| 1 | //=-- SystemZHazardRecognizer.h - SystemZ Hazard Recognizer -----*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file defines a hazard recognizer for the SystemZ scheduler. |
| 10 | // |
| 11 | // This class is used by the SystemZ scheduling strategy to maintain |
| 12 | // the state during scheduling, and provide cost functions for |
| 13 | // scheduling candidates. This includes: |
| 14 | // |
| 15 | // * Decoder grouping. A decoder group can maximally hold 3 uops, and |
| 16 | // instructions that always begin a new group should be scheduled when |
| 17 | // the current decoder group is empty. |
| 18 | // * Processor resources usage. It is beneficial to balance the use of |
| 19 | // resources. |
| 20 | // |
| 21 | // A goal is to consider all instructions, also those outside of any |
| 22 | // scheduling region. Such instructions are "advanced" past and include |
| 23 | // single instructions before a scheduling region, branches etc. |
| 24 | // |
| 25 | // A block that has only one predecessor continues scheduling with the state |
| 26 | // of it (which may be updated by emitting branches). |
| 27 | // |
| 28 | // ===---------------------------------------------------------------------===// |
| 29 | |
| 30 | #include "SystemZHazardRecognizer.h" |
| 31 | #include "llvm/ADT/Statistic.h" |
| 32 | |
| 33 | using namespace llvm; |
| 34 | |
| 35 | #define DEBUG_TYPE "machine-scheduler" |
| 36 | |
| 37 | // This is the limit of processor resource usage at which the |
| 38 | // scheduler should try to look for other instructions (not using the |
| 39 | // critical resource). |
| 40 | static cl::opt<int> ProcResCostLim("procres-cost-lim" , cl::Hidden, |
| 41 | cl::desc("The OOO window for processor " |
| 42 | "resources during scheduling." ), |
| 43 | cl::init(Val: 8)); |
| 44 | |
| 45 | unsigned SystemZHazardRecognizer:: |
| 46 | getNumDecoderSlots(SUnit *SU) const { |
| 47 | const MCSchedClassDesc *SC = getSchedClass(SU); |
| 48 | if (!SC->isValid()) |
| 49 | return 0; // IMPLICIT_DEF / KILL -- will not make impact in output. |
| 50 | |
| 51 | assert((SC->NumMicroOps != 2 || (SC->BeginGroup && !SC->EndGroup)) && |
| 52 | "Only cracked instruction can have 2 uops." ); |
| 53 | assert((SC->NumMicroOps < 3 || (SC->BeginGroup && SC->EndGroup)) && |
| 54 | "Expanded instructions always group alone." ); |
| 55 | assert((SC->NumMicroOps < 3 || (SC->NumMicroOps % 3 == 0)) && |
| 56 | "Expanded instructions fill the group(s)." ); |
| 57 | |
| 58 | return SC->NumMicroOps; |
| 59 | } |
| 60 | |
| 61 | unsigned SystemZHazardRecognizer::getCurrCycleIdx(SUnit *SU) const { |
| 62 | unsigned Idx = CurrGroupSize; |
| 63 | if (GrpCount % 2) |
| 64 | Idx += 3; |
| 65 | |
| 66 | if (SU != nullptr && !fitsIntoCurrentGroup(SU)) { |
| 67 | if (Idx == 1 || Idx == 2) |
| 68 | Idx = 3; |
| 69 | else if (Idx == 4 || Idx == 5) |
| 70 | Idx = 0; |
| 71 | } |
| 72 | |
| 73 | return Idx; |
| 74 | } |
| 75 | |
| 76 | ScheduleHazardRecognizer::HazardType SystemZHazardRecognizer:: |
