| 1 | //-- SystemZMachineScheduler.cpp - SystemZ Scheduler Interface -*- 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 | #include "SystemZMachineScheduler.h" |
| 10 | #include "llvm/CodeGen/MachineLoopInfo.h" |
| 11 | |
| 12 | using namespace llvm; |
| 13 | |
| 14 | #define DEBUG_TYPE "machine-scheduler" |
| 15 | |
| 16 | /// Pre-RA scheduling /// |
| 17 | |
| 18 | static bool isRegDef(const MachineOperand &MO) { |
| 19 | return MO.isReg() && MO.isDef(); |
| 20 | } |
| 21 | |
| 22 | bool SystemZPreRASchedStrategy::definesCmp0Src(const MachineInstr *MI, |
| 23 | bool CCDef) const { |
| 24 | if (Cmp0SrcReg != SystemZ::NoRegister && MI->getNumOperands() && |
| 25 | (MI->getDesc().hasImplicitDefOfPhysReg(Reg: SystemZ::CC) || !CCDef)) { |
| 26 | const MachineOperand &MO0 = MI->getOperand(i: 0); |
| 27 | if (isRegDef(MO: MO0) && MO0.getReg() == Cmp0SrcReg) |
| 28 | return true; |
| 29 | } |
| 30 | return false; |
| 31 | } |
| 32 | |
| 33 | bool SystemZPreRASchedStrategy::closesLiveRange(const SUnit *SU, |
| 34 | ScheduleDAGMILive *DAG) const { |
| 35 | if (SU->getInstr()->isCopy()) |
| 36 | return false; |
| 37 | |
| 38 | // Extract the PressureChanges that all fp/vector or GR64/GR32/GRH32 regs |
| 39 | // affect respectively. misched-prera-pdiffs.mir tests against any future |
| 40 | // change in the PressureSets modelling, so simply hard-code them here. |
| 41 | int VR16PChange = 0, GRX32PChange = 0; |
| 42 | const PressureDiff &PDiff = DAG->getPressureDiff(SU); |
| 43 | for (const PressureChange &PC : PDiff) { |
| 44 | if (!PC.isValid()) |
| 45 | break; |
| 46 | if (PC.getPSet() == SystemZ::VR16Bit) |
| 47 | VR16PChange = PC.getUnitInc(); |
| 48 | else if (PC.getPSet() == SystemZ::GRX32Bit) |
| 49 | GRX32PChange = PC.getUnitInc(); |
| 50 | } |
| 51 | |
| 52 | // Return true for a (vreg) def when register pressure is reduced. Prioritize |
| 53 | // FP/vector regs over GPRs. |
| 54 | return VR16PChange < 0 || (!VR16PChange && GRX32PChange < 0); |
| 55 | } |
| 56 | |
| 57 | bool SystemZPreRASchedStrategy::tryCandidate(SchedCandidate &Cand, |
| 58 | SchedCandidate &TryCand, |
| 59 | SchedBoundary *Zone) const { |
| 60 | assert(Zone && !Zone->isTop() && "Bottom-Up scheduling only." ); |
| 61 | |
| 62 | // Initialize the candidate if needed. |
| 63 | if (!Cand.isValid()) { |
| 64 | TryCand.Reason = FirstValid; |
| 65 | return true; |
| 66 | } |
| 67 | |
| 68 | // Bias physreg defs and copies to their uses and definitions respectively. |
| 69 | if (tryBiasPhysRegs(TryCand, Cand, Zone, /*BiasPRegsExtra=*/true)) |
| 70 | return TryCand.Reason != NoCand; |
| 71 | |
| 72 | if (RegionPolicy.ShouldTrackPressure) { |
| 73 | auto schedLow = [&](const SUnit *SU) { |
| 74 | return SU->getHeight() <= Zone->getScheduledLatency() && |
| 75 | SU->getHeight() < LivenessHeightCutOff && closesLiveRange(SU, DAG); |
| 76 | }; |
| 77 | // One SU closes a live range while preserving the scheduled latency. |
| 78 | if (tryGreater(TryVal: schedLow(TryCand.SU), CandVal: schedLow(Cand.SU), TryCand, Cand, |
| 79 | Reason: RegExcess)) |
| 80 | return TryCand.Reason != NoCand; |
| 81 | } |
| 82 | |
| 83 | // Do latency scheduling by trying to not increase the scheduled latency |
| 84 | // (going bottom-up). This is important for performance, but it is best to |
| 85 | // only do it when likely beneficial and not "move everything around": Only |
| 86 | // apply this to SUs that have somewhat longer latencies, and don't push an |
