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