1 //===- SIInsertWaitcnts.cpp - Insert Wait Instructions --------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 /// \file 10 /// Insert wait instructions for memory reads and writes. 11 /// 12 /// Memory reads and writes are issued asynchronously, so we need to insert 13 /// S_WAITCNT instructions when we want to access any of their results or 14 /// overwrite any register that's used asynchronously. 15 /// 16 /// TODO: This pass currently keeps one timeline per hardware counter. A more 17 /// finely-grained approach that keeps one timeline per event type could 18 /// sometimes get away with generating weaker s_waitcnt instructions. For 19 /// example, when both SMEM and LDS are in flight and we need to wait for 20 /// the i-th-last LDS instruction, then an lgkmcnt(i) is actually sufficient, 21 /// but the pass will currently generate a conservative lgkmcnt(0) because 22 /// multiple event types are in flight. 23 // 24 //===----------------------------------------------------------------------===// 25 26 #include "AMDGPU.h" 27 #include "GCNSubtarget.h" 28 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 29 #include "SIMachineFunctionInfo.h" 30 #include "Utils/AMDGPUBaseInfo.h" 31 #include "llvm/ADT/MapVector.h" 32 #include "llvm/ADT/PostOrderIterator.h" 33 #include "llvm/ADT/Sequence.h" 34 #include "llvm/CodeGen/MachineLoopInfo.h" 35 #include "llvm/CodeGen/MachinePostDominators.h" 36 #include "llvm/InitializePasses.h" 37 #include "llvm/Support/DebugCounter.h" 38 #include "llvm/Support/TargetParser.h" 39 using namespace llvm; 40 41 #define DEBUG_TYPE "si-insert-waitcnts" 42 43 DEBUG_COUNTER(ForceExpCounter, DEBUG_TYPE"-forceexp", 44 "Force emit s_waitcnt expcnt(0) instrs"); 45 DEBUG_COUNTER(ForceLgkmCounter, DEBUG_TYPE"-forcelgkm", 46 "Force emit s_waitcnt lgkmcnt(0) instrs"); 47 DEBUG_COUNTER(ForceVMCounter, DEBUG_TYPE"-forcevm", 48 "Force emit s_waitcnt vmcnt(0) instrs"); 49 50 static cl::opt<bool> ForceEmitZeroFlag( 51 "amdgpu-waitcnt-forcezero", 52 cl::desc("Force all waitcnt instrs to be emitted as s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)"), 53 cl::init(false), cl::Hidden); 54 55 namespace { 56 // Class of object that encapsulates latest instruction counter score 57 // associated with the operand. Used for determining whether 58 // s_waitcnt instruction needs to be emitted. 59 60 #define CNT_MASK(t) (1u << (t)) 61 62 enum InstCounterType { VM_CNT = 0, LGKM_CNT, EXP_CNT, VS_CNT, NUM_INST_CNTS }; 63 } // namespace 64 65 namespace llvm { 66 template <> struct enum_iteration_traits<InstCounterType> { 67 static constexpr bool is_iterable = true; 68 }; 69 } // namespace llvm 70 71 namespace { 72 auto inst_counter_types() { return enum_seq(VM_CNT, NUM_INST_CNTS); } 73 74 using RegInterval = std::pair<int, int>; 75 76 struct HardwareLimits { 77 unsigned VmcntMax; 78 unsigned ExpcntMax; 79 unsigned LgkmcntMax; 80 unsigned VscntMax; 81 }; 82 83 struct RegisterEncoding { 84 unsigned VGPR0; 85 unsigned VGPRL; 86 unsigned SGPR0; 87 unsigned SGPRL; 88 }; 89 90 enum WaitEventType { 91 VMEM_ACCESS, // vector-memory read & write 92 VMEM_READ_ACCESS, // vector-memory read 93 VMEM_WRITE_ACCESS, // vector-memory write 94 LDS_ACCESS, // lds read & write 95 GDS_ACCESS, // gds read & write 96 SQ_MESSAGE, // send message 97 SMEM_ACCESS, // scalar-memory read & write 98 EXP_GPR_LOCK, // export holding on its data src 99 GDS_GPR_LOCK, // GDS holding on its data and addr src 100 EXP_POS_ACCESS, // write to export position 101 EXP_PARAM_ACCESS, // write to export parameter 102 VMW_GPR_LOCK, // vector-memory write holding on its data src 103 EXP_LDS_ACCESS, // read by ldsdir counting as export 104 NUM_WAIT_EVENTS, 105 }; 106 107 static const unsigned WaitEventMaskForInst[NUM_INST_CNTS] = { 108 (1 << VMEM_ACCESS) | (1 << VMEM_READ_ACCESS), 109 (1 << SMEM_ACCESS) | (1 << LDS_ACCESS) | (1 << GDS_ACCESS) | 110 (1 << SQ_MESSAGE), 111 (1 << EXP_GPR_LOCK) | (1 << GDS_GPR_LOCK) | (1 << VMW_GPR_LOCK) | 112 (1 << EXP_PARAM_ACCESS) | (1 << EXP_POS_ACCESS) | (1 << EXP_LDS_ACCESS), 113 (1 << VMEM_WRITE_ACCESS)}; 114 115 // The mapping is: 116 // 0 .. SQ_MAX_PGM_VGPRS-1 real VGPRs 117 // SQ_MAX_PGM_VGPRS .. NUM_ALL_VGPRS-1 extra VGPR-like slots 118 // NUM_ALL_VGPRS .. NUM_ALL_VGPRS+SQ_MAX_PGM_SGPRS-1 real SGPRs 119 // We reserve a fixed number of VGPR slots in the scoring tables for 120 // special tokens like SCMEM_LDS (needed for buffer load to LDS). 121 enum RegisterMapping { 122 SQ_MAX_PGM_VGPRS = 512, // Maximum programmable VGPRs across all targets. 123 AGPR_OFFSET = 256, // Maximum programmable ArchVGPRs across all targets. 124 SQ_MAX_PGM_SGPRS = 256, // Maximum programmable SGPRs across all targets. 125 NUM_EXTRA_VGPRS = 1, // A reserved slot for DS. 126 EXTRA_VGPR_LDS = 0, // An artificial register to track LDS writes. 127 NUM_ALL_VGPRS = SQ_MAX_PGM_VGPRS + NUM_EXTRA_VGPRS, // Where SGPR starts. 128 }; 129 130 // Enumerate different types of result-returning VMEM operations. Although 131 // s_waitcnt orders them all with a single vmcnt counter, in the absence of 132 // s_waitcnt only instructions of the same VmemType are guaranteed to write 133 // their results in order -- so there is no need to insert an s_waitcnt between 134 // two instructions of the same type that write the same vgpr. 135 enum VmemType { 136 // BUF instructions and MIMG instructions without a sampler. 137 VMEM_NOSAMPLER, 138 // MIMG instructions with a sampler. 139 VMEM_SAMPLER, 140 // BVH instructions 141 VMEM_BVH 142 }; 143 144 static bool updateVMCntOnly(const MachineInstr &Inst) { 145 return SIInstrInfo::isVMEM(Inst) || SIInstrInfo::isFLATGlobal(Inst) || 146 SIInstrInfo::isFLATScratch(Inst); 147 } 148 149 VmemType getVmemType(const MachineInstr &Inst) { 150 assert(updateVMCntOnly(Inst)); 151 if (!SIInstrInfo::isMIMG(Inst)) 152 return VMEM_NOSAMPLER; 153 const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(Inst.getOpcode()); 154 const AMDGPU::MIMGBaseOpcodeInfo *BaseInfo = 155 AMDGPU::getMIMGBaseOpcodeInfo(Info->BaseOpcode); 156 return BaseInfo->BVH ? VMEM_BVH 157 : BaseInfo->Sampler ? VMEM_SAMPLER : VMEM_NOSAMPLER; 158 } 159 160 void addWait(AMDGPU::Waitcnt &Wait, InstCounterType T, unsigned Count) { 161 switch (T) { 162 case VM_CNT: 163 Wait.VmCnt = std::min(Wait.VmCnt, Count); 164 break; 165 case EXP_CNT: 166 Wait.ExpCnt = std::min(Wait.ExpCnt, Count); 167 break; 168 case LGKM_CNT: 169 Wait.LgkmCnt = std::min(Wait.LgkmCnt, Count); 170 break; 171 case VS_CNT: 172 Wait.VsCnt = std::min(Wait.VsCnt, Count); 173 break; 174 default: 175 llvm_unreachable("bad InstCounterType"); 176 } 177 } 178 179 // This objects maintains the current score brackets of each wait counter, and 180 // a per-register scoreboard for each wait counter. 181 // 182 // We also maintain the latest score for every event type that can change the 183 // waitcnt in order to know if there are multiple types of events within 184 // the brackets. When multiple types of event happen in the bracket, 185 // wait count may get decreased out of order, therefore we need to put in 186 // "s_waitcnt 0" before use. 187 class WaitcntBrackets { 188 public: 189 WaitcntBrackets(const GCNSubtarget *SubTarget, HardwareLimits Limits, 190 RegisterEncoding Encoding) 191 : ST(SubTarget), Limits(Limits), Encoding(Encoding) {} 192 193 unsigned getWaitCountMax(InstCounterType T) const { 194 switch (T) { 195 case VM_CNT: 196 return Limits.VmcntMax; 197 case LGKM_CNT: 198 return Limits.LgkmcntMax; 199 case EXP_CNT: 200 return Limits.ExpcntMax; 201 case VS_CNT: 202 return Limits.VscntMax; 203 default: 204 break; 205 } 206 return 0; 207 } 208 209 unsigned getScoreLB(InstCounterType T) const { 210 assert(T < NUM_INST_CNTS); 211 return ScoreLBs[T]; 212 } 213 214 unsigned getScoreUB(InstCounterType T) const { 215 assert(T < NUM_INST_CNTS); 216 return ScoreUBs[T]; 217 } 218 219 unsigned getScoreRange(InstCounterType T) const { 220 return getScoreUB(T) - getScoreLB(T); 221 } 222 223 // Mapping from event to counter. 