ScheduleDAGInstrs.cpp revision ec6906ba47d6d32cc817e85eddb87b320d6ae18c
1//===---- ScheduleDAGInstrs.cpp - MachineInstr Rescheduling ---------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This implements the ScheduleDAGInstrs class, which implements re-scheduling 11// of MachineInstrs. 12// 13//===----------------------------------------------------------------------===// 14 15#define DEBUG_TYPE "sched-instrs" 16#include "ScheduleDAGInstrs.h" 17#include "llvm/Operator.h" 18#include "llvm/Analysis/AliasAnalysis.h" 19#include "llvm/CodeGen/MachineFunctionPass.h" 20#include "llvm/CodeGen/MachineMemOperand.h" 21#include "llvm/CodeGen/MachineRegisterInfo.h" 22#include "llvm/CodeGen/PseudoSourceValue.h" 23#include "llvm/Target/TargetMachine.h" 24#include "llvm/Target/TargetInstrInfo.h" 25#include "llvm/Target/TargetRegisterInfo.h" 26#include "llvm/Target/TargetSubtarget.h" 27#include "llvm/Support/Debug.h" 28#include "llvm/Support/raw_ostream.h" 29#include "llvm/ADT/SmallSet.h" 30using namespace llvm; 31 32ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf, 33 const MachineLoopInfo &mli, 34 const MachineDominatorTree &mdt) 35 : ScheduleDAG(mf), MLI(mli), MDT(mdt), MFI(mf.getFrameInfo()), 36 InstrItins(mf.getTarget().getInstrItineraryData()), 37 Defs(TRI->getNumRegs()), Uses(TRI->getNumRegs()), LoopRegs(MLI, MDT) { 38 DbgValueVec.clear(); 39} 40 41/// Run - perform scheduling. 42/// 43void ScheduleDAGInstrs::Run(MachineBasicBlock *bb, 44 MachineBasicBlock::iterator begin, 45 MachineBasicBlock::iterator end, 46 unsigned endcount) { 47 BB = bb; 48 Begin = begin; 49 InsertPosIndex = endcount; 50 51 ScheduleDAG::Run(bb, end); 52} 53 54/// getUnderlyingObjectFromInt - This is the function that does the work of 55/// looking through basic ptrtoint+arithmetic+inttoptr sequences. 56static const Value *getUnderlyingObjectFromInt(const Value *V) { 57 do { 58 if (const Operator *U = dyn_cast<Operator>(V)) { 59 // If we find a ptrtoint, we can transfer control back to the 60 // regular getUnderlyingObjectFromInt. 61 if (U->getOpcode() == Instruction::PtrToInt) 62 return U->getOperand(0); 63 // If we find an add of a constant or a multiplied value, it's 64 // likely that the other operand will lead us to the base 65 // object. We don't have to worry about the case where the 66 // object address is somehow being computed by the multiply, 67 // because our callers only care when the result is an 68 // identifibale object. 69 if (U->getOpcode() != Instruction::Add || 70 (!isa<ConstantInt>(U->getOperand(1)) && 71 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul)) 72 return V; 73 V = U->getOperand(0); 74 } else { 75 return V; 76 } 77 assert(V->getType()->isIntegerTy() && "Unexpected operand type!"); 78 } while (1); 79} 80 81/// getUnderlyingObject - This is a wrapper around Value::getUnderlyingObject 82/// and adds support for basic ptrtoint+arithmetic+inttoptr sequences. 83static const Value *getUnderlyingObject(const Value *V) { 84 // First just call Value::getUnderlyingObject to let it do what it does. 85 do { 86 V = V->getUnderlyingObject(); 87 // If it found an inttoptr, use special code to continue climing. 88 if (Operator::getOpcode(V) != Instruction::IntToPtr) 89 break; 90 const Value *O = getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0)); 91 // If that succeeded in finding a pointer, continue the search. 