| 77 | getHazardType(SUnit *SU, int Stalls) { |
| 78 | return (fitsIntoCurrentGroup(SU) ? NoHazard : Hazard); |
| 79 | } |
| 80 | |
| 81 | void SystemZHazardRecognizer::Reset() { |
| 82 | CurrGroupSize = 0; |
| 83 | CurrGroupHas4RegOps = false; |
| 84 | clearProcResCounters(); |
| 85 | GrpCount = 0; |
| 86 | LastFPdOpCycleIdx = UINT_MAX; |
| 87 | LastEmittedMI = nullptr; |
| 88 | LLVM_DEBUG(CurGroupDbg = "" ;); |
| 89 | } |
| 90 | |
| 91 | bool |
| 92 | SystemZHazardRecognizer::fitsIntoCurrentGroup(SUnit *SU) const { |
| 93 | const MCSchedClassDesc *SC = getSchedClass(SU); |
| 94 | if (!SC->isValid()) |
| 95 | return true; |
| 96 | |
| 97 | // A cracked instruction only fits into schedule if the current |
| 98 | // group is empty. |
| 99 | if (SC->BeginGroup) |
| 100 | return (CurrGroupSize == 0); |
| 101 | |
| 102 | // An instruction with 4 register operands will not fit in last slot. |
| 103 | assert ((CurrGroupSize < 2 || !CurrGroupHas4RegOps) && |
| 104 | "Current decoder group is already full!" ); |
| 105 | if (CurrGroupSize == 2 && has4RegOps(MI: SU->getInstr())) |
| 106 | return false; |
| 107 | |
| 108 | // Since a full group is handled immediately in EmitInstruction(), |
| 109 | // SU should fit into current group. NumSlots should be 1 or 0, |
| 110 | // since it is not a cracked or expanded instruction. |
| 111 | assert ((getNumDecoderSlots(SU) <= 1) && (CurrGroupSize < 3) && |
| 112 | "Expected normal instruction to fit in non-full group!" ); |
| 113 | |
| 114 | return true; |
| 115 | } |
| 116 | |
| 117 | bool SystemZHazardRecognizer::has4RegOps(const MachineInstr *MI) const { |
| 118 | const MachineFunction &MF = *MI->getParent()->getParent(); |
| 119 | const TargetRegisterInfo *TRI = &TII->getRegisterInfo(); |
| 120 | const MCInstrDesc &MID = MI->getDesc(); |
| 121 | unsigned Count = 0; |
| 122 | for (unsigned OpIdx = 0; OpIdx < MID.getNumOperands(); OpIdx++) { |
| 123 | const TargetRegisterClass *RC = TII->getRegClass(MCID: MID, OpNum: OpIdx, TRI, MF); |
| 124 | if (RC == nullptr) |
| 125 | continue; |
| 126 | if (OpIdx >= MID.getNumDefs() && |
| 127 | MID.getOperandConstraint(OpNum: OpIdx, Constraint: MCOI::TIED_TO) != -1) |
| 128 | continue; |
| 129 | Count++; |
| 130 | } |
| 131 | return Count >= 4; |
| 132 | } |
| 133 | |
| 134 | void SystemZHazardRecognizer::nextGroup() { |
| 135 | if (CurrGroupSize == 0) |
| 136 | return; |
| 137 | |
| 138 | LLVM_DEBUG(dumpCurrGroup("Completed decode group" )); |
| 139 | LLVM_DEBUG(CurGroupDbg = "" ;); |
| 140 | |
| 141 | int NumGroups = ((CurrGroupSize > 3) ? (CurrGroupSize / 3) : 1); |
| 142 | assert((CurrGroupSize <= 3 || CurrGroupSize % 3 == 0) && |
| 143 | "Current decoder group bad." ); |
| 144 | |
| 145 | // Reset counter for next group. |
| 146 | CurrGroupSize = 0; |
| 147 | CurrGroupHas4RegOps = false; |
| 148 | |
| 149 | GrpCount += ((unsigned) NumGroups); |
| 150 | |
| 151 | // Decrease counters for execution units. |
| 152 | for (unsigned i = 0; i < SchedModel->getNumProcResourceKinds(); ++i) |