224 InstCounterType eventCounter(WaitEventType E) const { 225 for (auto T : inst_counter_types()) { 226 if (WaitEventMaskForInst[T] & (1 << E)) 227 return T; 228 } 229 llvm_unreachable("event type has no associated counter"); 230 } 231 232 unsigned getRegScore(int GprNo, InstCounterType T) const { 233 if (GprNo < NUM_ALL_VGPRS) { 234 return VgprScores[T][GprNo]; 235 } 236 assert(T == LGKM_CNT); 237 return SgprScores[GprNo - NUM_ALL_VGPRS]; 238 } 239 240 bool merge(const WaitcntBrackets &Other); 241 242 RegInterval getRegInterval(const MachineInstr *MI, const SIInstrInfo *TII, 243 const MachineRegisterInfo *MRI, 244 const SIRegisterInfo *TRI, unsigned OpNo) const; 245 246 bool counterOutOfOrder(InstCounterType T) const; 247 void simplifyWaitcnt(AMDGPU::Waitcnt &Wait) const; 248 void simplifyWaitcnt(InstCounterType T, unsigned &Count) const; 249 void determineWait(InstCounterType T, int RegNo, AMDGPU::Waitcnt &Wait) const; 250 void applyWaitcnt(const AMDGPU::Waitcnt &Wait); 251 void applyWaitcnt(InstCounterType T, unsigned Count); 252 void updateByEvent(const SIInstrInfo *TII, const SIRegisterInfo *TRI, 253 const MachineRegisterInfo *MRI, WaitEventType E, 254 MachineInstr &MI); 255 256 unsigned hasPendingEvent() const { return PendingEvents; } 257 unsigned hasPendingEvent(WaitEventType E) const { 258 return PendingEvents & (1 << E); 259 } 260 unsigned hasPendingEvent(InstCounterType T) const { 261 unsigned HasPending = PendingEvents & WaitEventMaskForInst[T]; 262 assert((HasPending != 0) == (getScoreRange(T) != 0)); 263 return HasPending; 264 } 265 266 bool hasMixedPendingEvents(InstCounterType T) const { 267 unsigned Events = hasPendingEvent(T); 268 // Return true if more than one bit is set in Events. 269 return Events & (Events - 1); 270 } 271 272 bool hasPendingFlat() const { 273 return ((LastFlat[LGKM_CNT] > ScoreLBs[LGKM_CNT] && 274 LastFlat[LGKM_CNT] <= ScoreUBs[LGKM_CNT]) || 275 (LastFlat[VM_CNT] > ScoreLBs[VM_CNT] && 276 LastFlat[VM_CNT] <= ScoreUBs[VM_CNT])); 277 } 278 279 void setPendingFlat() { 280 LastFlat[VM_CNT] = ScoreUBs[VM_CNT]; 281 LastFlat[LGKM_CNT] = ScoreUBs[LGKM_CNT]; 282 } 283 284 // Return true if there might be pending writes to the specified vgpr by VMEM 285 // instructions with types different from V. 286 bool hasOtherPendingVmemTypes(int GprNo, VmemType V) const { 287 assert(GprNo < NUM_ALL_VGPRS); 288 return VgprVmemTypes[GprNo] & ~(1 << V); 289 } 290 291 void clearVgprVmemTypes(int GprNo) { 292 assert(GprNo < NUM_ALL_VGPRS); 293 VgprVmemTypes[GprNo] = 0; 294 } 295 296 void print(raw_ostream &); 297 void dump() { print(dbgs()); } 298 299 private: 300 struct MergeInfo { 301 unsigned OldLB; 302 unsigned OtherLB; 303 unsigned MyShift; 304 unsigned OtherShift; 305 }; 306 static bool mergeScore(const MergeInfo &M, unsigned &Score, 307 unsigned OtherScore); 308 309 void setScoreLB(InstCounterType T, unsigned Val) { 310 assert(T < NUM_INST_CNTS); 311 ScoreLBs[T] = Val; 312 } 313 314 void setScoreUB(InstCounterType T, unsigned Val) { 315 assert(T < NUM_INST_CNTS); 316 ScoreUBs[T] = Val; 317 318 if (T != EXP_CNT) 319 return; 320 321 if (getScoreRange(EXP_CNT) > getWaitCountMax(EXP_CNT)) 322 ScoreLBs[EXP_CNT] = ScoreUBs[EXP_CNT] - getWaitCountMax(EXP_CNT); 323 } 324 325 void setRegScore(int GprNo, InstCounterType T, unsigned Val) { 326 if (GprNo < NUM_ALL_VGPRS) { 327 VgprUB = std::max(VgprUB, GprNo); 328 VgprScores[T][GprNo] = Val; 329 } else { 330 assert(T == LGKM_CNT); 331 SgprUB = std::max(SgprUB, GprNo - NUM_ALL_VGPRS); 332 SgprScores[GprNo - NUM_ALL_VGPRS] = Val; 333 } 334 } 335 336 void setExpScore(const MachineInstr *MI, const SIInstrInfo *TII, 337 const SIRegisterInfo *TRI, const MachineRegisterInfo *MRI, 338 unsigned OpNo, unsigned Val); 339 340 const GCNSubtarget *ST = nullptr; 341 HardwareLimits Limits = {}; 342 RegisterEncoding Encoding = {}; 343 unsigned ScoreLBs[NUM_INST_CNTS] = {0}; 344 unsigned ScoreUBs[NUM_INST_CNTS] = {0}; 345 unsigned PendingEvents = 0; 346 // Remember the last flat memory operation. 347 unsigned LastFlat[NUM_INST_CNTS] = {0}; 348 // wait_cnt scores for every vgpr. 349 // Keep track of the VgprUB and SgprUB to make merge at join efficient. 350 int VgprUB = -1; 351 int SgprUB = -1; 352 unsigned VgprScores[NUM_INST_CNTS][NUM_ALL_VGPRS] = {{0}}; 353 // Wait cnt scores for every sgpr, only lgkmcnt is relevant. 354 unsigned SgprScores[SQ_MAX_PGM_SGPRS] = {0}; 355 // Bitmask of the VmemTypes of VMEM instructions that might have a pending 356 // write to each vgpr. 357 unsigned char VgprVmemTypes[NUM_ALL_VGPRS] = {0}; 358 }; 359 360 class SIInsertWaitcnts : public MachineFunctionPass { 361 private: 362 const GCNSubtarget *ST = nullptr; 363 const SIInstrInfo *TII = nullptr; 364 const SIRegisterInfo *TRI = nullptr; 365 const MachineRegisterInfo *MRI = nullptr; 366 AMDGPU::IsaVersion IV; 367 368 DenseSet<MachineInstr *> TrackedWaitcntSet; 369 DenseMap<const Value *, MachineBasicBlock *> SLoadAddresses; 370 DenseMap<MachineBasicBlock *, bool> PreheadersToFlush; 371 MachineLoopInfo *MLI; 372 MachinePostDominatorTree *PDT; 373 374 struct BlockInfo { 375 MachineBasicBlock *MBB; 376 std::unique_ptr<WaitcntBrackets> Incoming; 377 bool Dirty = true; 378 379 explicit BlockInfo(MachineBasicBlock *MBB) : MBB(MBB) {} 380 }; 381 382 MapVector<MachineBasicBlock *, BlockInfo> BlockInfos; 383 384 // ForceEmitZeroWaitcnts: force all waitcnts insts to be s_waitcnt 0 385 // because of amdgpu-waitcnt-forcezero flag 386 bool ForceEmitZeroWaitcnts; 387 bool ForceEmitWaitcnt[NUM_INST_CNTS]; 388 389 public: 390 static char ID; 391 392 SIInsertWaitcnts() : MachineFunctionPass(ID) { 393 (void)ForceExpCounter; 394 (void)ForceLgkmCounter; 395 (void)ForceVMCounter; 396 } 397 398 bool shouldFlushVmCnt(MachineLoop *ML, WaitcntBrackets &Brackets); 399 bool isPreheaderToFlush(MachineBasicBlock &MBB, 400 WaitcntBrackets &ScoreBrackets); 401 bool runOnMachineFunction(MachineFunction &MF) override; 402 403 StringRef getPassName() const override { 404 return "SI insert wait instructions"; 405 } 406 407 void getAnalysisUsage(AnalysisUsage &AU) const override { 408 AU.setPreservesCFG(); 409 AU.addRequired<MachineLoopInfo>(); 410 AU.addRequired<MachinePostDominatorTree>(); 411 MachineFunctionPass::getAnalysisUsage(AU); 412 } 413 414 bool isForceEmitWaitcnt() const { 415 for (auto T : inst_counter_types()) 416 if (ForceEmitWaitcnt[T]) 417 return true; 418 return false; 419 } 420 421 AMDGPU::Waitcnt allZeroWaitcnt() const { 422 return AMDGPU::Waitcnt::allZero(ST->hasVscnt()); 423 } 424 425 void setForceEmitWaitcnt() { 426 // For non-debug builds, ForceEmitWaitcnt has been initialized to false; 427 // For debug builds, get the debug counter info and adjust if need be 428 #ifndef NDEBUG 429 if (DebugCounter::isCounterSet(ForceExpCounter) && 430 DebugCounter::shouldExecute(ForceExpCounter)) { 431 ForceEmitWaitcnt[EXP_CNT] = true; 432 } else { 433 ForceEmitWaitcnt[EXP_CNT] = false; 434 } 435 436 if (DebugCounter::isCounterSet(ForceLgkmCounter) && 437 DebugCounter::shouldExecute(ForceLgkmCounter)) { 438 ForceEmitWaitcnt[LGKM_CNT] = true; 439 } else { 440 ForceEmitWaitcnt[LGKM_CNT] = false; 441 } 442 443 if (DebugCounter::isCounterSet(ForceVMCounter) && 444 DebugCounter::shouldExecute(ForceVMCounter)) { 445 ForceEmitWaitcnt[VM_CNT] = true; 446 } else { 447 ForceEmitWaitcnt[VM_CNT] = false; 448 } 449 #endif // NDEBUG 450 } 451 452 // Return the appropriate VMEM_*_ACCESS type for Inst, which must be a VMEM or 453 // FLAT instruction. 454 WaitEventType getVmemWaitEventType(const MachineInstr &Inst) const { 455 assert(SIInstrInfo::isVMEM(Inst) || SIInstrInfo::isFLAT(Inst)); 456 if (!ST->hasVscnt()) 457 return VMEM_ACCESS; 458 if (Inst.mayStore() && !SIInstrInfo::isAtomicRet(Inst)) 459 return VMEM_WRITE_ACCESS; 460 return VMEM_READ_ACCESS; 461 } 462 463 bool mayAccessVMEMThroughFlat(const MachineInstr &MI) const; 464 bool mayAccessLDSThroughFlat(const MachineInstr &MI) const; 465 bool generateWaitcntInstBefore(MachineInstr &MI, 466 WaitcntBrackets &ScoreBrackets, 467 MachineInstr *OldWaitcntInstr, 468 bool FlushVmCnt); 469 bool generateWaitcntBlockEnd(MachineBasicBlock &Block, 470 WaitcntBrackets &ScoreBrackets, 471 MachineInstr *OldWaitcntInstr); 472 bool generateWaitcnt(AMDGPU::Waitcnt Wait, 473 MachineBasicBlock::instr_iterator It, 474 MachineBasicBlock &Block, WaitcntBrackets &ScoreBrackets, 475 MachineInstr *OldWaitcntInstr); 476 void updateEventWaitcntAfter(MachineInstr &Inst, 477 WaitcntBrackets *ScoreBrackets); 478 bool insertWaitcntInBlock(MachineFunction &MF, MachineBasicBlock &Block, 479 WaitcntBrackets &ScoreBrackets); 480 bool applyPreexistingWaitcnt(WaitcntBrackets &ScoreBrackets, 481 MachineInstr &OldWaitcntInstr, 482 AMDGPU::Waitcnt &Wait, 483 MachineBasicBlock::instr_iterator It) const; 484 }; 485 486 } // end anonymous namespace 487 488 RegInterval WaitcntBrackets::getRegInterval(const MachineInstr *MI, 489 const SIInstrInfo *TII, 490 const MachineRegisterInfo *MRI, 491 const SIRegisterInfo *TRI, 492 unsigned OpNo) const { 493 const MachineOperand &Op = MI->getOperand(OpNo); 494 if (!TRI->isInAllocatableClass(Op.getReg())) 495 return {-1, -1}; 496 497 // A use via a PW operand does not need a waitcnt. 