92 if (!O->getType()->isPointerTy()) 93 break; 94 V = O; 95 } while (1); 96 return V; 97} 98 99/// getUnderlyingObjectForInstr - If this machine instr has memory reference 100/// information and it can be tracked to a normal reference to a known 101/// object, return the Value for that object. Otherwise return null. 102static const Value *getUnderlyingObjectForInstr(const MachineInstr *MI, 103 const MachineFrameInfo *MFI, 104 bool &MayAlias) { 105 MayAlias = true; 106 if (!MI->hasOneMemOperand() || 107 !(*MI->memoperands_begin())->getValue() || 108 (*MI->memoperands_begin())->isVolatile()) 109 return 0; 110 111 const Value *V = (*MI->memoperands_begin())->getValue(); 112 if (!V) 113 return 0; 114 115 V = getUnderlyingObject(V); 116 if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) { 117 // For now, ignore PseudoSourceValues which may alias LLVM IR values 118 // because the code that uses this function has no way to cope with 119 // such aliases. 120 if (PSV->isAliased(MFI)) 121 return 0; 122 123 MayAlias = PSV->mayAlias(MFI); 124 return V; 125 } 126 127 if (isIdentifiedObject(V)) 128 return V; 129 130 return 0; 131} 132 133void ScheduleDAGInstrs::StartBlock(MachineBasicBlock *BB) { 134 if (MachineLoop *ML = MLI.getLoopFor(BB)) 135 if (BB == ML->getLoopLatch()) { 136 MachineBasicBlock *Header = ML->getHeader(); 137 for (MachineBasicBlock::livein_iterator I = Header->livein_begin(), 138 E = Header->livein_end(); I != E; ++I) 139 LoopLiveInRegs.insert(*I); 140 LoopRegs.VisitLoop(ML); 141 } 142} 143 144/// AddSchedBarrierDeps - Add dependencies from instructions in the current 145/// list of instructions being scheduled to scheduling barrier by adding 146/// the exit SU to the register defs and use list. This is because we want to 147/// make sure instructions which define registers that are either used by 148/// the terminator or are live-out are properly scheduled. This is 149/// especially important when the definition latency of the return value(s) 150/// are too high to be hidden by the branch or when the liveout registers 151/// used by instructions in the fallthrough block. 152void ScheduleDAGInstrs::AddSchedBarrierDeps() { 153 MachineInstr *ExitMI = InsertPos != BB->end() ? &*InsertPos : 0; 154 ExitSU.setInstr(ExitMI); 155 bool AllDepKnown = ExitMI && 156 (ExitMI->getDesc().isCall() || ExitMI->getDesc().isBarrier()); 157 if (ExitMI && AllDepKnown) { 158 // If it's a call or a barrier, add dependencies on the defs and uses of 159 // instruction. 160 for (unsigned i = 0, e = ExitMI->getNumOperands(); i != e; ++i) { 161 const MachineOperand &MO = ExitMI->getOperand(i); 162 if (!MO.isReg() || MO.isDef()) continue; 163 unsigned Reg = MO.getReg(); 164 if (Reg == 0) continue; 165 166 assert(TRI->isPhysicalRegister(Reg) && "Virtual register encountered!"); 167 Uses[Reg].push_back(&ExitSU); 168 } 169 } else { 170 // For others, e.g. fallthrough, conditional branch, assume the exit 171 // uses all the registers. 172 for (int i = 0, e = TRI->getNumRegs(); i != e; ++i) 173 Uses[i].push_back(&ExitSU); 174 } 175} 176 177void ScheduleDAGInstrs::BuildSchedGraph(AliasAnalysis *AA) { 178 // We'll be allocating one SUnit for each instruction, plus one for 179 // the region exit node. 180 SUnits.reserve(BB->size()); 181 182 // We build scheduling units by walking a block's instruction list from bottom 183 // to top. 184 185 // Remember where a generic side-effecting instruction is as we procede. 186 SUnit *BarrierChain = 0, *AliasChain = 0; 187 188 // Memory references to specific known memory locations are tracked 189 // so that they can be given more precise dependencies. We track 190 // separately the known memory locations that may alias and those 191 // that are known not to alias 192 std::map<const Value *, SUnit *> AliasMemDefs, NonAliasMemDefs; 193 std::map<const Value *, std::vector<SUnit *> > AliasMemUses, NonAliasMemUses; 194 195 // Keep track of dangling debug references to registers. 