacyVALUFn, 1423 VALUWritesExecWaitStates); 1424 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1425 1426 int SrcCIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2); 1427 1428 // Loop for both DGEMM and S/HGEMM 2nd instruction. 1429 for (const MachineOperand &Use : MI->explicit_uses()) { 1430 const int LegacyVALUNotDotWritesVGPRWaitStates = 2; 1431 const int SMFMA4x4WritesVGPROverlappedSMFMASrcCWaitStates = 2; 1432 const int SMFMA16x16WritesVGPROverlappedSMFMASrcCWaitStates = 8; 1433 const int SMFMA32x32WritesVGPROverlappedSMFMASrcCWaitStates = 16; 1434 const int SMFMA4x4WritesVGPROverlappedDMFMASrcCWaitStates = 3; 1435 const int SMFMA16x16WritesVGPROverlappedDMFMASrcCWaitStates = 9; 1436 const int SMFMA32x32WritesVGPROverlappedDMFMASrcCWaitStates = 17; 1437 const int DMFMA16x16WritesVGPROverlappedSrcCWaitStates = 9; 1438 const int DMFMA4x4WritesVGPROverlappedSrcCWaitStates = 4; 1439 const int SMFMA4x4WritesVGPROverlappedSrcABWaitStates = 5; 1440 const int SMFMA16x16WritesVGPROverlappedSrcABWaitStates = 11; 1441 const int SMFMA32x32WritesVGPROverlappedSrcABWaitStates = 19; 1442 const int DMFMA4x4WritesVGPROverlappedMFMASrcABWaitStates = 6; 1443 const int DMFMA16x16WritesVGPROverlappedMFMASrcABWaitStates = 11; 1444 const int DMFMA4x4WritesVGPRFullSrcCWaitStates = 4; 1445 const int MaxWaitStates = 19; 1446 1447 if (!Use.isReg()) 1448 continue; 1449 unsigned Reg = Use.getReg(); 1450 bool FullReg; 1451 const MachineInstr *MI1; 1452 1453 auto IsOverlappedDGEMMorXDLFn = [Reg, &IsMFMAFn, &FullReg, &MI1, 1454 this](const MachineInstr &MI) { 1455 if (!IsMFMAFn(MI)) 1456 return false; 1457 if (!isDGEMM(MI.getOpcode()) && !isXDL(ST, MI)) 1458 return false; 1459 Register DstReg = MI.getOperand(0).getReg(); 1460 FullReg = (DstReg == Reg); 1461 MI1 = &MI; 1462 return TRI.regsOverlap(DstReg, Reg); 1463 }; 1464 1465 WaitStatesNeededForUse = LegacyVALUNotDotWritesVGPRWaitStates - 1466 getWaitStatesSinceDef(Reg, IsLegacyVALUNotDotFn, MaxWaitStates); 1467 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1468 1469 int NumWaitStates = getWaitStatesSinceDef(Reg, IsOverlappedDGEMMorXDLFn, 1470 MaxWaitStates); 1471 if (NumWaitStates == std::numeric_limits<int>::max()) 1472 continue; 1473 1474 int OpNo = MI->getOperandNo(&Use); 1475 unsigned Opc1 = MI1->getOpcode(); 1476 int NeedWaitStates = 0; 1477 if (OpNo == SrcCIdx) { 1478 if (!isDGEMM(Opc) && isDGEMM(Opc1)) { 1479 NeedWaitStates = 0; 1480 } else if (FullReg) { 1481 if ((Opc == AMDGPU::V_MFMA_F64_4X4X4F64_e64 || 1482 Opc == AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64) && 1483 (Opc1 == AMDGPU::V_MFMA_F64_4X4X4F64_e64 || 1484 Opc1 == AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64)) 1485 NeedWaitStates = DMFMA4x4WritesVGPRFullSrcCWaitStates; 1486 } else { 1487 switch (Opc1) { 1488 case AMDGPU::V_MFMA_F64_16X16X4F64_e64: 1489 case AMDGPU::V_MFMA_F64_16X16X4F64_vgprcd_e64: 1490 if (!isXDL(ST, *MI)) 1491 NeedWaitStates = DMFMA16x16WritesVGPROverlappedSrcCWaitStates; 