498 // A partial write is not a WAW. 499 assert(!Op.getSubReg() || !Op.isUndef()); 500 501 RegInterval Result; 502 503 unsigned Reg = TRI->getEncodingValue(AMDGPU::getMCReg(Op.getReg(), *ST)); 504 505 if (TRI->isVectorRegister(*MRI, Op.getReg())) { 506 assert(Reg >= Encoding.VGPR0 && Reg <= Encoding.VGPRL); 507 Result.first = Reg - Encoding.VGPR0; 508 if (TRI->isAGPR(*MRI, Op.getReg())) 509 Result.first += AGPR_OFFSET; 510 assert(Result.first >= 0 && Result.first < SQ_MAX_PGM_VGPRS); 511 } else if (TRI->isSGPRReg(*MRI, Op.getReg())) { 512 assert(Reg >= Encoding.SGPR0 && Reg < SQ_MAX_PGM_SGPRS); 513 Result.first = Reg - Encoding.SGPR0 + NUM_ALL_VGPRS; 514 assert(Result.first >= NUM_ALL_VGPRS && 515 Result.first < SQ_MAX_PGM_SGPRS + NUM_ALL_VGPRS); 516 } 517 // TODO: Handle TTMP 518 // else if (TRI->isTTMP(*MRI, Reg.getReg())) ... 519 else 520 return {-1, -1}; 521 522 const TargetRegisterClass *RC = TII->getOpRegClass(*MI, OpNo); 523 unsigned Size = TRI->getRegSizeInBits(*RC); 524 Result.second = Result.first + ((Size + 16) / 32); 525 526 return Result; 527 } 528 529 void WaitcntBrackets::setExpScore(const MachineInstr *MI, 530 const SIInstrInfo *TII, 531 const SIRegisterInfo *TRI, 532 const MachineRegisterInfo *MRI, unsigned OpNo, 533 unsigned Val) { 534 RegInterval Interval = getRegInterval(MI, TII, MRI, TRI, OpNo); 535 assert(TRI->isVectorRegister(*MRI, MI->getOperand(OpNo).getReg())); 536 for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo) { 537 setRegScore(RegNo, EXP_CNT, Val); 538 } 539 } 540 541 // MUBUF and FLAT LDS DMA operations need a wait on vmcnt before LDS written 542 // can be accessed. A load from LDS to VMEM does not need a wait. 543 static bool mayWriteLDSThroughDMA(const MachineInstr &MI) { 544 return SIInstrInfo::isVALU(MI) && 545 (SIInstrInfo::isMUBUF(MI) || SIInstrInfo::isFLAT(MI)) && 546 MI.getOpcode() != AMDGPU::BUFFER_STORE_LDS_DWORD; 547 } 548 549 void WaitcntBrackets::updateByEvent(const SIInstrInfo *TII, 550 const SIRegisterInfo *TRI, 551 const MachineRegisterInfo *MRI, 552 WaitEventType E, MachineInstr &Inst) { 553 InstCounterType T = eventCounter(E); 554 unsigned CurrScore = getScoreUB(T) + 1; 555 if (CurrScore == 0) 556 report_fatal_error("InsertWaitcnt score wraparound"); 557 // PendingEvents and ScoreUB need to be update regardless if this event 558 // changes the score of a register or not. 559 // Examples including vm_cnt when buffer-store or lgkm_cnt when send-message. 560 PendingEvents |= 1 << E; 561 setScoreUB(T, CurrScore); 562 563 if (T == EXP_CNT) { 564 // Put score on the source vgprs. If this is a store, just use those 565 // specific register(s). 566 if (TII->isDS(Inst) && (Inst.mayStore() || Inst.mayLoad())) { 567 int AddrOpIdx = 568 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::addr); 569 // All GDS operations must protect their address register (same as 570 // export.) 571 if (AddrOpIdx != -1) { 572 setExpScore(&Inst, TII, TRI, MRI, AddrOpIdx, CurrScore); 573 } 574 575 if (Inst.mayStore()) { 576 if (AMDGPU::hasNamedOperand(Inst.getOpcode(), AMDGPU::OpName::data0)) { 577 setExpScore( 578 &Inst, TII, TRI, MRI, 579 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data0), 580 CurrScore); 581 } 582 if (AMDGPU::hasNamedOperand(Inst.getOpcode(), AMDGPU::OpName::data1)) { 583 setExpScore(&Inst, TII, TRI, MRI, 584 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), 585 AMDGPU::OpName::data1), 586 CurrScore); 587 } 588 } else if (SIInstrInfo::isAtomicRet(Inst) && 589 Inst.getOpcode() != AMDGPU::DS_GWS_INIT && 590 Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_V && 591 Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_BR && 592 Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_P && 593 Inst.getOpcode() != AMDGPU::DS_GWS_BARRIER && 594 Inst.getOpcode() != AMDGPU::DS_APPEND && 595 Inst.getOpcode() != AMDGPU::DS_CONSUME && 596 Inst.getOpcode() != AMDGPU::DS_ORDERED_COUNT) { 597 for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) { 598 const MachineOperand &Op = Inst.getOperand(I); 599 if (Op.isReg() && !Op.isDef() && 600 TRI->isVectorRegister(*MRI, Op.getReg())) { 601 setExpScore(&Inst, TII, TRI, MRI, I, CurrScore); 602 } 603 } 604 } 605 } else if (TII->isFLAT(Inst)) { 606 if (Inst.mayStore()) { 607 setExpScore( 608 &Inst, TII, TRI, MRI, 609 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data), 610 CurrScore); 611 } else if (SIInstrInfo::isAtomicRet(Inst)) { 612 setExpScore( 613 &Inst, TII, TRI, MRI, 614 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data), 615 CurrScore); 616 } 617 } else if (TII->isMIMG(Inst)) { 618 if (Inst.mayStore()) { 619 setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore); 620 } else if (SIInstrInfo::isAtomicRet(Inst)) { 621 setExpScore( 622 &Inst, TII, TRI, MRI, 623 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data), 624 CurrScore); 625 } 626 } else if (TII->isMTBUF(Inst)) { 627 if (Inst.mayStore()) { 628 setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore); 629 } 630 } else if (TII->isMUBUF(Inst)) { 631 if (Inst.mayStore()) { 632 setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore); 633 } else if (SIInstrInfo::isAtomicRet(Inst)) { 634 setExpScore( 635 &Inst, TII, TRI, MRI, 636 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data), 637 CurrScore); 638 } 639 } else if (TII->isLDSDIR(Inst)) { 640 // LDSDIR instructions attach the score to the destination. 641 setExpScore( 642 &Inst, TII, TRI, MRI, 643 AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::vdst), 644 CurrScore); 645 } else { 646 if (TII->isEXP(Inst)) { 647 // For export the destination registers are really temps that 648 // can be used as the actual source after export patching, so 649 // we need to treat them like sources and set the EXP_CNT 650 // score. 651 for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) { 652 MachineOperand &DefMO = Inst.getOperand(I); 653 if (DefMO.isReg() && DefMO.isDef() && 654 TRI->isVGPR(*MRI, DefMO.getReg())) { 655 setRegScore( 656 TRI->getEncodingValue(AMDGPU::getMCReg(DefMO.getReg(), *ST)), 657 EXP_CNT, CurrScore); 658 } 659 } 660 } 661 for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) { 662 MachineOperand &MO = Inst.getOperand(I); 663 if (MO.isReg() && !MO.isDef() && 664 TRI->isVectorRegister(*MRI, MO.getReg())) { 665 setExpScore(&Inst, TII, TRI, MRI, I, CurrScore); 666 } 667 } 668 } 669 #if 0 // TODO: check if this is handled by MUBUF code above. 670 } else if (Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORD || 671 Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORDX2 || 672 Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORDX4) { 673 MachineOperand *MO = TII->getNamedOperand(Inst, AMDGPU::OpName::data); 674 unsigned OpNo;//TODO: find the OpNo for this operand; 675 RegInterval Interval = getRegInterval(&Inst, TII, MRI, TRI, OpNo); 676 for (int RegNo = Interval.first; RegNo < Interval.second; 677 ++RegNo) { 678 setRegScore(RegNo + NUM_ALL_VGPRS, t, CurrScore); 679 } 680 #endif 681 } else { 682 // Match the score to the destination registers. 683 for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) { 684 auto &Op = Inst.getOperand(I); 685 if (!Op.isReg() || !Op.isDef()) 686 continue; 687 RegInterval Interval = getRegInterval(&Inst, TII, MRI, TRI, I); 688 if (T == VM_CNT) { 689 if (Interval.first >= NUM_ALL_VGPRS) 690 continue; 691 if (updateVMCntOnly(Inst)) { 692 VmemType V = getVmemType(Inst); 693 for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo) 694 VgprVmemTypes[RegNo] |= 1 << V; 695 } 696 } 697 for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo) { 698 setRegScore(RegNo, T, CurrScore); 699 } 700 } 701 if (Inst.mayStore() && (TII->isDS(Inst) || mayWriteLDSThroughDMA(Inst))) { 702 setRegScore(SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS, T, CurrScore); 703 } 704 } 705 } 706 707 void WaitcntBrackets::print(raw_ostream &OS) { 708 OS << '\n'; 709 for (auto T : inst_counter_types()) { 710 unsigned SR = getScoreRange(T); 711 712 switch (T) { 713 case VM_CNT: 714 OS << " VM_CNT(" << SR << "): "; 715 break; 716 case LGKM_CNT: 717 OS << " LGKM_CNT(" << SR << "): "; 718 break; 719 case EXP_CNT: 720 OS << " EXP_CNT(" << SR << "): "; 721 break; 722 case VS_CNT: 723 OS << " VS_CNT(" << SR << "): "; 724 break; 725 default: 726 OS << " UNKNOWN(" << SR << "): "; 727 break; 728 } 729 730 if (SR != 0) { 731 // Print vgpr scores. 