| 153 | ProcResourceCounters[i] = ((ProcResourceCounters[i] > NumGroups) |
| 154 | ? (ProcResourceCounters[i] - NumGroups) |
| 155 | : 0); |
| 156 | |
| 157 | // Clear CriticalResourceIdx if it is now below the threshold. |
| 158 | if (CriticalResourceIdx != UINT_MAX && |
| 159 | (ProcResourceCounters[CriticalResourceIdx] <= |
| 160 | ProcResCostLim)) |
| 161 | CriticalResourceIdx = UINT_MAX; |
| 162 | |
| 163 | LLVM_DEBUG(dumpState();); |
| 164 | } |
| 165 | |
| 166 | #ifndef NDEBUG // Debug output |
| 167 | void SystemZHazardRecognizer::dumpSU(SUnit *SU, raw_ostream &OS) const { |
| 168 | OS << "SU(" << SU->NodeNum << "):" ; |
| 169 | OS << TII->getName(SU->getInstr()->getOpcode()); |
| 170 | |
| 171 | const MCSchedClassDesc *SC = getSchedClass(SU); |
| 172 | if (!SC->isValid()) |
| 173 | return; |
| 174 | |
| 175 | for (TargetSchedModel::ProcResIter |
| 176 | PI = SchedModel->getWriteProcResBegin(SC), |
| 177 | PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) { |
| 178 | const MCProcResourceDesc &PRD = |
| 179 | *SchedModel->getProcResource(PI->ProcResourceIdx); |
| 180 | std::string FU(PRD.Name); |
| 181 | // trim e.g. Z13_FXaUnit -> FXa |
| 182 | FU = FU.substr(FU.find('_') + 1); |
| 183 | size_t Pos = FU.find("Unit" ); |
| 184 | if (Pos != std::string::npos) |
| 185 | FU.resize(Pos); |
| 186 | if (FU == "LS" ) // LSUnit -> LSU |
| 187 | FU = "LSU" ; |
| 188 | OS << "/" << FU; |
| 189 | |
| 190 | if (PI->ReleaseAtCycle> 1) |
| 191 | OS << "(" << PI->ReleaseAtCycle << "cyc)" ; |
| 192 | } |
| 193 | |
| 194 | if (SC->NumMicroOps > 1) |
| 195 | OS << "/" << SC->NumMicroOps << "uops" ; |
| 196 | if (SC->BeginGroup && SC->EndGroup) |
| 197 | OS << "/GroupsAlone" ; |
| 198 | else if (SC->BeginGroup) |
| 199 | OS << "/BeginsGroup" ; |
| 200 | else if (SC->EndGroup) |
| 201 | OS << "/EndsGroup" ; |
| 202 | if (SU->isUnbuffered) |
| 203 | OS << "/Unbuffered" ; |
| 204 | if (has4RegOps(SU->getInstr())) |
| 205 | OS << "/4RegOps" ; |
| 206 | } |
| 207 | |
| 208 | void SystemZHazardRecognizer::dumpCurrGroup(std::string Msg) const { |
| 209 | dbgs() << "++ " << Msg; |
| 210 | dbgs() << ": " ; |
| 211 | |
| 212 | if (CurGroupDbg.empty()) |
| 213 | dbgs() << " <empty>\n" ; |
| 214 | else { |
| 215 | dbgs() << "{ " << CurGroupDbg << " }" ; |
| 216 | dbgs() << " (" << CurrGroupSize << " decoder slot" |
| 217 | << (CurrGroupSize > 1 ? "s" :"" ) |
| 218 | << (CurrGroupHas4RegOps ? ", 4RegOps" : "" ) |
| 219 | << ")\n" ; |
| 220 | } |
| 221 | } |
| 222 | |
| 223 | void SystemZHazardRecognizer::dumpProcResourceCounters() const { |
| 224 | bool any = false; |
| 225 | |
| 226 | for (unsigned i = 0; i < SchedModel->getNumProcResourceKinds(); ++i) |
| 227 | if (ProcResourceCounters[i] > 0) { |
| 228 | any = true; |
| 229 | break; |
| 230 | } |
| 231 | |
| 232 | if (!any) |
| 233 | return; |
| 234 | |
| 235 | dbgs() << "++ | Resource counters: " ; |
| 236 | for (unsigned i = 0; i < SchedModel->getNumProcResourceKinds(); ++i) |