1492 break; 1493 case AMDGPU::V_MFMA_F64_4X4X4F64_e64: 1494 case AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64: 1495 if (!isXDL(ST, *MI)) 1496 NeedWaitStates = DMFMA4x4WritesVGPROverlappedSrcCWaitStates; 1497 break; 1498 default: 1499 switch (TSchedModel.computeInstrLatency(MI1)) { 1500 case 2: 1501 NeedWaitStates = isDGEMM(Opc) 1502 ? SMFMA4x4WritesVGPROverlappedDMFMASrcCWaitStates 1503 : SMFMA4x4WritesVGPROverlappedSMFMASrcCWaitStates; 1504 break; 1505 case 8: 1506 NeedWaitStates = isDGEMM(Opc) 1507 ? SMFMA16x16WritesVGPROverlappedDMFMASrcCWaitStates 1508 : SMFMA16x16WritesVGPROverlappedSMFMASrcCWaitStates; 1509 break; 1510 case 16: LLVM_FALLTHROUGH; 1511 default: 1512 NeedWaitStates = isDGEMM(Opc) 1513 ? SMFMA32x32WritesVGPROverlappedDMFMASrcCWaitStates 1514 : SMFMA32x32WritesVGPROverlappedSMFMASrcCWaitStates; 1515 } 1516 } 1517 } 1518 } else { 1519 switch (Opc1) { 1520 case AMDGPU::V_MFMA_F64_16X16X4F64_e64: 1521 case AMDGPU::V_MFMA_F64_16X16X4F64_vgprcd_e64: 1522 NeedWaitStates = DMFMA16x16WritesVGPROverlappedMFMASrcABWaitStates; 1523 break; 1524 case AMDGPU::V_MFMA_F64_4X4X4F64_e64: 1525 case AMDGPU::V_MFMA_F64_4X4X4F64_vgprcd_e64: 1526 NeedWaitStates = DMFMA4x4WritesVGPROverlappedMFMASrcABWaitStates; 1527 break; 1528 default: 1529 switch (TSchedModel.computeInstrLatency(MI1)) { 1530 case 2: 1531 NeedWaitStates = SMFMA4x4WritesVGPROverlappedSrcABWaitStates; 1532 break; 1533 case 8: 1534 NeedWaitStates = SMFMA16x16WritesVGPROverlappedSrcABWaitStates; 1535 break; 1536 case 16: LLVM_FALLTHROUGH; 1537 default: 1538 NeedWaitStates = SMFMA32x32WritesVGPROverlappedSrcABWaitStates; 1539 } 1540 } 1541 } 1542 if (WaitStatesNeeded >= NeedWaitStates) 1543 continue; 1544 1545 WaitStatesNeededForUse = NeedWaitStates - NumWaitStates; 1546 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1547 1548 if (WaitStatesNeeded == MaxWaitStates) 1549 break; 1550 } 1551 1552 return WaitStatesNeeded; 1553 } 1554 1555 int GCNHazardRecognizer::checkMAILdStHazards(MachineInstr *MI) { 1556 // On gfx90a+ releveant hazards are checked in checkMAIVALUHazards() 1557 if (!ST.hasMAIInsts() || ST.hasGFX90AInsts()) 1558 return 0; 1559 1560 int WaitStatesNeeded = 0; 1561 1562 auto IsAccVgprReadFn = [](const MachineInstr &MI) { 1563 return MI.getOpcode() == AMDGPU::V_ACCVGPR_READ_B32_e64; 1564 }; 1565 1566 for (const MachineOperand &Op : MI->explicit_uses()) { 1567 if (!Op.isReg() || !TRI.isVGPR(MF.getRegInfo(), Op.getReg())) 1568 continue; 1569 1570 Register Reg = Op.getReg(); 1571 1572 const int AccVgprReadLdStWaitStates = 2; 1573 const int VALUWriteAccVgprRdWrLdStDepVALUWaitStates = 1; 1574 const int MaxWaitStates = 2; 1575 1576 int WaitStatesNeededForUse = AccVgprReadLdStWaitStates - 1577 getWaitStatesSinceDef(Reg, IsAccVgprReadFn, MaxWaitStates); 1578 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1579 1580 if (WaitStatesNeeded == MaxWaitStates) 1581 return WaitStatesNeeded; // Early exit. 1582 1583 auto IsVALUAccVgprRdWrCheckFn = [Reg, this](const MachineInstr &MI) { 1584 if (MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64 && 1585 MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64) 1586 return false; 1587 auto IsVALUFn = [](const MachineInstr &MI) { 1588 return SIInstrInfo::isVALU(MI) && !SIInstrInfo::isMAI(MI); 1589 }; 1590 return getWaitStatesSinceDef(Reg, IsVALUFn, 2 /*MaxWaitStates*/) < 1591 std::numeric_limits<int>::max(); 1592 }; 1593 1594 WaitStatesNeededForUse = VALUWriteAccVgprRdWrLdStDepVALUWaitStates - 1595 getWaitStatesSince(IsVALUAccVgprRdWrCheckFn, MaxWaitStates); 1596 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1597 } 1598 1599 return WaitStatesNeeded; 1600 } 1601 1602 int GCNHazardRecognizer::checkMAIVALUHazards(MachineInstr *MI) { 1603 if (!ST.hasGFX90AInsts()) 1604 return 0; 1605 1606 auto IsMFMAFn = [](const MachineInstr &MI) -> bool { 1607 return SIInstrInfo::isMAI(MI) && 1608 MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 && 1609 MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64; 1610 }; 1611 1612 auto IsDGEMMFn = [](const MachineInstr &MI) -> bool { 1613 return isDGEMM(MI.getOpcode()); 1614 }; 1615 1616 // This is checked in checkMAIHazards90A() 1617 if (IsMFMAFn(*MI)) 1618 return 0; 1619 1620 int WaitStatesNeeded = 0; 1621 1622 bool IsMemOrExport = SIInstrInfo::isVMEM(*MI) || 1623 SIInstrInfo::isFLAT(*MI) || 1624 SIInstrInfo::isDS(*MI) || 1625 SIInstrInfo::isEXP(*MI); 1626 bool IsVALU = SIInstrInfo::isVALU(*MI); 1627 1628 const MachineInstr *MFMA = nullptr; 1629 unsigned Reg; 1630 auto IsDGEMMorXDLWriteFn = [&Reg, &IsMFMAFn, &MFMA, 1631 this](const MachineInstr &MI) { 1632 if (!IsMFMAFn(MI) || !TRI.regsOverlap(MI.getOperand(0).getReg(), Reg)) 1633 return false; 1634 if (!isDGEMM(MI.getOpcode()) && !isXDL(ST, MI)) 1635 return false; 1636 MFMA = &MI; 1637 return true; 1638 }; 1639 1640 const MachineInstr *DOT = nullptr; 1641 auto IsDotWriteFn = [&Reg, &DOT, this](const MachineInstr &MI) { 1642 if (!SIInstrInfo::isDOT(MI) || 1643 !TRI.regsOverlap(MI.getOperand(0).getReg(), Reg)) 1644 return false; 1645 DOT = &MI; 1646 return true; 1647 }; 1648 1649 int SrcCIdx = AMDGPU::getNamedOperandIdx(MI->getOpcode(), 1650 AMDGPU::OpName::src2); 1651 1652 if (IsMemOrExport || IsVALU) { 1653 const int SMFMA4x4WriteVgprVALUMemExpReadWaitStates = 5; 1654 const int SMFMA16x16WriteVgprVALUMemExpReadWaitStates = 11; 1655 const int SMFMA32x32WriteVgprVALUMemExpReadWaitStates = 19; 1656 const int DMFMA4x4WriteVgprMemExpReadWaitStates = 9; 1657 const int DMFMA16x16WriteVgprMemExpReadWaitStates = 18; 1658 const int DMFMA4x4WriteVgprVALUReadWaitStates = 6; 1659 const int DMFMA16x16WriteVgprVALUReadWaitStates = 11; 1660 const int DotWriteSameDotReadSrcAB = 3; 1661 const int DotWriteDifferentVALURead = 3; 1662 const int MaxWaitStates = 19; 1663 1664 for (const MachineOperand &Use : MI->explicit_uses()) { 1665 if (!Use.isReg()) 1666 continue; 1667 Reg = Use.getReg(); 1668 1669 DOT = nullptr; 1670 