732 unsigned LB = getScoreLB(T); 733 734 for (int J = 0; J <= VgprUB; J++) { 735 unsigned RegScore = getRegScore(J, T); 736 if (RegScore <= LB) 737 continue; 738 unsigned RelScore = RegScore - LB - 1; 739 if (J < SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS) { 740 OS << RelScore << ":v" << J << " "; 741 } else { 742 OS << RelScore << ":ds "; 743 } 744 } 745 // Also need to print sgpr scores for lgkm_cnt. 746 if (T == LGKM_CNT) { 747 for (int J = 0; J <= SgprUB; J++) { 748 unsigned RegScore = getRegScore(J + NUM_ALL_VGPRS, LGKM_CNT); 749 if (RegScore <= LB) 750 continue; 751 unsigned RelScore = RegScore - LB - 1; 752 OS << RelScore << ":s" << J << " "; 753 } 754 } 755 } 756 OS << '\n'; 757 } 758 OS << '\n'; 759 } 760 761 /// Simplify the waitcnt, in the sense of removing redundant counts, and return 762 /// whether a waitcnt instruction is needed at all. 763 void WaitcntBrackets::simplifyWaitcnt(AMDGPU::Waitcnt &Wait) const { 764 simplifyWaitcnt(VM_CNT, Wait.VmCnt); 765 simplifyWaitcnt(EXP_CNT, Wait.ExpCnt); 766 simplifyWaitcnt(LGKM_CNT, Wait.LgkmCnt); 767 simplifyWaitcnt(VS_CNT, Wait.VsCnt); 768 } 769 770 void WaitcntBrackets::simplifyWaitcnt(InstCounterType T, 771 unsigned &Count) const { 772 // The number of outstanding events for this type, T, can be calculated 773 // as (UB - LB). If the current Count is greater than or equal to the number 774 // of outstanding events, then the wait for this counter is redundant. 775 if (Count >= getScoreRange(T)) 776 Count = ~0u; 777 } 778 779 void WaitcntBrackets::determineWait(InstCounterType T, int RegNo, 780 AMDGPU::Waitcnt &Wait) const { 781 unsigned ScoreToWait = getRegScore(RegNo, T); 782 783 // If the score of src_operand falls within the bracket, we need an 784 // s_waitcnt instruction. 785 const unsigned LB = getScoreLB(T); 786 const unsigned UB = getScoreUB(T); 787 if ((UB >= ScoreToWait) && (ScoreToWait > LB)) { 788 if ((T == VM_CNT || T == LGKM_CNT) && 789 hasPendingFlat() && 790 !ST->hasFlatLgkmVMemCountInOrder()) { 791 // If there is a pending FLAT operation, and this is a VMem or LGKM 792 // waitcnt and the target can report early completion, then we need 793 // to force a waitcnt 0. 794 addWait(Wait, T, 0); 795 } else if (counterOutOfOrder(T)) { 796 // Counter can get decremented out-of-order when there 797 // are multiple types event in the bracket. Also emit an s_wait counter 798 // with a conservative value of 0 for the counter. 799 addWait(Wait, T, 0); 800 } else { 801 // If a counter has been maxed out avoid overflow by waiting for 802 // MAX(CounterType) - 1 instead. 803 unsigned NeededWait = std::min(UB - ScoreToWait, getWaitCountMax(T) - 1); 804 addWait(Wait, T, NeededWait); 805 } 806 } 807 } 808 809 void WaitcntBrackets::applyWaitcnt(const AMDGPU::Waitcnt &Wait) { 810 applyWaitcnt(VM_CNT, Wait.VmCnt); 811 applyWaitcnt(EXP_CNT, Wait.ExpCnt); 812 applyWaitcnt(LGKM_CNT, Wait.LgkmCnt); 813 applyWaitcnt(VS_CNT, Wait.VsCnt); 814 } 815 816 void WaitcntBrackets::applyWaitcnt(InstCounterType T, unsigned Count) { 817 const unsigned UB = getScoreUB(T); 818 if (Count >= UB) 819 return; 820 if (Count != 0) { 821 if (counterOutOfOrder(T)) 822 return; 823 setScoreLB(T, std::max(getScoreLB(T), UB - Count)); 824 } else { 825 setScoreLB(T, UB); 826 PendingEvents &= ~WaitEventMaskForInst[T]; 827 } 828 } 829 830 // Where there are multiple types of event in the bracket of a counter, 831 // the decrement may go out of order. 832 bool WaitcntBrackets::counterOutOfOrder(InstCounterType T) const { 833 // Scalar memory read always can go out of order. 834 if (T == LGKM_CNT && hasPendingEvent(SMEM_ACCESS)) 835 return true; 836 return hasMixedPendingEvents(T); 837 } 838 839 INITIALIZE_PASS_BEGIN(SIInsertWaitcnts, DEBUG_TYPE, "SI Insert Waitcnts", false, 840 false) 841 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 842 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree) 843 INITIALIZE_PASS_END(SIInsertWaitcnts, DEBUG_TYPE, "SI Insert Waitcnts", false, 844 false) 845 846 char SIInsertWaitcnts::ID = 0; 847 848 char &llvm::SIInsertWaitcntsID = SIInsertWaitcnts::ID; 849 850 FunctionPass *llvm::createSIInsertWaitcntsPass() { 851 return new SIInsertWaitcnts(); 852 } 853 854 static bool updateOperandIfDifferent(MachineInstr &MI, uint16_t OpName, 855 unsigned NewEnc) { 856 int OpIdx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OpName); 857 assert(OpIdx >= 0); 858 859 MachineOperand &MO = MI.getOperand(OpIdx); 860 861 if (NewEnc == MO.getImm()) 862 return false; 863 864 MO.setImm(NewEnc); 865 return true; 866 } 867 868 /// Combine consecutive waitcnt instructions that precede \p It and follow 869 /// \p OldWaitcntInstr and apply any extra wait from waitcnt that were added 870 /// by previous passes. Currently this pass conservatively assumes that these 871 /// preexisting waitcnt are required for correctness. 872 bool SIInsertWaitcnts::applyPreexistingWaitcnt( 873 WaitcntBrackets &ScoreBrackets, MachineInstr &OldWaitcntInstr, 874 AMDGPU::Waitcnt &Wait, MachineBasicBlock::instr_iterator It) const { 875 bool Modified = false; 876 MachineInstr *WaitcntInstr = nullptr; 877 MachineInstr *WaitcntVsCntInstr = nullptr; 878 879 for (auto &II : 880 make_early_inc_range(make_range(OldWaitcntInstr.getIterator(), It))) { 881 if (II.isMetaInstruction()) 882 continue; 883 884 if (II.getOpcode() == AMDGPU::S_WAITCNT) { 885 // Conservatively update required wait if this waitcnt was added in an 886 // earlier pass. In this case it will not exist in the tracked waitcnt 887 // set. 888 if (!TrackedWaitcntSet.count(&II)) { 889 unsigned IEnc = II.getOperand(0).getImm(); 890 AMDGPU::Waitcnt OldWait = AMDGPU::decodeWaitcnt(IV, IEnc); 891 Wait = Wait.combined(OldWait); 892 } 893 894 // Merge consecutive waitcnt of the same type by erasing multiples. 895 if (!WaitcntInstr) { 896 WaitcntInstr = &II; 897 } else { 898 II.eraseFromParent(); 899 Modified = true; 900 } 901 902 } else { 903 assert(II.getOpcode() == AMDGPU::S_WAITCNT_VSCNT); 904 assert(II.getOperand(0).getReg() == AMDGPU::SGPR_NULL); 905 if (!TrackedWaitcntSet.count(&II)) { 906 unsigned OldVSCnt = 907 TII->getNamedOperand(II, AMDGPU::OpName::simm16)->getImm(); 908 Wait.VsCnt = std::min(Wait.VsCnt, OldVSCnt); 909 } 910 911 if (!WaitcntVsCntInstr) { 912 WaitcntVsCntInstr = &II; 913 } else { 914 II.eraseFromParent(); 915 Modified = true; 916 } 917 } 918 } 919 920 // Updated encoding of merged waitcnt with the required wait. 921 if (WaitcntInstr) { 922 if (Wait.hasWaitExceptVsCnt()) { 923 Modified |= 924 updateOperandIfDifferent(*WaitcntInstr, AMDGPU::OpName::simm16, 925 AMDGPU::encodeWaitcnt(IV, Wait)); 926 ScoreBrackets.applyWaitcnt(Wait); 927 Wait.VmCnt = ~0u; 928 Wait.LgkmCnt = ~0u; 929 Wait.ExpCnt = ~0u; 930 931 LLVM_DEBUG(It == OldWaitcntInstr.getParent()->end() 932 ? dbgs() << "applyPreexistingWaitcnt\n" 933 << "New Instr at block end: " << *WaitcntInstr 934 << '\n' 935 : dbgs() << "applyPreexistingWaitcnt\n" 936 << "Old Instr: " << *It 937 << "New Instr: " << *WaitcntInstr << '\n'); 938 939 } else { 940 WaitcntInstr->eraseFromParent(); 941 Modified = true; 942 } 943 } 944 945 if (WaitcntVsCntInstr) { 946 if (Wait.hasWaitVsCnt()) { 947 assert(ST->hasVscnt()); 948 Modified |= updateOperandIfDifferent(*WaitcntVsCntInstr, 949 AMDGPU::OpName::simm16, Wait.VsCnt); 950 ScoreBrackets.applyWaitcnt(Wait); 951 Wait.VsCnt = ~0u; 952 953 LLVM_DEBUG(It == OldWaitcntInstr.getParent()->end() 954 ? dbgs() << "applyPreexistingWaitcnt\n" 955 << "New Instr at block end: " 956 << *WaitcntVsCntInstr << '\n' 957 : dbgs() << "applyPreexistingWaitcnt\n" 958 << "Old Instr: " << *It 959 << "New Instr: " << *WaitcntVsCntInstr << '\n'); 960 } else { 961 WaitcntVsCntInstr->eraseFromParent(); 962 Modified = true; 963 } 964 } 965 966 return Modified; 967 } 968 969 static bool readsVCCZ(const MachineInstr &MI) { 970 unsigned Opc = MI.getOpcode(); 971 return (Opc == AMDGPU::S_CBRANCH_VCCNZ || Opc == AMDGPU::S_CBRANCH_VCCZ) && 972 !MI.getOperand(1).isUndef(); 973 } 974 975 /// \returns true if the callee inserts an s_waitcnt 0 on function entry. 976 static bool callWaitsOnFunctionEntry(const MachineInstr &MI) { 977 // Currently all conventions wait, but this may not always be the case. 978 // 979 // TODO: If IPRA is enabled, and the callee is isSafeForNoCSROpt, it may make 980 // senses to omit the wait and do it in the caller. 981 return true; 982 } 983 984 /// \returns true if the callee is expected to wait for any outstanding waits 985 /// before returning. 986 static bool callWaitsOnFunctionReturn(const MachineInstr &MI) { 987 return true; 988 } 989 990 /// Generate s_waitcnt instruction to be placed before cur_Inst. 991 /// Instructions of a given type are returned in order, 992 /// but instructions of different types can complete out of order. 993 /// We rely on this in-order completion 994 /// and simply assign a score to the memory access instructions. 995 /// We keep track of the active "score bracket" to determine 996 /// if an access of a memory read requires an s_waitcnt 997 /// and if so what the value of each counter is. 998 /// The "score bracket" is bound by the lower bound and upper bound 999 /// scores (*_score_LB and *_score_ub respectively). 