| 237 | if (ProcResourceCounters[i] > 0) |
| 238 | dbgs() << SchedModel->getProcResource(i)->Name |
| 239 | << ":" << ProcResourceCounters[i] << " " ; |
| 240 | dbgs() << "\n" ; |
| 241 | |
| 242 | if (CriticalResourceIdx != UINT_MAX) |
| 243 | dbgs() << "++ | Critical resource: " |
| 244 | << SchedModel->getProcResource(CriticalResourceIdx)->Name |
| 245 | << "\n" ; |
| 246 | } |
| 247 | |
| 248 | void SystemZHazardRecognizer::dumpState() const { |
| 249 | dumpCurrGroup("| Current decoder group" ); |
| 250 | dbgs() << "++ | Current cycle index: " |
| 251 | << getCurrCycleIdx() << "\n" ; |
| 252 | dumpProcResourceCounters(); |
| 253 | if (LastFPdOpCycleIdx != UINT_MAX) |
| 254 | dbgs() << "++ | Last FPd cycle index: " << LastFPdOpCycleIdx << "\n" ; |
| 255 | } |
| 256 | |
| 257 | #endif //NDEBUG |
| 258 | |
| 259 | void SystemZHazardRecognizer::clearProcResCounters() { |
| 260 | ProcResourceCounters.assign(NumElts: SchedModel->getNumProcResourceKinds(), Elt: 0); |
| 261 | CriticalResourceIdx = UINT_MAX; |
| 262 | } |
| 263 | |
| 264 | static inline bool isBranchRetTrap(MachineInstr *MI) { |
| 265 | return (MI->isBranch() || MI->isReturn() || |
| 266 | MI->getOpcode() == SystemZ::CondTrap); |
| 267 | } |
| 268 | |
| 269 | // Update state with SU as the next scheduled unit. |
| 270 | void SystemZHazardRecognizer:: |
| 271 | EmitInstruction(SUnit *SU) { |
| 272 | const MCSchedClassDesc *SC = getSchedClass(SU); |
| 273 | LLVM_DEBUG(dbgs() << "++ HazardRecognizer emitting " ; dumpSU(SU, dbgs()); |
| 274 | dbgs() << "\n" ;); |
| 275 | LLVM_DEBUG(dumpCurrGroup("Decode group before emission" );); |
| 276 | |
| 277 | // If scheduling an SU that must begin a new decoder group, move on |
| 278 | // to next group. |
| 279 | if (!fitsIntoCurrentGroup(SU)) |
| 280 | nextGroup(); |
| 281 | |
| 282 | LLVM_DEBUG(raw_string_ostream cgd(CurGroupDbg); |
| 283 | if (CurGroupDbg.length()) cgd << ", " ; dumpSU(SU, cgd);); |
| 284 | |
| 285 | LastEmittedMI = SU->getInstr(); |
| 286 | |
| 287 | // After returning from a call, we don't know much about the state. |
| 288 | if (SU->isCall) { |
| 289 | LLVM_DEBUG(dbgs() << "++ Clearing state after call.\n" ;); |
| 290 | Reset(); |
| 291 | LastEmittedMI = SU->getInstr(); |
| 292 | return; |
| 293 | } |
| 294 | |
| 295 | // Increase counter for execution unit(s). |
| 296 | for (TargetSchedModel::ProcResIter |
| 297 | PI = SchedModel->getWriteProcResBegin(SC), |
| 298 | PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) { |
| 299 | // Don't handle FPd together with the other resources. |
| 300 | if (SchedModel->getProcResource(PIdx: PI->ProcResourceIdx)->BufferSize == 1) |
| 301 | continue; |
| 302 | int &CurrCounter = |
| 303 | ProcResourceCounters[PI->ProcResourceIdx]; |
| 304 | CurrCounter += PI->ReleaseAtCycle; |
| 305 | // Check if this is now the new critical resource. |
| 306 | if ((CurrCounter > ProcResCostLim) && |
| 307 | (CriticalResourceIdx == UINT_MAX || |
| 308 | (PI->ProcResourceIdx != CriticalResourceIdx && |