| 87 | // SU upwards if it's on the critical path. |
| 88 | if (!RegionPolicy.DisableLatencyHeuristic && |
| 89 | std::max(a: TryCand.SU->Latency, b: Cand.SU->Latency) >= 5) |
| 90 | if (const SUnit *HigherSU = |
| 91 | TryCand.SU->getHeight() > Cand.SU->getHeight() ? TryCand.SU |
| 92 | : TryCand.SU->getHeight() < Cand.SU->getHeight() ? Cand.SU |
| 93 | : nullptr) |
| 94 | if (HigherSU->getHeight() > Zone->getScheduledLatency() && |
| 95 | HigherSU->getDepth() < computeRemLatency(CurrZone&: *Zone)) { |
| 96 | // HigherSU would increase the scheduled latency and there is room |
| 97 | // above for it next to the critical path, so wait with it. |
| 98 | tryLess(TryVal: TryCand.SU->getHeight(), CandVal: Cand.SU->getHeight(), TryCand, Cand, |
| 99 | Reason: GenericSchedulerBase::BotHeightReduce); |
| 100 | return TryCand.Reason != NoCand; |
| 101 | } |
| 102 | |
| 103 | // Weak edges help copy coalescing. |
| 104 | if (tryLess(TryVal: TryCand.SU->WeakSuccsLeft, CandVal: Cand.SU->WeakSuccsLeft, TryCand, Cand, |
| 105 | Reason: Weak)) |
| 106 | return TryCand.Reason != NoCand; |
| 107 | |
| 108 | // Help compare with zero elimination. |
| 109 | if (tryGreater(TryVal: definesCmp0Src(MI: TryCand.SU->getInstr()), |
| 110 | CandVal: definesCmp0Src(MI: Cand.SU->getInstr()), TryCand, Cand, Reason: Weak)) |
| 111 | return TryCand.Reason != NoCand; |
| 112 | |
| 113 | // Fall through to original instruction order. |
| 114 | if (TryCand.SU->NodeNum > Cand.SU->NodeNum) { |
| 115 | TryCand.Reason = NodeOrder; |
| 116 | return true; |
| 117 | } |
| 118 | |
| 119 | return false; |
| 120 | } |
| 121 | |
| 122 | void SystemZPreRASchedStrategy::initPolicy(MachineBasicBlock::iterator Begin, |
| 123 | MachineBasicBlock::iterator End, |
| 124 | unsigned NumRegionInstrs) { |
| 125 | // TopRegionSUs is the number of SUs that is considered to be part of the |
| 126 | // "top" of a region. Liveness reduction is not done in regions smaller than |
| 127 | // this. The idea is to prioritize latency more after branches and help |
| 128 | // liveness only when the decoder is ahead of execution anyway. |
| 129 | static const unsigned TopRegionSUs = 36; |
| 130 | |
| 131 | // Avoid setting up the register pressure tracker unless needed to save |
| 132 | // compile time. |
| 133 | RegionPolicy.ShouldTrackPressure = NumRegionInstrs > TopRegionSUs; |
| 134 | |
| 135 | // These heuristics has so far seemed to work better without adding a |
| 136 | // top-down boundary. |
| 137 | RegionPolicy.OnlyBottomUp = true; |
| 138 | |
| 139 | BotIdx = NumRegionInstrs - 1; |
| 140 | this->NumRegionInstrs = NumRegionInstrs; |
| 141 | } |
| 142 | |
| 143 | void SystemZPreRASchedStrategy::initialize(ScheduleDAGMI *dag) { |
| 144 | GenericScheduler::initialize(dag); |
| 145 | |
| 146 | Cmp0SrcReg = SystemZ::NoRegister; |
| 147 | |
| 148 | unsigned DAGHeight = 0; |
| 149 | for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) |
| 150 | DAGHeight = std::max(a: DAGHeight, b: DAG->SUnits[Idx].getHeight()); |
| 151 | |
| 152 | if (RegionPolicy.ShouldTrackPressure) |
| 153 | LivenessHeightCutOff = DAGHeight / (DAG->SUnits.size() < 50 ? 4 : 2); |
| 154 | |
| 155 | // Disable latency reduction if region has many SUs relative to the |
| 156 | // overall height. |
| 157 | RegionPolicy.DisableLatencyHeuristic = |
| 158 | DAG->SUnits.size() >= 3 * std::max(a: DAGHeight, b: 1u); |
| 159 | } |
| 160 | |