196 std::vector<std::pair<MachineInstr*, unsigned> > 197 DanglingDebugValue(TRI->getNumRegs(), 198 std::make_pair(static_cast<MachineInstr*>(0), 0)); 199 200 // Check to see if the scheduler cares about latencies. 201 bool UnitLatencies = ForceUnitLatencies(); 202 203 // Ask the target if address-backscheduling is desirable, and if so how much. 204 const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>(); 205 unsigned SpecialAddressLatency = ST.getSpecialAddressLatency(); 206 207 // Remove any stale debug info; sometimes BuildSchedGraph is called again 208 // without emitting the info from the previous call. 209 DbgValueVec.clear(); 210 211 // Model data dependencies between instructions being scheduled and the 212 // ExitSU. 213 AddSchedBarrierDeps(); 214 215 // Walk the list of instructions, from bottom moving up. 216 for (MachineBasicBlock::iterator MII = InsertPos, MIE = Begin; 217 MII != MIE; --MII) { 218 MachineInstr *MI = prior(MII); 219 // DBG_VALUE does not have SUnit's built, so just remember these for later 220 // reinsertion. 221 if (MI->isDebugValue()) { 222 if (MI->getNumOperands()==3 && MI->getOperand(0).isReg() && 223 MI->getOperand(0).getReg()) 224 DanglingDebugValue[MI->getOperand(0).getReg()] = 225 std::make_pair(MI, DbgValueVec.size()); 226 DbgValueVec.push_back(MI); 227 continue; 228 } 229 const TargetInstrDesc &TID = MI->getDesc(); 230 assert(!TID.isTerminator() && !MI->isLabel() && 231 "Cannot schedule terminators or labels!"); 232 // Create the SUnit for this MI. 233 SUnit *SU = NewSUnit(MI); 234 235 // Assign the Latency field of SU using target-provided information. 236 if (UnitLatencies) 237 SU->Latency = 1; 238 else 239 ComputeLatency(SU); 240 241 // Add register-based dependencies (data, anti, and output). 242 for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) { 243 const MachineOperand &MO = MI->getOperand(j); 244 if (!MO.isReg()) continue; 245 unsigned Reg = MO.getReg(); 246 if (Reg == 0) continue; 247 248 assert(TRI->isPhysicalRegister(Reg) && "Virtual register encountered!"); 249 250 if (MO.isDef() && DanglingDebugValue[Reg].first!=0) { 251 SU->DbgInstrList.push_back(DanglingDebugValue[Reg].first); 252 DbgValueVec[DanglingDebugValue[Reg].second] = 0; 253 DanglingDebugValue[Reg] = std::make_pair((MachineInstr*)0, 0); 254 } 255 256 std::vector<SUnit *> &UseList = Uses[Reg]; 257 std::vector<SUnit *> &DefList = Defs[Reg]; 258 // Optionally add output and anti dependencies. For anti 259 // dependencies we use a latency of 0 because for a multi-issue 260 // target we want to allow the defining instruction to issue 261 // in the same cycle as the using instruction. 262 // TODO: Using a latency of 1 here for output dependencies assumes 263 // there's no cost for reusing registers. 264 SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output; 265 unsigned AOLatency = (Kind == SDep::Anti) ? 0 : 1; 266 for (unsigned i = 0, e = DefList.size(); i != e; ++i) { 267 SUnit *DefSU = DefList[i]; 268 if (DefSU == &ExitSU) 269 continue; 270 if (DefSU != SU && 271 (Kind != SDep::Output || !MO.isDead() || 272 !DefSU->getInstr()->registerDefIsDead(Reg))) 273 DefSU->addPred(SDep(SU, Kind, AOLatency, /*Reg=*/Reg)); 274 } 275 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) { 276 std::vector<SUnit *> &DefList = Defs[*Alias]; 277 for (unsigned i = 0, e = DefList.size(); i != e; ++i) { 278 SUnit *DefSU = DefList[i]; 279 if (DefSU == &ExitSU) 280 continue; 281 if (DefSU != SU && 282 (Kind != SDep::Output || !MO.isDead() || 283 !DefSU->getInstr()->registerDefIsDead(*Alias))) 284 DefSU->addPred(SDep(SU, Kind, AOLatency, /*Reg=*/ *Alias)); 285 } 286 } 287 288 if (MO.isDef()) { 289 // Add any data dependencies. 