1000 /// If FlushVmCnt is true, that means that we want to generate a s_waitcnt to 1001 /// flush the vmcnt counter here. 1002 bool SIInsertWaitcnts::generateWaitcntInstBefore(MachineInstr &MI, 1003 WaitcntBrackets &ScoreBrackets, 1004 MachineInstr *OldWaitcntInstr, 1005 bool FlushVmCnt) { 1006 setForceEmitWaitcnt(); 1007 1008 if (MI.isMetaInstruction()) 1009 return false; 1010 1011 AMDGPU::Waitcnt Wait; 1012 1013 // FIXME: This should have already been handled by the memory legalizer. 1014 // Removing this currently doesn't affect any lit tests, but we need to 1015 // verify that nothing was relying on this. The number of buffer invalidates 1016 // being handled here should not be expanded. 1017 if (MI.getOpcode() == AMDGPU::BUFFER_WBINVL1 || 1018 MI.getOpcode() == AMDGPU::BUFFER_WBINVL1_SC || 1019 MI.getOpcode() == AMDGPU::BUFFER_WBINVL1_VOL || 1020 MI.getOpcode() == AMDGPU::BUFFER_GL0_INV || 1021 MI.getOpcode() == AMDGPU::BUFFER_GL1_INV) { 1022 Wait.VmCnt = 0; 1023 } 1024 1025 // All waits must be resolved at call return. 1026 // NOTE: this could be improved with knowledge of all call sites or 1027 // with knowledge of the called routines. 1028 if (MI.getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG || 1029 MI.getOpcode() == AMDGPU::SI_RETURN || 1030 MI.getOpcode() == AMDGPU::S_SETPC_B64_return || 1031 (MI.isReturn() && MI.isCall() && !callWaitsOnFunctionEntry(MI))) { 1032 Wait = Wait.combined(allZeroWaitcnt()); 1033 } 1034 // Resolve vm waits before gs-done. 1035 else if ((MI.getOpcode() == AMDGPU::S_SENDMSG || 1036 MI.getOpcode() == AMDGPU::S_SENDMSGHALT) && 1037 ST->hasLegacyGeometry() && 1038 ((MI.getOperand(0).getImm() & AMDGPU::SendMsg::ID_MASK_PreGFX11_) == 1039 AMDGPU::SendMsg::ID_GS_DONE_PreGFX11)) { 1040 Wait.VmCnt = 0; 1041 } 1042 #if 0 // TODO: the following blocks of logic when we have fence. 1043 else if (MI.getOpcode() == SC_FENCE) { 1044 const unsigned int group_size = 1045 context->shader_info->GetMaxThreadGroupSize(); 1046 // group_size == 0 means thread group size is unknown at compile time 1047 const bool group_is_multi_wave = 1048 (group_size == 0 || group_size > target_info->GetWaveFrontSize()); 1049 const bool fence_is_global = !((SCInstInternalMisc*)Inst)->IsGroupFence(); 1050 1051 for (unsigned int i = 0; i < Inst->NumSrcOperands(); i++) { 1052 SCRegType src_type = Inst->GetSrcType(i); 1053 switch (src_type) { 1054 case SCMEM_LDS: 1055 if (group_is_multi_wave || 1056 context->OptFlagIsOn(OPT_R1100_LDSMEM_FENCE_CHICKEN_BIT)) { 1057 EmitWaitcnt |= ScoreBrackets->updateByWait(LGKM_CNT, 1058 ScoreBrackets->getScoreUB(LGKM_CNT)); 1059 // LDS may have to wait for VM_CNT after buffer load to LDS 1060 if (target_info->HasBufferLoadToLDS()) { 1061 EmitWaitcnt |= ScoreBrackets->updateByWait(VM_CNT, 1062 ScoreBrackets->getScoreUB(VM_CNT)); 1063 } 1064 } 1065 break; 1066 1067 case SCMEM_GDS: 1068 if (group_is_multi_wave || fence_is_global) { 1069 EmitWaitcnt |= ScoreBrackets->updateByWait(EXP_CNT, 1070 ScoreBrackets->getScoreUB(EXP_CNT)); 1071 EmitWaitcnt |= ScoreBrackets->updateByWait(LGKM_CNT, 1072 ScoreBrackets->getScoreUB(LGKM_CNT)); 1073 } 1074 break; 1075 1076 case SCMEM_UAV: 1077 case SCMEM_TFBUF: 1078 case SCMEM_RING: 1079 case SCMEM_SCATTER: 1080 if (group_is_multi_wave || fence_is_global) { 1081 EmitWaitcnt |= ScoreBrackets->updateByWait(EXP_CNT, 1082 ScoreBrackets->getScoreUB(EXP_CNT)); 1083 EmitWaitcnt |= ScoreBrackets->updateByWait(VM_CNT, 1084 ScoreBrackets->getScoreUB(VM_CNT)); 1085 } 1086 break; 1087 1088 case SCMEM_SCRATCH: 1089 default: 1090 break; 1091 } 1092 } 1093 } 1094 #endif 1095 1096 // Export & GDS instructions do not read the EXEC mask until after the export 1097 // is granted (which can occur well after the instruction is issued). 1098 // The shader program must flush all EXP operations on the export-count 1099 // before overwriting the EXEC mask. 1100 else { 1101 if (MI.modifiesRegister(AMDGPU::EXEC, TRI)) { 1102 // Export and GDS are tracked individually, either may trigger a waitcnt 1103 // for EXEC. 1104 if (ScoreBrackets.hasPendingEvent(EXP_GPR_LOCK) || 1105 ScoreBrackets.hasPendingEvent(EXP_PARAM_ACCESS) || 1106 ScoreBrackets.hasPendingEvent(EXP_POS_ACCESS) || 1107 ScoreBrackets.hasPendingEvent(GDS_GPR_LOCK)) { 1108 Wait.ExpCnt = 0; 1109 } 1110 } 1111 1112 if (MI.isCall() && callWaitsOnFunctionEntry(MI)) { 1113 // The function is going to insert a wait on everything in its prolog. 1114 // This still needs to be careful if the call target is a load (e.g. a GOT 1115 // load). We also need to check WAW dependency with saved PC. 1116 Wait = AMDGPU::Waitcnt(); 1117 1118 int CallAddrOpIdx = 1119 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::src0); 1120 1121 if (MI.getOperand(CallAddrOpIdx).isReg()) { 1122 RegInterval CallAddrOpInterval = 1123 ScoreBrackets.getRegInterval(&MI, TII, MRI, TRI, CallAddrOpIdx); 1124 1125 for (int RegNo = CallAddrOpInterval.first; 1126 RegNo < CallAddrOpInterval.second; ++RegNo) 1127 ScoreBrackets.determineWait(LGKM_CNT, RegNo, Wait); 1128 1129 int RtnAddrOpIdx = 1130 AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::dst); 1131 if (RtnAddrOpIdx != -1) { 1132 RegInterval RtnAddrOpInterval = 1133 ScoreBrackets.getRegInterval(&MI, TII, MRI, TRI, RtnAddrOpIdx); 1134 1135 for (int RegNo = RtnAddrOpInterval.first; 1136 RegNo < RtnAddrOpInterval.second; ++RegNo) 1137 ScoreBrackets.determineWait(LGKM_CNT, RegNo, Wait); 1138 } 1139 } 1140 } else { 1141 // FIXME: Should not be relying on memoperands. 1142 // Look at the source operands of every instruction to see if 1143 // any of them results from a previous memory operation that affects 1144 // its current usage. If so, an s_waitcnt instruction needs to be 1145 // emitted. 1146 // If the source operand was defined by a load, add the s_waitcnt 1147 // instruction. 1148 // 1149 // Two cases are handled for destination operands: 1150 // 1) If the destination operand was defined by a load, add the s_waitcnt 1151 // instruction to guarantee the right WAW order. 1152 // 2) If a destination operand that was used by a recent export/store ins, 1153 // add s_waitcnt on exp_cnt to guarantee the WAR order. 1154 for (const MachineMemOperand *Memop : MI.memoperands()) { 1155 const Value *Ptr = Memop->getValue(); 1156 if (Memop->isStore() && SLoadAddresses.count(Ptr)) { 1157 addWait(Wait, LGKM_CNT, 0); 1158 if (PDT->dominates(MI.getParent(), SLoadAddresses.find(Ptr)->second)) 1159 SLoadAddresses.erase(Ptr); 1160 } 1161 unsigned AS = Memop->getAddrSpace(); 1162 if (AS != AMDGPUAS::LOCAL_ADDRESS && AS != AMDGPUAS::FLAT_ADDRESS) 1163 continue; 1164 // No need to wait before load from VMEM to LDS. 1165 if (mayWriteLDSThroughDMA(MI)) 1166 continue; 1167 unsigned RegNo = SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS; 1168 // VM_CNT is only relevant to vgpr or LDS. 1169 ScoreBrackets.determineWait(VM_CNT, RegNo, Wait); 1170 if (Memop->isStore()) { 1171 ScoreBrackets.determineWait(EXP_CNT, RegNo, Wait); 1172 } 1173 } 1174 1175 // Loop over use and def operands. 1176 for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) { 1177 MachineOperand &Op = MI.getOperand(I); 1178 if (!Op.isReg()) 1179 continue; 1180 1181 // If the instruction does not read tied source, skip the operand. 1182 if (Op.isTied() && Op.isUse() && TII->doesNotReadTiedSource(MI)) 1183 continue; 1184 1185 RegInterval Interval = 1186 ScoreBrackets.getRegInterval(&MI, TII, MRI, TRI, I); 1187 1188 const bool IsVGPR = TRI->isVectorRegister(*MRI, Op.getReg()); 1189 for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo) { 1190 if (IsVGPR) { 1191 // RAW always needs an s_waitcnt. WAW needs an s_waitcnt unless the 1192 // previous write and this write are the same type of VMEM 1193 // instruction, in which case they're guaranteed to write their 1194 // results in order anyway. 1195 if (Op.isUse() || !updateVMCntOnly(MI) || 1196 ScoreBrackets.hasOtherPendingVmemTypes(RegNo, 1197 getVmemType(MI))) { 1198 ScoreBrackets.determineWait(VM_CNT, RegNo, Wait); 1199 ScoreBrackets.clearVgprVmemTypes(RegNo); 1200 } 1201 if (Op.isDef() || ScoreBrackets.hasPendingEvent(EXP_LDS_ACCESS)) { 1202 ScoreBrackets.determineWait(EXP_CNT, RegNo, Wait); 1203 } 1204 } 1205 ScoreBrackets.determineWait(LGKM_CNT, RegNo, Wait); 1206 } 1207 } 1208 } 1209 } 1210 1211 // The subtarget may have an implicit S_WAITCNT 0 before barriers. If it does 1212 // not, we need to ensure the subtarget is capable of backing off barrier 1213 // instructions in case there are any outstanding memory operations that may 1214 // cause an exception. Otherwise, insert an explicit S_WAITCNT 0 here. 1215 if (MI.getOpcode() == AMDGPU::S_BARRIER && 1216 !ST->hasAutoWaitcntBeforeBarrier() && !ST->supportsBackOffBarrier()) { 1217 Wait = Wait.combined(allZeroWaitcnt()); 1218 } 1219 1220 // TODO: Remove this work-around, enable the assert for Bug 457939 1221 // after fixing the scheduler. Also, the Shader Compiler code is 1222 // independent of target. 