| 161 | void SystemZPreRASchedStrategy::schedNode(SUnit *SU, bool IsTopNode) { |
| 162 | GenericScheduler::schedNode(SU, IsTopNode); |
| 163 | |
| 164 | const SystemZInstrInfo *TII = static_cast<const SystemZInstrInfo *>(DAG->TII); |
| 165 | MachineInstr *MI = SU->getInstr(); |
| 166 | if (TII->isCompareZero(Compare: *MI)) |
| 167 | Cmp0SrcReg = TII->getCompareSourceReg(Compare: *MI); |
| 168 | else if (MI->getDesc().hasImplicitDefOfPhysReg(Reg: SystemZ::CC) || |
| 169 | definesCmp0Src(MI, /*CCDef=*/false)) |
| 170 | Cmp0SrcReg = SystemZ::NoRegister; |
| 171 | } |
| 172 | |
| 173 | /// Post-RA scheduling /// |
| 174 | |
| 175 | #ifndef NDEBUG |
| 176 | // Print the set of SUs |
| 177 | void SystemZPostRASchedStrategy::SUSet:: |
| 178 | dump(SystemZHazardRecognizer &HazardRec) const { |
| 179 | dbgs() << "{" ; |
| 180 | for (auto &SU : *this) { |
| 181 | HazardRec.dumpSU(SU, dbgs()); |
| 182 | if (SU != *rbegin()) |
| 183 | dbgs() << ", " ; |
| 184 | } |
| 185 | dbgs() << "}\n" ; |
| 186 | } |
| 187 | #endif |
| 188 | |
| 189 | // Try to find a single predecessor that would be interesting for the |
| 190 | // scheduler in the top-most region of MBB. |
| 191 | static MachineBasicBlock *getSingleSchedPred(MachineBasicBlock *MBB, |
| 192 | const MachineLoop *Loop) { |
| 193 | MachineBasicBlock *PredMBB = nullptr; |
| 194 | if (MBB->pred_size() == 1) |
| 195 | PredMBB = *MBB->pred_begin(); |
| 196 | |
| 197 | // The loop header has two predecessors, return the latch, but not for a |
| 198 | // single block loop. |
| 199 | if (MBB->pred_size() == 2 && Loop != nullptr && Loop->getHeader() == MBB) { |
| 200 | for (MachineBasicBlock *Pred : MBB->predecessors()) |
| 201 | if (Loop->contains(BB: Pred)) |
| 202 | PredMBB = (Pred == MBB ? nullptr : Pred); |
| 203 | } |
| 204 | |
| 205 | assert ((PredMBB == nullptr || !Loop || Loop->contains(PredMBB)) |
| 206 | && "Loop MBB should not consider predecessor outside of loop." ); |
| 207 | |
| 208 | return PredMBB; |
| 209 | } |
| 210 | |
| 211 | void SystemZPostRASchedStrategy:: |
| 212 | advanceTo(MachineBasicBlock::iterator NextBegin) { |
| 213 | MachineBasicBlock::iterator LastEmittedMI = HazardRec->getLastEmittedMI(); |
| 214 | MachineBasicBlock::iterator I = |
| 215 | ((LastEmittedMI != nullptr && LastEmittedMI->getParent() == MBB) ? |
| 216 | std::next(x: LastEmittedMI) : MBB->begin()); |
| 217 | |
| 218 | for (; I != NextBegin; ++I) { |
| 219 | if (I->isPosition() || I->isDebugInstr()) |
| 220 | continue; |
| 221 | HazardRec->emitInstruction(MI: &*I); |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | void SystemZPostRASchedStrategy::initialize(ScheduleDAGMI *dag) { |
| 226 | Available.clear(); // -misched-cutoff. |
| 227 | LLVM_DEBUG(HazardRec->dumpState();); |
| 228 | } |
| 229 | |
| 230 | void SystemZPostRASchedStrategy::enterMBB(MachineBasicBlock *NextMBB) { |
| 231 | assert ((SchedStates.find(NextMBB) == SchedStates.end()) && |
| 232 | "Entering MBB twice?" ); |
| 233 | LLVM_DEBUG(dbgs() << "** Entering " << printMBBReference(*NextMBB)); |
| 234 | |
| 235 | MBB = NextMBB; |
| 236 | |
| 237 | /// Create a HazardRec for MBB, save it in SchedStates and set HazardRec to |
| 238 | /// point to it. |
| 239 | HazardRec = SchedStates[MBB] = new SystemZHazardRecognizer(TII, &SchedModel); |
| 240 | LLVM_DEBUG(const MachineLoop *Loop = MLI->getLoopFor(MBB); |
| 241 | if (Loop && Loop->getHeader() == MBB) dbgs() << " (Loop header)" ; |