| 309 | CurrCounter > |
| 310 | ProcResourceCounters[CriticalResourceIdx]))) { |
| 311 | LLVM_DEBUG( |
| 312 | dbgs() << "++ New critical resource: " |
| 313 | << SchedModel->getProcResource(PI->ProcResourceIdx)->Name |
| 314 | << "\n" ;); |
| 315 | CriticalResourceIdx = PI->ProcResourceIdx; |
| 316 | } |
| 317 | } |
| 318 | |
| 319 | // Make note of an instruction that uses a blocking resource (FPd). |
| 320 | if (SU->isUnbuffered) { |
| 321 | LastFPdOpCycleIdx = getCurrCycleIdx(SU); |
| 322 | LLVM_DEBUG(dbgs() << "++ Last FPd cycle index: " << LastFPdOpCycleIdx |
| 323 | << "\n" ;); |
| 324 | } |
| 325 | |
| 326 | // Insert SU into current group by increasing number of slots used |
| 327 | // in current group. |
| 328 | CurrGroupSize += getNumDecoderSlots(SU); |
| 329 | CurrGroupHas4RegOps |= has4RegOps(MI: SU->getInstr()); |
| 330 | unsigned GroupLim = (CurrGroupHas4RegOps ? 2 : 3); |
| 331 | assert((CurrGroupSize <= GroupLim || CurrGroupSize == getNumDecoderSlots(SU)) |
| 332 | && "SU does not fit into decoder group!" ); |
| 333 | |
| 334 | // Check if current group is now full/ended. If so, move on to next |
| 335 | // group to be ready to evaluate more candidates. |
| 336 | if (CurrGroupSize >= GroupLim || SC->EndGroup) |
| 337 | nextGroup(); |
| 338 | } |
| 339 | |
| 340 | int SystemZHazardRecognizer::groupingCost(SUnit *SU) const { |
| 341 | const MCSchedClassDesc *SC = getSchedClass(SU); |
| 342 | if (!SC->isValid()) |
| 343 | return 0; |
| 344 | |
| 345 | // If SU begins new group, it can either break a current group early |
| 346 | // or fit naturally if current group is empty (negative cost). |
| 347 | if (SC->BeginGroup) { |
| 348 | if (CurrGroupSize) |
| 349 | return 3 - CurrGroupSize; |
| 350 | return -1; |
| 351 | } |
| 352 | |
| 353 | // Similarly, a group-ending SU may either fit well (last in group), or |
| 354 | // end the group prematurely. |
| 355 | if (SC->EndGroup) { |
| 356 | unsigned resultingGroupSize = |
| 357 | (CurrGroupSize + getNumDecoderSlots(SU)); |
| 358 | if (resultingGroupSize < 3) |
| 359 | return (3 - resultingGroupSize); |
| 360 | return -1; |
| 361 | } |
| 362 | |
| 363 | // An instruction with 4 register operands will not fit in last slot. |
| 364 | if (CurrGroupSize == 2 && has4RegOps(MI: SU->getInstr())) |
| 365 | return 1; |
| 366 | |
| 367 | // Most instructions can be placed in any decoder slot. |
| 368 | return 0; |
| 369 | } |
| 370 | |
| 371 | bool SystemZHazardRecognizer::isFPdOpPreferred_distance(SUnit *SU) const { |
| 372 | assert (SU->isUnbuffered); |
| 373 | // If this is the first FPd op, it should be scheduled high. |
| 374 | if (LastFPdOpCycleIdx == UINT_MAX) |
| 375 | return true; |
| 376 | // If this is not the first PFd op, it should go into the other side |
| 377 | // of the processor to use the other FPd unit there. This should |
| 378 | // generally happen if two FPd ops are placed with 2 other |
| 379 | // instructions between them (modulo 6). |
| 380 | unsigned SUCycleIdx = getCurrCycleIdx(SU); |