1223 if (readsVCCZ(MI) && ST->hasReadVCCZBug()) { 1224 if (ScoreBrackets.hasPendingEvent(SMEM_ACCESS)) { 1225 Wait.LgkmCnt = 0; 1226 } 1227 } 1228 1229 // Verify that the wait is actually needed. 1230 ScoreBrackets.simplifyWaitcnt(Wait); 1231 1232 if (ForceEmitZeroWaitcnts) 1233 Wait = allZeroWaitcnt(); 1234 1235 if (ForceEmitWaitcnt[VM_CNT]) 1236 Wait.VmCnt = 0; 1237 if (ForceEmitWaitcnt[EXP_CNT]) 1238 Wait.ExpCnt = 0; 1239 if (ForceEmitWaitcnt[LGKM_CNT]) 1240 Wait.LgkmCnt = 0; 1241 if (ForceEmitWaitcnt[VS_CNT]) 1242 Wait.VsCnt = 0; 1243 1244 if (FlushVmCnt) { 1245 if (ScoreBrackets.hasPendingEvent(VM_CNT)) 1246 Wait.VmCnt = 0; 1247 } 1248 1249 return generateWaitcnt(Wait, MI.getIterator(), *MI.getParent(), ScoreBrackets, 1250 OldWaitcntInstr); 1251 } 1252 1253 // Add a waitcnt to flush the vmcnt counter at the end of the given block if 1254 // needed. 1255 bool SIInsertWaitcnts::generateWaitcntBlockEnd(MachineBasicBlock &Block, 1256 WaitcntBrackets &ScoreBrackets, 1257 MachineInstr *OldWaitcntInstr) { 1258 AMDGPU::Waitcnt Wait; 1259 1260 if (!ScoreBrackets.hasPendingEvent(VM_CNT)) 1261 return false; 1262 1263 Wait.VmCnt = 0; 1264 1265 return generateWaitcnt(Wait, Block.instr_end(), Block, ScoreBrackets, 1266 OldWaitcntInstr); 1267 } 1268 1269 bool SIInsertWaitcnts::generateWaitcnt(AMDGPU::Waitcnt Wait, 1270 MachineBasicBlock::instr_iterator It, 1271 MachineBasicBlock &Block, 1272 WaitcntBrackets &ScoreBrackets, 1273 MachineInstr *OldWaitcntInstr) { 1274 bool Modified = false; 1275 const DebugLoc &DL = Block.findDebugLoc(It); 1276 1277 if (OldWaitcntInstr) 1278 // Try to merge the required wait with preexisting waitcnt instructions. 1279 // Also erase redundant waitcnt. 1280 Modified = 1281 applyPreexistingWaitcnt(ScoreBrackets, *OldWaitcntInstr, Wait, It); 1282 else 1283 ScoreBrackets.applyWaitcnt(Wait); 1284 1285 // ExpCnt can be merged into VINTERP. 1286 if (Wait.ExpCnt != ~0u && It != Block.instr_end() && 1287 SIInstrInfo::isVINTERP(*It)) { 1288 MachineOperand *WaitExp = 1289 TII->getNamedOperand(*It, AMDGPU::OpName::waitexp); 1290 if (Wait.ExpCnt < WaitExp->getImm()) { 1291 WaitExp->setImm(Wait.ExpCnt); 1292 Modified = true; 1293 } 1294 Wait.ExpCnt = ~0u; 1295 1296 LLVM_DEBUG(dbgs() << "generateWaitcntInstBefore\n" 1297 << "Update Instr: " << *It); 1298 } 1299 1300 // Build new waitcnt instructions unless no wait is needed or the old waitcnt 1301 // instruction was modified to handle the required wait. 1302 if (Wait.hasWaitExceptVsCnt()) { 1303 unsigned Enc = AMDGPU::encodeWaitcnt(IV, Wait); 1304 auto SWaitInst = 1305 BuildMI(Block, It, DL, TII->get(AMDGPU::S_WAITCNT)).addImm(Enc); 1306 TrackedWaitcntSet.insert(SWaitInst); 1307 Modified = true; 1308 1309 LLVM_DEBUG(dbgs() << "generateWaitcnt\n"; 1310 if (It != Block.instr_end()) dbgs() << "Old Instr: " << *It; 1311 dbgs() << "New Instr: " << *SWaitInst << '\n'); 1312 } 1313 1314 if (Wait.hasWaitVsCnt()) { 1315 assert(ST->hasVscnt()); 1316 1317 auto SWaitInst = BuildMI(Block, It, DL, TII->get(AMDGPU::S_WAITCNT_VSCNT)) 1318 .addReg(AMDGPU::SGPR_NULL, RegState::Undef) 1319 .addImm(Wait.VsCnt); 1320 TrackedWaitcntSet.insert(SWaitInst); 1321 Modified = true; 1322 1323 LLVM_DEBUG(dbgs() << "generateWaitcnt\n"; 1324 if (It != Block.instr_end()) dbgs() << "Old Instr: " << *It; 1325 dbgs() << "New Instr: " << *SWaitInst << '\n'); 1326 } 1327 return Modified; 1328 } 1329 1330 // This is a flat memory operation. Check to see if it has memory tokens other 1331 // than LDS. Other address spaces supported by flat memory operations involve 1332 // global memory. 1333 bool SIInsertWaitcnts::mayAccessVMEMThroughFlat(const MachineInstr &MI) const { 1334 assert(TII->isFLAT(MI)); 1335 1336 // All flat instructions use the VMEM counter. 1337 assert(TII->usesVM_CNT(MI)); 1338 1339 // If there are no memory operands then conservatively assume the flat 1340 // operation may access VMEM. 1341 if (MI.memoperands_empty()) 1342 return true; 1343 1344 // See if any memory operand specifies an address space that involves VMEM. 1345 // Flat operations only supported FLAT, LOCAL (LDS), or address spaces 1346 // involving VMEM such as GLOBAL, CONSTANT, PRIVATE (SCRATCH), etc. The REGION 1347 // (GDS) address space is not supported by flat operations. Therefore, simply 1348 // return true unless only the LDS address space is found. 1349 for (const MachineMemOperand *Memop : MI.memoperands()) { 1350 unsigned AS = Memop->getAddrSpace(); 1351 assert(AS != AMDGPUAS::REGION_ADDRESS); 1352 if (AS != AMDGPUAS::LOCAL_ADDRESS) 1353 return true; 1354 } 1355 1356 return false; 1357 } 1358 1359 // This is a flat memory operation. Check to see if it has memory tokens for 1360 // either LDS or FLAT. 1361 bool SIInsertWaitcnts::mayAccessLDSThroughFlat(const MachineInstr &MI) const { 1362 assert(TII->isFLAT(MI)); 1363 1364 // Flat instruction such as SCRATCH and GLOBAL do not use the lgkm counter. 1365 if (!TII->usesLGKM_CNT(MI)) 1366 return false; 1367 1368 // If in tgsplit mode then there can be no use of LDS. 1369 if (ST->isTgSplitEnabled()) 1370 return false; 1371 1372 // If there are no memory operands then conservatively assume the flat 1373 // operation may access LDS. 1374 if (MI.memoperands_empty()) 1375 return true; 1376 1377 // See if any memory operand specifies an address space that involves LDS. 1378 for (const MachineMemOperand *Memop : MI.memoperands()) { 1379 unsigned AS = Memop->getAddrSpace(); 1380 if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS) 1381 return true; 1382 } 1383 1384 return false; 1385 } 1386 1387 void SIInsertWaitcnts::updateEventWaitcntAfter(MachineInstr &Inst, 1388 WaitcntBrackets *ScoreBrackets) { 1389 // Now look at the instruction opcode. If it is a memory access 1390 // instruction, update the upper-bound of the appropriate counter's 1391 // bracket and the destination operand scores. 1392 // TODO: Use the (TSFlags & SIInstrFlags::LGKM_CNT) property everywhere. 1393 if (TII->isDS(Inst) && TII->usesLGKM_CNT(Inst)) { 1394 if (TII->isAlwaysGDS(Inst.getOpcode()) || 1395 TII->hasModifiersSet(Inst, AMDGPU::OpName::gds)) { 1396 ScoreBrackets->updateByEvent(TII, TRI, MRI, GDS_ACCESS, Inst); 1397 ScoreBrackets->updateByEvent(TII, TRI, MRI, GDS_GPR_LOCK, Inst); 1398 } else { 1399 ScoreBrackets->updateByEvent(TII, TRI, MRI, LDS_ACCESS, Inst); 1400 } 1401 } else if (TII->isFLAT(Inst)) { 1402 assert(Inst.mayLoadOrStore()); 1403 1404 int FlatASCount = 0; 1405 1406 if (mayAccessVMEMThroughFlat(Inst)) { 1407 ++FlatASCount; 1408 ScoreBrackets->updateByEvent(TII, TRI, MRI, getVmemWaitEventType(Inst), 1409 Inst); 1410 } 1411 1412 if (mayAccessLDSThroughFlat(Inst)) { 1413 ++FlatASCount; 1414 ScoreBrackets->updateByEvent(TII, TRI, MRI, LDS_ACCESS, Inst); 1415 } 1416 1417 // A Flat memory operation must access at least one address space. 1418 assert(FlatASCount); 1419 1420 // This is a flat memory operation that access both VMEM and LDS, so note it 1421 // - it will require that both the VM and LGKM be flushed to zero if it is 1422 // pending when a VM or LGKM dependency occurs. 1423 if (FlatASCount > 1) 1424 ScoreBrackets->setPendingFlat(); 1425 } else if (SIInstrInfo::isVMEM(Inst) && 1426 !llvm::AMDGPU::getMUBUFIsBufferInv(Inst.getOpcode())) { 1427 ScoreBrackets->updateByEvent(TII, TRI, MRI, getVmemWaitEventType(Inst), 1428 Inst); 1429 1430 if (ST->vmemWriteNeedsExpWaitcnt() && 1431 (Inst.mayStore() || SIInstrInfo::isAtomicRet(Inst))) { 1432 ScoreBrackets->updateByEvent(TII, TRI, MRI, VMW_GPR_LOCK, Inst); 1433 } 1434 } else if (TII->isSMRD(Inst)) { 1435 ScoreBrackets->updateByEvent(TII, TRI, MRI, SMEM_ACCESS, Inst); 1436 } else if (Inst.isCall()) { 1437 if (callWaitsOnFunctionReturn(Inst)) { 1438 // Act as a wait on everything 1439 ScoreBrackets->applyWaitcnt(allZeroWaitcnt()); 1440 } else { 1441 // May need to way wait for anything. 