290 unsigned DataLatency = SU->Latency; 291 for (unsigned i = 0, e = UseList.size(); i != e; ++i) { 292 SUnit *UseSU = UseList[i]; 293 if (UseSU == SU) 294 continue; 295 unsigned LDataLatency = DataLatency; 296 // Optionally add in a special extra latency for nodes that 297 // feed addresses. 298 // TODO: Do this for register aliases too. 299 // TODO: Perhaps we should get rid of 300 // SpecialAddressLatency and just move this into 301 // adjustSchedDependency for the targets that care about it. 302 if (SpecialAddressLatency != 0 && !UnitLatencies && 303 UseSU != &ExitSU) { 304 MachineInstr *UseMI = UseSU->getInstr(); 305 const TargetInstrDesc &UseTID = UseMI->getDesc(); 306 int RegUseIndex = UseMI->findRegisterUseOperandIdx(Reg); 307 assert(RegUseIndex >= 0 && "UseMI doesn's use register!"); 308 if (RegUseIndex >= 0 && 309 (UseTID.mayLoad() || UseTID.mayStore()) && 310 (unsigned)RegUseIndex < UseTID.getNumOperands() && 311 UseTID.OpInfo[RegUseIndex].isLookupPtrRegClass()) 312 LDataLatency += SpecialAddressLatency; 313 } 314 // Adjust the dependence latency using operand def/use 315 // information (if any), and then allow the target to 316 // perform its own adjustments. 317 const SDep& dep = SDep(SU, SDep::Data, LDataLatency, Reg); 318 if (!UnitLatencies) { 319 ComputeOperandLatency(SU, UseSU, const_cast<SDep &>(dep)); 320 ST.adjustSchedDependency(SU, UseSU, const_cast<SDep &>(dep)); 321 } 322 UseSU->addPred(dep); 323 } 324 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) { 325 std::vector<SUnit *> &UseList = Uses[*Alias]; 326 for (unsigned i = 0, e = UseList.size(); i != e; ++i) { 327 SUnit *UseSU = UseList[i]; 328 if (UseSU == SU) 329 continue; 330 const SDep& dep = SDep(SU, SDep::Data, DataLatency, *Alias); 331 if (!UnitLatencies) { 332 ComputeOperandLatency(SU, UseSU, const_cast<SDep &>(dep)); 333 ST.adjustSchedDependency(SU, UseSU, const_cast<SDep &>(dep)); 334 } 335 UseSU->addPred(dep); 336 } 337 } 338 339 // If a def is going to wrap back around to the top of the loop, 340 // backschedule it. 341 if (!UnitLatencies && DefList.empty()) { 342 LoopDependencies::LoopDeps::iterator I = LoopRegs.Deps.find(Reg); 343 if (I != LoopRegs.Deps.end()) { 344 const MachineOperand *UseMO = I->second.first; 345 unsigned Count = I->second.second; 346 const MachineInstr *UseMI = UseMO->getParent(); 347 unsigned UseMOIdx = UseMO - &UseMI->getOperand(0); 348 const TargetInstrDesc &UseTID = UseMI->getDesc(); 349 // TODO: If we knew the total depth of the region here, we could 350 // handle the case where the whole loop is inside the region but 351 // is large enough that the isScheduleHigh trick isn't needed. 352 if (UseMOIdx < UseTID.getNumOperands()) { 353 // Currently, we only support scheduling regions consisting of 354 // single basic blocks. Check to see if the instruction is in 355 // the same region by checking to see if it has the same parent. 356 if (UseMI->getParent() != MI->getParent()) { 357 unsigned Latency = SU->Latency; 358 if (UseTID.OpInfo[UseMOIdx].isLookupPtrRegClass()) 359 Latency += SpecialAddressLatency; 360 // This is a wild guess as to the portion of the latency which 361 // will be overlapped by work done outside the current 362 // scheduling region. 