| 381 | if (LastFPdOpCycleIdx > SUCycleIdx) |
| 382 | return ((LastFPdOpCycleIdx - SUCycleIdx) == 3); |
| 383 | return ((SUCycleIdx - LastFPdOpCycleIdx) == 3); |
| 384 | } |
| 385 | |
| 386 | int SystemZHazardRecognizer:: |
| 387 | resourcesCost(SUnit *SU) { |
| 388 | int Cost = 0; |
| 389 | |
| 390 | const MCSchedClassDesc *SC = getSchedClass(SU); |
| 391 | if (!SC->isValid()) |
| 392 | return 0; |
| 393 | |
| 394 | // For a FPd op, either return min or max value as indicated by the |
| 395 | // distance to any prior FPd op. |
| 396 | if (SU->isUnbuffered) |
| 397 | Cost = (isFPdOpPreferred_distance(SU) ? INT_MIN : INT_MAX); |
| 398 | // For other instructions, give a cost to the use of the critical resource. |
| 399 | else if (CriticalResourceIdx != UINT_MAX) { |
| 400 | for (TargetSchedModel::ProcResIter |
| 401 | PI = SchedModel->getWriteProcResBegin(SC), |
| 402 | PE = SchedModel->getWriteProcResEnd(SC); PI != PE; ++PI) |
| 403 | if (PI->ProcResourceIdx == CriticalResourceIdx) |
| 404 | Cost = PI->ReleaseAtCycle; |
| 405 | } |
| 406 | |
| 407 | return Cost; |
| 408 | } |
| 409 | |
| 410 | void SystemZHazardRecognizer::emitInstruction(MachineInstr *MI, |
| 411 | bool TakenBranch) { |
| 412 | // Make a temporary SUnit. |
| 413 | SUnit SU(MI, 0); |
| 414 | |
| 415 | // Set interesting flags. |
| 416 | SU.isCall = MI->isCall(); |
| 417 | |
| 418 | const MCSchedClassDesc *SC = SchedModel->resolveSchedClass(MI); |
| 419 | for (const MCWriteProcResEntry &PRE : |
| 420 | make_range(x: SchedModel->getWriteProcResBegin(SC), |
| 421 | y: SchedModel->getWriteProcResEnd(SC))) { |
| 422 | switch (SchedModel->getProcResource(PIdx: PRE.ProcResourceIdx)->BufferSize) { |
| 423 | case 0: |
| 424 | SU.hasReservedResource = true; |
| 425 | break; |
| 426 | case 1: |
| 427 | SU.isUnbuffered = true; |
| 428 | break; |
| 429 | default: |
| 430 | break; |
| 431 | } |
| 432 | } |
| 433 | |
| 434 | unsigned GroupSizeBeforeEmit = CurrGroupSize; |
| 435 | EmitInstruction(SU: &SU); |
| 436 | |
| 437 | if (!TakenBranch && isBranchRetTrap(MI)) { |
| 438 | // NT Branch on second slot ends group. |
| 439 | if (GroupSizeBeforeEmit == 1) |
| 440 | nextGroup(); |
| 441 | } |
| 442 | |
| 443 | if (TakenBranch && CurrGroupSize > 0) |
| 444 | nextGroup(); |
| 445 | |
| 446 | assert ((!MI->isTerminator() || isBranchRetTrap(MI)) && |
| 447 | "Scheduler: unhandled terminator!" ); |
| 448 | } |
| 449 | |
| 450 | void SystemZHazardRecognizer:: |
| 451 | copyState(SystemZHazardRecognizer *Incoming) { |
| 452 | // Current decoder group |
| 453 | CurrGroupSize = Incoming->CurrGroupSize; |
| 454 | LLVM_DEBUG(CurGroupDbg = Incoming->CurGroupDbg;); |
| 455 | |
| 456 | // Processor resources |
| 457 | ProcResourceCounters = Incoming->ProcResourceCounters; |
| 458 | CriticalResourceIdx = Incoming->CriticalResourceIdx; |
| 459 | |
| 460 | // FPd |
| 461 | LastFPdOpCycleIdx = Incoming->LastFPdOpCycleIdx; |
| 462 | GrpCount = Incoming->GrpCount; |
| 463 | } |
| 464 | |