int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsDotWriteFn, 1671 MaxWaitStates); 1672 if (DOT) { 1673 int NeedWaitStates = 0; 1674 if (DOT->getOpcode() == MI->getOpcode()) { 1675 if (&Use - &MI->getOperand(0) != SrcCIdx) 1676 NeedWaitStates = DotWriteSameDotReadSrcAB; 1677 } else { 1678 NeedWaitStates = DotWriteDifferentVALURead; 1679 } 1680 1681 int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef; 1682 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1683 } 1684 1685 MFMA = nullptr; 1686 WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsDGEMMorXDLWriteFn, 1687 MaxWaitStates); 1688 if (!MFMA) 1689 continue; 1690 1691 unsigned HazardDefLatency = TSchedModel.computeInstrLatency(MFMA); 1692 int NeedWaitStates = MaxWaitStates; 1693 switch (HazardDefLatency) { 1694 case 2: 1695 NeedWaitStates = SMFMA4x4WriteVgprVALUMemExpReadWaitStates; 1696 break; 1697 case 4: 1698 assert(isDGEMM(MFMA->getOpcode())); 1699 NeedWaitStates = 1700 IsMemOrExport ? DMFMA4x4WriteVgprMemExpReadWaitStates 1701 : DMFMA4x4WriteVgprVALUReadWaitStates; 1702 break; 1703 case 8: 1704 NeedWaitStates = SMFMA16x16WriteVgprVALUMemExpReadWaitStates; 1705 break; 1706 case 16: LLVM_FALLTHROUGH; 1707 default: 1708 NeedWaitStates = 1709 isDGEMM(MFMA->getOpcode()) 1710 ? IsMemOrExport ? DMFMA16x16WriteVgprMemExpReadWaitStates 1711 : DMFMA16x16WriteVgprVALUReadWaitStates 1712 : SMFMA32x32WriteVgprVALUMemExpReadWaitStates; 1713 break; 1714 } 1715 1716 int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef; 1717 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1718 1719 if (WaitStatesNeeded == MaxWaitStates) 1720 break; 1721 } 1722 } 1723 1724 unsigned Opc = MI->getOpcode(); 1725 const int DMFMAToFMA64WaitStates = 2; 1726 if ((Opc == AMDGPU::V_FMA_F64_e64 || 1727 Opc == AMDGPU::V_FMAC_F64_e32 || Opc == AMDGPU::V_FMAC_F64_e64 || 1728 Opc == AMDGPU::V_FMAC_F64_dpp) && 1729 WaitStatesNeeded < DMFMAToFMA64WaitStates) { 1730 int WaitStatesNeededForUse = DMFMAToFMA64WaitStates - 1731 getWaitStatesSince(IsDGEMMFn, DMFMAToFMA64WaitStates); 1732 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1733 } 1734 1735 if (!IsVALU && !IsMemOrExport) 1736 return WaitStatesNeeded; 1737 1738 for (const MachineOperand &Def : MI->defs()) { 1739 const int SMFMA4x4WriteVgprVALUWawWaitStates = 5; 1740 const int SMFMA16x16WriteVgprVALUWawWaitStates = 11; 1741 const int SMFMA32x32WriteVgprVALUWawWaitStates = 19; 1742 const int SMFMA4x4ReadVgprVALUWarWaitStates = 1; 1743 const int SMFMA16x16ReadVgprVALUWarWaitStates = 7; 1744 const int SMFMA32x32ReadVgprVALUWarWaitStates = 15; 1745 const int DMFMA4x4WriteVgprVALUWriteWaitStates = 6; 1746 const int DMFMA16x16WriteVgprVALUWriteWaitStates = 11; 1747 const int DotWriteDifferentVALUWrite = 3; 1748 const int MaxWaitStates = 19; 1749 const int MaxWarWaitStates = 15; 1750 1751 Reg = Def.getReg(); 1752 1753 DOT = nullptr; 1754 int WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsDotWriteFn, 1755 MaxWaitStates); 1756 if (DOT && DOT->getOpcode() != MI->getOpcode()) 1757 WaitStatesNeeded = std::max(WaitStatesNeeded, DotWriteDifferentVALUWrite - 1758 WaitStatesSinceDef); 1759 1760 MFMA = nullptr; 1761 WaitStatesSinceDef = getWaitStatesSinceDef(Reg, IsDGEMMorXDLWriteFn, 1762 MaxWaitStates); 1763 if (MFMA) { 1764 int NeedWaitStates = MaxWaitStates; 1765 switch (TSchedModel.computeInstrLatency(MFMA)) { 1766 case 2: 1767 NeedWaitStates = SMFMA4x4WriteVgprVALUWawWaitStates; 1768 break; 1769 case 4: 1770 assert(isDGEMM(MFMA->getOpcode())); 1771 NeedWaitStates = DMFMA4x4WriteVgprVALUWriteWaitStates; 1772 break; 1773 case 8: 1774 NeedWaitStates = SMFMA16x16WriteVgprVALUWawWaitStates; 1775 break; 1776 case 16: LLVM_FALLTHROUGH; 1777 default: 1778 NeedWaitStates = isDGEMM(MFMA->getOpcode()) 1779 ? DMFMA16x16WriteVgprVALUWriteWaitStates 1780 : SMFMA32x32WriteVgprVALUWawWaitStates; 1781 break; 1782 } 1783 1784 int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceDef; 1785 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1786 1787 if (WaitStatesNeeded == MaxWaitStates) 1788 break; 1789 } 1790 1791 auto IsSMFMAReadAsCFn = [&Reg, &IsMFMAFn, &MFMA, 1792 this](const MachineInstr &MI) { 1793 if (!IsMFMAFn(MI) || isDGEMM(MI.getOpcode()) || 1794 !MI.readsRegister(Reg, &TRI)) 1795 return false; 1796 1797 const MachineOperand *SrcC = 1798 TII.getNamedOperand(MI, AMDGPU::OpName::src2); 1799 assert(SrcC); 1800 if (!SrcC->isReg() || !TRI.regsOverlap(SrcC->getReg(), Reg)) 1801 return false; 1802 1803 MFMA = &MI; 1804 return true; 1805 }; 1806 1807 MFMA = nullptr; 1808 int WaitStatesSinceUse = getWaitStatesSince(IsSMFMAReadAsCFn, 1809 MaxWarWaitStates); 1810 if (!MFMA) 1811 continue; 1812 1813 unsigned HazardDefLatency = TSchedModel.computeInstrLatency(MFMA); 1814 int NeedWaitStates = MaxWaitStates; 1815 switch (HazardDefLatency) { 1816 case 2: NeedWaitStates = SMFMA4x4ReadVgprVALUWarWaitStates; 1817 break; 1818 case 8: NeedWaitStates = SMFMA16x16ReadVgprVALUWarWaitStates; 1819 break; 1820 case 16: LLVM_FALLTHROUGH; 1821 default: NeedWaitStates = SMFMA32x32ReadVgprVALUWarWaitStates; 1822 break; 1823 } 1824 1825 int WaitStatesNeededForUse = NeedWaitStates - WaitStatesSinceUse; 1826 WaitStatesNeeded = std::max(WaitStatesNeeded, WaitStatesNeededForUse); 1827 } 1828 1829 return WaitStatesNeeded; 1830 } 1831 1832 bool GCNHazardRecognizer::ShouldPreferAnother(SUnit *SU) { 1833 if (!SU->isInstr()) 1834 return false; 1835 1836 const MachineInstr *MAI = nullptr; 1837 auto IsMFMAFn = [&MAI](const MachineInstr &MI) { 1838 MAI = nullptr; 1839 if (SIInstrInfo::isMAI(MI) && 1840 MI.getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64 && 1841 MI.getOpcode() != AMDGPU::V_ACCVGPR_READ_B32_e64) 1842 MAI = &MI; 1843 return MAI != nullptr; 1844 }; 1845 1846 MachineInstr *MI = SU->getInstr(); 1847 if (IsMFMAFn(*MI)) { 1848 int W = getWaitStatesSince(IsMFMAFn, 16); 1849 if (MAI) 1850 return W < (int)TSchedModel.computeInstrLatency(MAI); 1851 } 1852 1853 return false; 1854 } 1855