363 Latency -= std::min(Latency, Count); 364 // Add the artifical edge. 365 ExitSU.addPred(SDep(SU, SDep::Order, Latency, 366 /*Reg=*/0, /*isNormalMemory=*/false, 367 /*isMustAlias=*/false, 368 /*isArtificial=*/true)); 369 } else if (SpecialAddressLatency > 0 && 370 UseTID.OpInfo[UseMOIdx].isLookupPtrRegClass()) { 371 // The entire loop body is within the current scheduling region 372 // and the latency of this operation is assumed to be greater 373 // than the latency of the loop. 374 // TODO: Recursively mark data-edge predecessors as 375 // isScheduleHigh too. 376 SU->isScheduleHigh = true; 377 } 378 } 379 LoopRegs.Deps.erase(I); 380 } 381 } 382 383 UseList.clear(); 384 if (!MO.isDead()) 385 DefList.clear(); 386 DefList.push_back(SU); 387 } else { 388 UseList.push_back(SU); 389 } 390 } 391 392 // Add chain dependencies. 393 // Chain dependencies used to enforce memory order should have 394 // latency of 0 (except for true dependency of Store followed by 395 // aliased Load... we estimate that with a single cycle of latency 396 // assuming the hardware will bypass) 397 // Note that isStoreToStackSlot and isLoadFromStackSLot are not usable 398 // after stack slots are lowered to actual addresses. 399 // TODO: Use an AliasAnalysis and do real alias-analysis queries, and 400 // produce more precise dependence information. 401#define STORE_LOAD_LATENCY 1 402 unsigned TrueMemOrderLatency = 0; 403 if (TID.isCall() || TID.hasUnmodeledSideEffects() || 404 (MI->hasVolatileMemoryRef() && 405 (!TID.mayLoad() || !MI->isInvariantLoad(AA)))) { 406 // Be conservative with these and add dependencies on all memory 407 // references, even those that are known to not alias. 408 for (std::map<const Value *, SUnit *>::iterator I = 409 NonAliasMemDefs.begin(), E = NonAliasMemDefs.end(); I != E; ++I) { 410 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0)); 411 } 412 for (std::map<const Value *, std::vector<SUnit *> >::iterator I = 413 NonAliasMemUses.begin(), E = NonAliasMemUses.end(); I != E; ++I) { 414 for (unsigned i = 0, e = I->second.size(); i != e; ++i) 415 I->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency)); 416 } 417 NonAliasMemDefs.clear(); 418 NonAliasMemUses.clear(); 419 // Add SU to the barrier chain. 420 if (BarrierChain) 421 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0)); 422 BarrierChain = SU; 423 424 // fall-through 425 new_alias_chain: 426 // Chain all possibly aliasing memory references though SU. 427 if (AliasChain) 428 AliasChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0)); 429 AliasChain = SU; 430 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k) 431 PendingLoads[k]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency)); 432 for (std::map<const Value *, SUnit *>::iterator I = AliasMemDefs.begin(), 433 E = AliasMemDefs.end(); I != E; ++I) { 434 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0)); 435 } 436 for (std::map<const Value *, std::vector<SUnit *> >::iterator I = 437 AliasMemUses.begin(), E = AliasMemUses.end(); I != E; ++I) { 438 for (unsigned i = 0, e = I->second.size(); i != e; ++i) 439 I->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency)); 440 } 441 PendingLoads.clear(); 442 AliasMemDefs.clear(); 443 AliasMemUses.clear(); 444 } else if (TID.mayStore()) { 445 bool MayAlias = true; 446 TrueMemOrderLatency = STORE_LOAD_LATENCY; 447 if (const Value *V = getUnderlyingObjectForInstr(MI, MFI, MayAlias)) { 448 // A store to a specific PseudoSourceValue. Add precise dependencies. 