1442 ScoreBrackets->applyWaitcnt(AMDGPU::Waitcnt()); 1443 } 1444 } else if (SIInstrInfo::isLDSDIR(Inst)) { 1445 ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_LDS_ACCESS, Inst); 1446 } else if (TII->isVINTERP(Inst)) { 1447 int64_t Imm = TII->getNamedOperand(Inst, AMDGPU::OpName::waitexp)->getImm(); 1448 ScoreBrackets->applyWaitcnt(EXP_CNT, Imm); 1449 } else if (SIInstrInfo::isEXP(Inst)) { 1450 unsigned Imm = TII->getNamedOperand(Inst, AMDGPU::OpName::tgt)->getImm(); 1451 if (Imm >= AMDGPU::Exp::ET_PARAM0 && Imm <= AMDGPU::Exp::ET_PARAM31) 1452 ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_PARAM_ACCESS, Inst); 1453 else if (Imm >= AMDGPU::Exp::ET_POS0 && Imm <= AMDGPU::Exp::ET_POS_LAST) 1454 ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_POS_ACCESS, Inst); 1455 else 1456 ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_GPR_LOCK, Inst); 1457 } else { 1458 switch (Inst.getOpcode()) { 1459 case AMDGPU::S_SENDMSG: 1460 case AMDGPU::S_SENDMSG_RTN_B32: 1461 case AMDGPU::S_SENDMSG_RTN_B64: 1462 case AMDGPU::S_SENDMSGHALT: 1463 ScoreBrackets->updateByEvent(TII, TRI, MRI, SQ_MESSAGE, Inst); 1464 break; 1465 case AMDGPU::S_MEMTIME: 1466 case AMDGPU::S_MEMREALTIME: 1467 ScoreBrackets->updateByEvent(TII, TRI, MRI, SMEM_ACCESS, Inst); 1468 break; 1469 } 1470 } 1471 } 1472 1473 bool WaitcntBrackets::mergeScore(const MergeInfo &M, unsigned &Score, 1474 unsigned OtherScore) { 1475 unsigned MyShifted = Score <= M.OldLB ? 0 : Score + M.MyShift; 1476 unsigned OtherShifted = 1477 OtherScore <= M.OtherLB ? 0 : OtherScore + M.OtherShift; 1478 Score = std::max(MyShifted, OtherShifted); 1479 return OtherShifted > MyShifted; 1480 } 1481 1482 /// Merge the pending events and associater score brackets of \p Other into 1483 /// this brackets status. 1484 /// 1485 /// Returns whether the merge resulted in a change that requires tighter waits 1486 /// (i.e. the merged brackets strictly dominate the original brackets). 1487 bool WaitcntBrackets::merge(const WaitcntBrackets &Other) { 1488 bool StrictDom = false; 1489 1490 VgprUB = std::max(VgprUB, Other.VgprUB); 1491 SgprUB = std::max(SgprUB, Other.SgprUB); 1492 1493 for (auto T : inst_counter_types()) { 1494 // Merge event flags for this counter 1495 const unsigned OldEvents = PendingEvents & WaitEventMaskForInst[T]; 1496 const unsigned OtherEvents = Other.PendingEvents & WaitEventMaskForInst[T]; 1497 if (OtherEvents & ~OldEvents) 1498 StrictDom = true; 1499 PendingEvents |= OtherEvents; 1500 1501 // Merge scores for this counter 1502 const unsigned MyPending = ScoreUBs[T] - ScoreLBs[T]; 1503 const unsigned OtherPending = Other.ScoreUBs[T] - Other.ScoreLBs[T]; 1504 const unsigned NewUB = ScoreLBs[T] + std::max(MyPending, OtherPending); 1505 if (NewUB < ScoreLBs[T]) 1506 report_fatal_error("waitcnt score overflow"); 1507 1508 MergeInfo M; 1509 M.OldLB = ScoreLBs[T]; 1510 M.OtherLB = Other.ScoreLBs[T]; 1511 M.MyShift = NewUB - ScoreUBs[T]; 1512 M.OtherShift = NewUB - Other.ScoreUBs[T]; 1513 1514 ScoreUBs[T] = NewUB; 1515 1516 StrictDom |= mergeScore(M, LastFlat[T], Other.LastFlat[T]); 1517 1518 for (int J = 0; J <= VgprUB; J++) 1519 StrictDom |= mergeScore(M, VgprScores[T][J], Other.VgprScores[T][J]); 1520 1521 if (T == LGKM_CNT) { 1522 for (int J = 0; J <= SgprUB; J++) 1523 StrictDom |= mergeScore(M, SgprScores[J], Other.SgprScores[J]); 1524 } 1525 } 1526 1527 for (int J = 0; J <= VgprUB; J++) { 1528 unsigned char NewVmemTypes = VgprVmemTypes[J] | Other.VgprVmemTypes[J]; 1529 StrictDom |= NewVmemTypes != VgprVmemTypes[J]; 1530 VgprVmemTypes[J] = NewVmemTypes; 1531 } 1532 1533 return StrictDom; 1534 } 1535 1536 static bool isWaitInstr(MachineInstr &Inst) { 1537 return Inst.getOpcode() == AMDGPU::S_WAITCNT || 1538 (Inst.getOpcode() == AMDGPU::S_WAITCNT_VSCNT && 1539 Inst.getOperand(0).isReg() && 1540 Inst.getOperand(0).getReg() == AMDGPU::SGPR_NULL); 1541 } 1542 1543 // Generate s_waitcnt instructions where needed. 1544 bool SIInsertWaitcnts::insertWaitcntInBlock(MachineFunction &MF, 1545 MachineBasicBlock &Block, 1546 WaitcntBrackets &ScoreBrackets) { 1547 bool Modified = false; 1548 1549 LLVM_DEBUG({ 1550 dbgs() << "*** Block" << Block.getNumber() << " ***"; 1551 ScoreBrackets.dump(); 1552 }); 1553 1554 // Track the correctness of vccz through this basic block. There are two 1555 // reasons why it might be incorrect; see ST->hasReadVCCZBug() and 1556 // ST->partialVCCWritesUpdateVCCZ(). 1557 bool VCCZCorrect = true; 1558 if (ST->hasReadVCCZBug()) { 1559 // vccz could be incorrect at a basic block boundary if a predecessor wrote 1560 // to vcc and then issued an smem load. 1561 VCCZCorrect = false; 1562 } else if (!ST->partialVCCWritesUpdateVCCZ()) { 1563 // vccz could be incorrect at a basic block boundary if a predecessor wrote 1564 // to vcc_lo or vcc_hi. 1565 VCCZCorrect = false; 1566 } 1567 1568 // Walk over the instructions. 1569 MachineInstr *OldWaitcntInstr = nullptr; 1570 1571 for (MachineBasicBlock::instr_iterator Iter = Block.instr_begin(), 1572 E = Block.instr_end(); 1573 Iter != E;) { 1574 MachineInstr &Inst = *Iter; 1575 1576 // Track pre-existing waitcnts that were added in earlier iterations or by 1577 // the memory legalizer. 1578 if (isWaitInstr(Inst)) { 1579 if (!OldWaitcntInstr) 1580 OldWaitcntInstr = &Inst; 1581 ++Iter; 1582 continue; 1583 } 1584 1585 bool FlushVmCnt = Block.getFirstTerminator() == Inst && 1586 isPreheaderToFlush(Block, ScoreBrackets); 1587 1588 // Generate an s_waitcnt instruction to be placed before Inst, if needed. 1589 Modified |= generateWaitcntInstBefore(Inst, ScoreBrackets, OldWaitcntInstr, 1590 FlushVmCnt); 1591 OldWaitcntInstr = nullptr; 1592 1593 // Restore vccz if it's not known to be correct already. 1594 bool RestoreVCCZ = !VCCZCorrect && readsVCCZ(Inst); 1595 1596 // Don't examine operands unless we need to track vccz correctness. 1597 if (ST->hasReadVCCZBug() || !ST->partialVCCWritesUpdateVCCZ()) { 1598 if (Inst.definesRegister(AMDGPU::VCC_LO) || 1599 Inst.definesRegister(AMDGPU::VCC_HI)) { 1600 // Up to gfx9, writes to vcc_lo and vcc_hi don't update vccz. 1601 if (!ST->partialVCCWritesUpdateVCCZ()) 1602 VCCZCorrect = false; 1603 } else if (Inst.definesRegister(AMDGPU::VCC)) { 1604 // There is a hardware bug on CI/SI where SMRD instruction may corrupt 1605 // vccz bit, so when we detect that an instruction may read from a 1606 // corrupt vccz bit, we need to: 1607 // 1. Insert s_waitcnt lgkm(0) to wait for all outstanding SMRD 1608 // operations to complete. 1609 // 2. Restore the correct value of vccz by writing the current value 1610 // of vcc back to vcc. 1611 if (ST->hasReadVCCZBug() && 1612 ScoreBrackets.hasPendingEvent(SMEM_ACCESS)) { 1613 // Writes to vcc while there's an outstanding smem read may get 1614 // clobbered as soon as any read completes. 1615 VCCZCorrect = false; 1616 } else { 1617 // Writes to vcc will fix any incorrect value in vccz. 1618 VCCZCorrect = true; 1619 } 1620 } 1621 } 1622 1623 if (TII->isSMRD(Inst)) { 1624 for (const MachineMemOperand *Memop : Inst.memoperands()) { 1625 // No need to handle invariant loads when avoiding WAR conflicts, as 1626 // there cannot be a vector store to the same memory location. 1627 if (!Memop->isInvariant()) { 1628 const Value *Ptr = Memop->getValue(); 1629 SLoadAddresses.insert(std::pair(Ptr, Inst.getParent())); 1630 } 1631 } 1632 if (ST->hasReadVCCZBug()) { 1633 // This smem read could complete and clobber vccz at any time. 1634 VCCZCorrect = false; 1635 } 1636 } 1637 1638 updateEventWaitcntAfter(Inst, &ScoreBrackets); 1639 1640 #if 0 // TODO: implement resource type check controlled by options with ub = LB. 1641 // If this instruction generates a S_SETVSKIP because it is an 1642 // indexed resource, and we are on Tahiti, then it will also force 1643 // an S_WAITCNT vmcnt(0) 1644 if (RequireCheckResourceType(Inst, context)) { 1645 // Force the score to as if an S_WAITCNT vmcnt(0) is emitted. 1646 ScoreBrackets->setScoreLB(VM_CNT, 1647 ScoreBrackets->getScoreUB(VM_CNT)); 1648 } 1649 #endif 1650 1651 LLVM_DEBUG({ 1652 Inst.print(dbgs()); 1653 ScoreBrackets.dump(); 1654 }); 1655 1656 // TODO: Remove this work-around after fixing the scheduler and enable the 1657 // assert above. 1658 if (RestoreVCCZ) { 1659 // Restore the vccz bit. Any time a value is written to vcc, the vcc 1660 // bit is updated, so we can restore the bit by reading the value of 1661 // vcc and then writing it back to the register. 1662 BuildMI(Block, Inst, Inst.getDebugLoc(), 1663 TII->get(ST->isWave32() ? AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64), 1664 TRI->getVCC()) 1665 .addReg(TRI->getVCC()); 1666 VCCZCorrect = true; 1667 Modified = true; 1668 } 1669 1670 ++Iter; 1671 } 1672 1673 if (Block.getFirstTerminator() == Block.end() && 1674 isPreheaderToFlush(Block, ScoreBrackets)) 1675 Modified |= generateWaitcntBlockEnd(Block, ScoreBrackets, OldWaitcntInstr); 1676 1677 return Modified; 1678 } 1679 1680 // Return true if the given machine basic block is a preheader of a loop in 1681 // which we want to flush the vmcnt counter, and false otherwise. 