| 242 | dbgs() << ":\n" ;); |
| 243 | |
| 244 | // Try to take over the state from a single predecessor, if it has been |
| 245 | // scheduled. If this is not possible, we are done. |
| 246 | MachineBasicBlock *SinglePredMBB = |
| 247 | getSingleSchedPred(MBB, Loop: MLI->getLoopFor(BB: MBB)); |
| 248 | if (SinglePredMBB == nullptr) |
| 249 | return; |
| 250 | auto It = SchedStates.find(x: SinglePredMBB); |
| 251 | if (It == SchedStates.end()) |
| 252 | return; |
| 253 | |
| 254 | LLVM_DEBUG(dbgs() << "** Continued scheduling from " |
| 255 | << printMBBReference(*SinglePredMBB) << "\n" ;); |
| 256 | |
| 257 | HazardRec->copyState(Incoming: It->second); |
| 258 | LLVM_DEBUG(HazardRec->dumpState();); |
| 259 | |
| 260 | // Emit incoming terminator(s). Be optimistic and assume that branch |
| 261 | // prediction will generally do "the right thing". |
| 262 | for (MachineInstr &MI : SinglePredMBB->terminators()) { |
| 263 | LLVM_DEBUG(dbgs() << "** Emitting incoming branch: " ; MI.dump();); |
| 264 | bool TakenBranch = (MI.isBranch() && |
| 265 | (TII->getBranchInfo(MI).isIndirect() || |
| 266 | TII->getBranchInfo(MI).getMBBTarget() == MBB)); |
| 267 | HazardRec->emitInstruction(MI: &MI, TakenBranch); |
| 268 | if (TakenBranch) |
| 269 | break; |
| 270 | } |
| 271 | } |
| 272 | |
| 273 | void SystemZPostRASchedStrategy::leaveMBB() { |
| 274 | LLVM_DEBUG(dbgs() << "** Leaving " << printMBBReference(*MBB) << "\n" ;); |
| 275 | |
| 276 | // Advance to first terminator. The successor block will handle terminators |
| 277 | // dependent on CFG layout (T/NT branch etc). |
| 278 | advanceTo(NextBegin: MBB->getFirstTerminator()); |
| 279 | } |
| 280 | |
| 281 | SystemZPostRASchedStrategy:: |
| 282 | SystemZPostRASchedStrategy(const MachineSchedContext *C) |
| 283 | : MLI(C->MLI), |
| 284 | TII(static_cast<const SystemZInstrInfo *> |
| 285 | (C->MF->getSubtarget().getInstrInfo())), |
| 286 | MBB(nullptr), HazardRec(nullptr) { |
| 287 | const TargetSubtargetInfo *ST = &C->MF->getSubtarget(); |
| 288 | SchedModel.init(TSInfo: ST); |
| 289 | } |
| 290 | |
| 291 | SystemZPostRASchedStrategy::~SystemZPostRASchedStrategy() { |
| 292 | // Delete hazard recognizers kept around for each MBB. |
| 293 | for (auto I : SchedStates) { |
| 294 | SystemZHazardRecognizer *hazrec = I.second; |
| 295 | delete hazrec; |
| 296 | } |
| 297 | } |
| 298 | |
| 299 | void SystemZPostRASchedStrategy::initPolicy(MachineBasicBlock::iterator Begin, |
| 300 | MachineBasicBlock::iterator End, |
| 301 | unsigned NumRegionInstrs) { |
| 302 | // Don't emit the terminators. |
| 303 | if (Begin->isTerminator()) |
| 304 | return; |
| 305 | |
| 306 | // Emit any instructions before start of region. |
| 307 | advanceTo(NextBegin: Begin); |
| 308 | } |
| 309 | |
| 310 | // Pick the next node to schedule. |
| 311 | SUnit *SystemZPostRASchedStrategy::pickNode(bool &IsTopNode) { |
| 312 | // Only scheduling top-down. |
| 313 | IsTopNode = true; |
| 314 | |
| 315 | if (Available.empty()) |
| 316 | return nullptr; |
| 317 | |
| 318 | // If only one choice, return it. |
| 319 | if (Available.size() == 1) { |
| 320 | LLVM_DEBUG(dbgs() << "** Only one: " ; |
| 321 | HazardRec->dumpSU(*Available.begin(), dbgs()); dbgs() << "\n" ;); |
| 322 | return *Available.begin(); |
| 323 | } |
| 324 | |
| 325 | // All nodes that are possible to schedule are stored in the Available set. |
| 326 | LLVM_DEBUG(dbgs() << "** Available: " ; Available.dump(*HazardRec);); |