449 // Record the def in MemDefs, first adding a dep if there is 450 // an existing def. 451 std::map<const Value *, SUnit *>::iterator I = 452 ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V)); 453 std::map<const Value *, SUnit *>::iterator IE = 454 ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end()); 455 if (I != IE) { 456 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0, /*Reg=*/0, 457 /*isNormalMemory=*/true)); 458 I->second = SU; 459 } else { 460 if (MayAlias) 461 AliasMemDefs[V] = SU; 462 else 463 NonAliasMemDefs[V] = SU; 464 } 465 // Handle the uses in MemUses, if there are any. 466 std::map<const Value *, std::vector<SUnit *> >::iterator J = 467 ((MayAlias) ? AliasMemUses.find(V) : NonAliasMemUses.find(V)); 468 std::map<const Value *, std::vector<SUnit *> >::iterator JE = 469 ((MayAlias) ? AliasMemUses.end() : NonAliasMemUses.end()); 470 if (J != JE) { 471 for (unsigned i = 0, e = J->second.size(); i != e; ++i) 472 J->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency, 473 /*Reg=*/0, /*isNormalMemory=*/true)); 474 J->second.clear(); 475 } 476 if (MayAlias) { 477 // Add dependencies from all the PendingLoads, i.e. loads 478 // with no underlying object. 479 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k) 480 PendingLoads[k]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency)); 481 // Add dependence on alias chain, if needed. 482 if (AliasChain) 483 AliasChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0)); 484 } 485 // Add dependence on barrier chain, if needed. 486 if (BarrierChain) 487 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0)); 488 } else { 489 // Treat all other stores conservatively. 490 goto new_alias_chain; 491 } 492 493 if (!ExitSU.isPred(SU)) 494 // Push store's up a bit to avoid them getting in between cmp 495 // and branches. 496 ExitSU.addPred(SDep(SU, SDep::Order, 0, 497 /*Reg=*/0, /*isNormalMemory=*/false, 498 /*isMustAlias=*/false, 499 /*isArtificial=*/true)); 500 } else if (TID.mayLoad()) { 501 bool MayAlias = true; 502 TrueMemOrderLatency = 0; 503 if (MI->isInvariantLoad(AA)) { 504 // Invariant load, no chain dependencies needed! 505 } else { 506 if (const Value *V = 507 getUnderlyingObjectForInstr(MI, MFI, MayAlias)) { 508 // A load from a specific PseudoSourceValue. Add precise dependencies. 509 std::map<const Value *, SUnit *>::iterator I = 510 ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V)); 511 std::map<const Value *, SUnit *>::iterator IE = 512 ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end()); 513 if (I != IE) 514 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0, /*Reg=*/0, 515 /*isNormalMemory=*/true)); 516 if (MayAlias) 517 AliasMemUses[V].push_back(SU); 518 else 519 NonAliasMemUses[V].push_back(SU); 520 } else { 521 // A load with no underlying object. Depend on all 522 // potentially aliasing stores. 523 for (std::map<const Value *, SUnit *>::iterator I = 524 AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I) 525 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0)); 526 527 PendingLoads.push_back(SU); 528 MayAlias = true; 529 } 530 531 // Add dependencies on alias and barrier chains, if needed. 532 if (MayAlias && AliasChain) 533 AliasChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0)); 534 if (BarrierChain) 535 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0)); 536 } 537 } 538 } 539 540 for (int i = 0, e = TRI->getNumRegs(); i != e; ++i) { 541 Defs[i].clear(); 542 Uses[i].clear(); 543 } 544 PendingLoads.clear(); 545} 546 547void ScheduleDAGInstrs::FinishBlock() { 548 // Nothing to do. 549} 550 551void ScheduleDAGInstrs::ComputeLatency(SUnit *SU) { 552 // Compute the latency for the node. 553 if (!InstrItins || InstrItins->isEmpty()) { 554 SU->Latency = 1; 555 556 // Simplistic target-independent heuristic: assume that loads take 557 // extra time. 558 if (SU->getInstr()->getDesc().mayLoad()) 559 SU->Latency += 2; 560 } else 561 SU->Latency = 562 InstrItins->getStageLatency(SU->getInstr()->getDesc().getSchedClass()); 563} 564 565void ScheduleDAGInstrs::ComputeOperandLatency(SUnit *Def, SUnit *Use, 566 SDep& dep) const { 567 if (!InstrItins || InstrItins->isEmpty()) 568 return; 569 570 // For a data dependency with a known register... 