1682 bool SIInsertWaitcnts::isPreheaderToFlush(MachineBasicBlock &MBB, 1683 WaitcntBrackets &ScoreBrackets) { 1684 if (PreheadersToFlush.count(&MBB)) 1685 return PreheadersToFlush[&MBB]; 1686 1687 auto UpdateCache = [&](bool val) { 1688 PreheadersToFlush[&MBB] = val; 1689 return val; 1690 }; 1691 1692 MachineBasicBlock *Succ = MBB.getSingleSuccessor(); 1693 if (!Succ) 1694 return UpdateCache(false); 1695 1696 MachineLoop *Loop = MLI->getLoopFor(Succ); 1697 if (!Loop) 1698 return UpdateCache(false); 1699 1700 if (Loop->getLoopPreheader() == &MBB && shouldFlushVmCnt(Loop, ScoreBrackets)) 1701 return UpdateCache(true); 1702 1703 return UpdateCache(false); 1704 } 1705 1706 // Return true if it is better to flush the vmcnt counter in the preheader of 1707 // the given loop. We currently decide to flush in two situations: 1708 // 1. The loop contains vmem store(s), no vmem load and at least one use of a 1709 // vgpr containing a value that is loaded outside of the loop. (Only on 1710 // targets with no vscnt counter). 1711 // 2. The loop contains vmem load(s), but the loaded values are not used in the 1712 // loop, and at least one use of a vgpr containing a value that is loaded 1713 // outside of the loop. 1714 bool SIInsertWaitcnts::shouldFlushVmCnt(MachineLoop *ML, 1715 WaitcntBrackets &Brackets) { 1716 bool HasVMemLoad = false; 1717 bool HasVMemStore = false; 1718 bool UsesVgprLoadedOutside = false; 1719 DenseSet<Register> VgprUse; 1720 DenseSet<Register> VgprDef; 1721 1722 for (MachineBasicBlock *MBB : ML->blocks()) { 1723 for (MachineInstr &MI : *MBB) { 1724 if (SIInstrInfo::isVMEM(MI)) { 1725 if (MI.mayLoad()) 1726 HasVMemLoad = true; 1727 if (MI.mayStore()) 1728 HasVMemStore = true; 1729 } 1730 for (unsigned I = 0; I < MI.getNumOperands(); I++) { 1731 MachineOperand &Op = MI.getOperand(I); 1732 if (!Op.isReg() || !TRI->isVectorRegister(*MRI, Op.getReg())) 1733 continue; 1734 RegInterval Interval = Brackets.getRegInterval(&MI, TII, MRI, TRI, I); 1735 // Vgpr use 1736 if (Op.isUse()) { 1737 for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo) { 1738 // If we find a register that is loaded inside the loop, 1. and 2. 1739 // are invalidated and we can exit. 1740 if (VgprDef.contains(RegNo)) 1741 return false; 1742 VgprUse.insert(RegNo); 1743 // If at least one of Op's registers is in the score brackets, the 1744 // value is likely loaded outside of the loop. 1745 if (Brackets.getRegScore(RegNo, VM_CNT) > Brackets.getScoreLB(VM_CNT)) { 1746 UsesVgprLoadedOutside = true; 1747 break; 1748 } 1749 } 1750 } 1751 // VMem load vgpr def 1752 else if (SIInstrInfo::isVMEM(MI) && MI.mayLoad() && Op.isDef()) 1753 for (int RegNo = Interval.first; RegNo < Interval.second; ++RegNo) { 1754 // If we find a register that is loaded inside the loop, 1. and 2. 1755 // are invalidated and we can exit. 1756 if (VgprUse.contains(RegNo)) 1757 return false; 1758 VgprDef.insert(RegNo); 1759 } 1760 } 1761 } 1762 } 1763 if (!ST->hasVscnt() && HasVMemStore && !HasVMemLoad && UsesVgprLoadedOutside) 1764 return true; 1765 return HasVMemLoad && UsesVgprLoadedOutside; 1766 } 1767 1768 bool SIInsertWaitcnts::runOnMachineFunction(MachineFunction &MF) { 1769 ST = &MF.getSubtarget<GCNSubtarget>(); 1770 TII = ST->getInstrInfo(); 1771 TRI = &TII->getRegisterInfo(); 1772 MRI = &MF.getRegInfo(); 1773 IV = AMDGPU::getIsaVersion(ST->getCPU()); 1774 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>(); 1775 MLI = &getAnalysis<MachineLoopInfo>(); 1776 PDT = &getAnalysis<MachinePostDominatorTree>(); 1777 1778 ForceEmitZeroWaitcnts = ForceEmitZeroFlag; 1779 for (auto T : inst_counter_types()) 1780 ForceEmitWaitcnt[T] = false; 1781 1782 HardwareLimits Limits = {}; 1783 Limits.VmcntMax = AMDGPU::getVmcntBitMask(IV); 1784 Limits.ExpcntMax = AMDGPU::getExpcntBitMask(IV); 1785 Limits.LgkmcntMax = AMDGPU::getLgkmcntBitMask(IV); 1786 Limits.VscntMax = ST->hasVscnt() ? 63 : 0; 1787 1788 unsigned NumVGPRsMax = ST->getAddressableNumVGPRs(); 1789 unsigned NumSGPRsMax = ST->getAddressableNumSGPRs(); 1790 assert(NumVGPRsMax <= SQ_MAX_PGM_VGPRS); 1791 assert(NumSGPRsMax <= SQ_MAX_PGM_SGPRS); 1792 1793 RegisterEncoding Encoding = {}; 1794 Encoding.VGPR0 = TRI->getEncodingValue(AMDGPU::VGPR0); 1795 Encoding.VGPRL = Encoding.VGPR0 + NumVGPRsMax - 1; 1796 Encoding.SGPR0 = TRI->getEncodingValue(AMDGPU::SGPR0); 1797 Encoding.SGPRL = Encoding.SGPR0 + NumSGPRsMax - 1; 1798 1799 TrackedWaitcntSet.clear(); 1800 BlockInfos.clear(); 1801 bool Modified = false; 1802 1803 if (!MFI->isEntryFunction()) { 1804 // Wait for any outstanding memory operations that the input registers may 1805 // depend on. We can't track them and it's better to do the wait after the 1806 // costly call sequence. 1807 1808 // TODO: Could insert earlier and schedule more liberally with operations 1809 // that only use caller preserved registers. 1810 MachineBasicBlock &EntryBB = MF.front(); 1811 MachineBasicBlock::iterator I = EntryBB.begin(); 1812 for (MachineBasicBlock::iterator E = EntryBB.end(); 1813 I != E && (I->isPHI() || I->isMetaInstruction()); ++I) 1814 ; 1815 BuildMI(EntryBB, I, DebugLoc(), TII->get(AMDGPU::S_WAITCNT)).addImm(0); 1816 if (ST->hasVscnt()) 1817 BuildMI(EntryBB, I, DebugLoc(), TII->get(AMDGPU::S_WAITCNT_VSCNT)) 1818 .addReg(AMDGPU::SGPR_NULL, RegState::Undef) 1819 .addImm(0); 1820 1821 Modified = true; 1822 } 1823 1824 // Keep iterating over the blocks in reverse post order, inserting and 1825 // updating s_waitcnt where needed, until a fix point is reached. 1826 for (auto *MBB : ReversePostOrderTraversal<MachineFunction *>(&MF)) 1827 BlockInfos.insert({MBB, BlockInfo(MBB)}); 1828 1829 std::unique_ptr<WaitcntBrackets> Brackets; 1830 bool Repeat; 1831 do { 1832 Repeat = false; 1833 1834 for (auto BII = BlockInfos.begin(), BIE = BlockInfos.end(); BII != BIE; 1835 ++BII) { 1836 BlockInfo &BI = BII->second; 1837 if (!BI.Dirty) 1838 continue; 1839 1840 if (BI.Incoming) { 1841 if (!Brackets) 1842 Brackets = std::make_unique<WaitcntBrackets>(*BI.Incoming); 1843 else 1844 *Brackets = *BI.Incoming; 1845 } else { 1846 if (!Brackets) 1847 Brackets = std::make_unique<WaitcntBrackets>(ST, Limits, Encoding); 1848 else 1849 *Brackets = WaitcntBrackets(ST, Limits, Encoding); 1850 } 1851 1852 Modified |= insertWaitcntInBlock(MF, *BI.MBB, *Brackets); 1853 BI.Dirty = false; 1854 1855 if (Brackets->hasPendingEvent()) { 1856 BlockInfo *MoveBracketsToSucc = nullptr; 1857 for (MachineBasicBlock *Succ : BI.MBB->successors()) { 1858 auto SuccBII = BlockInfos.find(Succ); 1859 BlockInfo &SuccBI = SuccBII->second; 1860 if (!SuccBI.Incoming) { 1861 SuccBI.Dirty = true; 1862 if (SuccBII <= BII) 1863 Repeat = true; 1864 if (!MoveBracketsToSucc) { 1865 MoveBracketsToSucc = &SuccBI; 1866 } else { 1867 SuccBI.Incoming = std::make_unique<WaitcntBrackets>(*Brackets); 1868 } 1869 } else if (SuccBI.Incoming->merge(*Brackets)) { 1870 SuccBI.Dirty = true; 1871 if (SuccBII <= BII) 1872 Repeat = true; 1873 } 1874 } 1875 if (MoveBracketsToSucc) 1876 MoveBracketsToSucc->Incoming = std::move(Brackets); 1877 } 1878 } 1879 } while (Repeat); 1880 1881 if (ST->hasScalarStores()) { 1882 SmallVector<MachineBasicBlock *, 4> EndPgmBlocks; 1883 bool HaveScalarStores = false; 1884 1885 for (MachineBasicBlock &MBB : MF) { 1886 for (MachineInstr &MI : MBB) { 1887 if (!HaveScalarStores && TII->isScalarStore(MI)) 1888 HaveScalarStores = true; 1889 1890 if (MI.getOpcode() == AMDGPU::S_ENDPGM || 1891 MI.getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG) 1892 EndPgmBlocks.push_back(&MBB); 1893 } 1894 } 1895 1896 if (HaveScalarStores) { 1897 // If scalar writes are used, the cache must be flushed or else the next 1898 // wave to reuse the same scratch memory can be clobbered. 1899 // 1900 // Insert s_dcache_wb at wave termination points if there were any scalar 1901 // stores, and only if the cache hasn't already been flushed. This could 1902 // be improved by looking across blocks for flushes in postdominating 1903 // blocks from the stores but an explicitly requested flush is probably 1904 // very rare. 1905 for (MachineBasicBlock *MBB : EndPgmBlocks) { 1906 bool SeenDCacheWB = false; 1907 1908 for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); 1909 I != E; ++I) { 1910 if (I->getOpcode() == AMDGPU::S_DCACHE_WB) 1911 SeenDCacheWB = true; 1912 else if (TII->isScalarStore(*I)) 1913 SeenDCacheWB = false; 1914 1915 // FIXME: It would be better to insert this before a waitcnt if any. 1916 if ((I->getOpcode() == AMDGPU::S_ENDPGM || 1917 I->getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG) && 1918 !SeenDCacheWB) { 1919 Modified = true; 1920 BuildMI(*MBB, I, I->getDebugLoc(), TII->get(AMDGPU::S_DCACHE_WB)); 1921 } 1922 } 1923 } 1924 } 1925 } 1926 1927 return Modified; 1928 } 1929