| 327 | |
| 328 | Candidate Best; |
| 329 | for (auto *SU : Available) { |
| 330 | |
| 331 | // SU is the next candidate to be compared against current Best. |
| 332 | Candidate c(SU, *HazardRec); |
| 333 | |
| 334 | // Remeber which SU is the best candidate. |
| 335 | if (Best.SU == nullptr || c < Best) { |
| 336 | Best = c; |
| 337 | LLVM_DEBUG(dbgs() << "** Best so far: " ;); |
| 338 | } else |
| 339 | LLVM_DEBUG(dbgs() << "** Tried : " ;); |
| 340 | LLVM_DEBUG(HazardRec->dumpSU(c.SU, dbgs()); c.dumpCosts(); |
| 341 | dbgs() << " Height:" << c.SU->getHeight(); dbgs() << "\n" ;); |
| 342 | |
| 343 | // Once we know we have seen all SUs that affect grouping or use unbuffered |
| 344 | // resources, we can stop iterating if Best looks good. |
| 345 | if (!SU->isScheduleHigh && Best.noCost()) |
| 346 | break; |
| 347 | } |
| 348 | |
| 349 | assert (Best.SU != nullptr); |
| 350 | return Best.SU; |
| 351 | } |
| 352 | |
| 353 | SystemZPostRASchedStrategy::Candidate:: |
| 354 | Candidate(SUnit *SU_, SystemZHazardRecognizer &HazardRec) : Candidate() { |
| 355 | SU = SU_; |
| 356 | |
| 357 | // Check the grouping cost. For a node that must begin / end a |
| 358 | // group, it is positive if it would do so prematurely, or negative |
| 359 | // if it would fit naturally into the schedule. |
| 360 | GroupingCost = HazardRec.groupingCost(SU); |
| 361 | |
| 362 | // Check the resources cost for this SU. |
| 363 | ResourcesCost = HazardRec.resourcesCost(SU); |
| 364 | } |
| 365 | |
| 366 | bool SystemZPostRASchedStrategy::Candidate:: |
| 367 | operator<(const Candidate &other) { |
| 368 | |
| 369 | // Check decoder grouping. |
| 370 | if (GroupingCost < other.GroupingCost) |
| 371 | return true; |
| 372 | if (GroupingCost > other.GroupingCost) |
| 373 | return false; |
| 374 | |
| 375 | // Compare the use of resources. |
| 376 | if (ResourcesCost < other.ResourcesCost) |
| 377 | return true; |
| 378 | if (ResourcesCost > other.ResourcesCost) |
| 379 | return false; |
| 380 | |
| 381 | // Higher SU is otherwise generally better. |
| 382 | if (SU->getHeight() > other.SU->getHeight()) |
| 383 | return true; |
| 384 | if (SU->getHeight() < other.SU->getHeight()) |
| 385 | return false; |
| 386 | |
| 387 | // If all same, fall back to original order. |
| 388 | if (SU->NodeNum < other.SU->NodeNum) |
| 389 | return true; |
| 390 | |
| 391 | return false; |
| 392 | } |
| 393 | |
| 394 | void SystemZPostRASchedStrategy::schedNode(SUnit *SU, bool IsTopNode) { |
| 395 | LLVM_DEBUG(dbgs() << "** Scheduling SU(" << SU->NodeNum << ") " ; |
| 396 | if (Available.size() == 1) dbgs() << "(only one) " ; |
| 397 | Candidate c(SU, *HazardRec); c.dumpCosts(); dbgs() << "\n" ;); |
| 398 | |
| 399 | // Remove SU from Available set and update HazardRec. |
| 400 | Available.erase(x: SU); |
| 401 | HazardRec->EmitInstruction(SU); |
| 402 | } |
| 403 | |
| 404 | void SystemZPostRASchedStrategy::releaseTopNode(SUnit *SU) { |
| 405 | // Set isScheduleHigh flag on all SUs that we want to consider first in |
| 406 | // pickNode(). |
| 407 | const MCSchedClassDesc *SC = HazardRec->getSchedClass(SU); |
| 408 | bool AffectsGrouping = (SC->isValid() && (SC->BeginGroup || SC->EndGroup)); |
| 409 | SU->isScheduleHigh = (AffectsGrouping || SU->isUnbuffered); |
| 410 | |
| 411 | // Put all released SUs in the Available set. |
| 412 | Available.insert(x: SU); |
| 413 | } |
| 414 | |