571 if ((dep.getKind() != SDep::Data) || (dep.getReg() == 0)) 572 return; 573 574 const unsigned Reg = dep.getReg(); 575 576 // ... find the definition of the register in the defining 577 // instruction 578 MachineInstr *DefMI = Def->getInstr(); 579 int DefIdx = DefMI->findRegisterDefOperandIdx(Reg); 580 if (DefIdx != -1) { 581 const MachineOperand &MO = DefMI->getOperand(DefIdx); 582 if (MO.isReg() && MO.isImplicit() && 583 DefIdx >= (int)DefMI->getDesc().getNumOperands()) { 584 // This is an implicit def, getOperandLatency() won't return the correct 585 // latency. e.g. 586 // %D6<def>, %D7<def> = VLD1q16 %R2<kill>, 0, ..., %Q3<imp-def> 587 // %Q1<def> = VMULv8i16 %Q1<kill>, %Q3<kill>, ... 588 // What we want is to compute latency between def of %D6/%D7 and use of 589 // %Q3 instead. 590 DefIdx = DefMI->findRegisterDefOperandIdx(Reg, false, true, TRI); 591 } 592 MachineInstr *UseMI = Use->getInstr(); 593 // For all uses of the register, calculate the maxmimum latency 594 int Latency = -1; 595 if (UseMI) { 596 for (unsigned i = 0, e = UseMI->getNumOperands(); i != e; ++i) { 597 const MachineOperand &MO = UseMI->getOperand(i); 598 if (!MO.isReg() || !MO.isUse()) 599 continue; 600 unsigned MOReg = MO.getReg(); 601 if (MOReg != Reg) 602 continue; 603 604 int UseCycle = TII->getOperandLatency(InstrItins, DefMI, DefIdx, 605 UseMI, i); 606 Latency = std::max(Latency, UseCycle); 607 } 608 } else { 609 // UseMI is null, then it must be a scheduling barrier. 610 if (!InstrItins || InstrItins->isEmpty()) 611 return; 612 unsigned DefClass = DefMI->getDesc().getSchedClass(); 613 Latency = InstrItins->getOperandCycle(DefClass, DefIdx); 614 } 615 616 // If we found a latency, then replace the existing dependence latency. 617 if (Latency >= 0) 618 dep.setLatency(Latency); 619 } 620} 621 622void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const { 623 SU->getInstr()->dump(); 624} 625 626std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const { 627 std::string s; 628 raw_string_ostream oss(s); 629 if (SU == &EntrySU) 630 oss << "<entry>"; 631 else if (SU == &ExitSU) 632 oss << "<exit>"; 633 else 634 SU->getInstr()->print(oss); 635 return oss.str(); 636} 637 638// EmitSchedule - Emit the machine code in scheduled order. 639MachineBasicBlock *ScheduleDAGInstrs::EmitSchedule() { 640 // For MachineInstr-based scheduling, we're rescheduling the instructions in 641 // the block, so start by removing them from the block. 642 while (Begin != InsertPos) { 643 MachineBasicBlock::iterator I = Begin; 644 ++Begin; 645 BB->remove(I); 646 } 647 648 // First reinsert any remaining debug_values; these are either constants, 649 // or refer to live-in registers. The beginning of the block is the right 650 // place for the latter. The former might reasonably be placed elsewhere 651 // using some kind of ordering algorithm, but right now it doesn't matter. 652 for (int i = DbgValueVec.size()-1; i>=0; --i) 653 if (DbgValueVec[i]) 654 BB->insert(InsertPos, DbgValueVec[i]); 655 656 // Then re-insert them according to the given schedule. 657 for (unsigned i = 0, e = Sequence.size(); i != e; i++) { 658 SUnit *SU = Sequence[i]; 659 if (!SU) { 660 // Null SUnit* is a noop. 661 EmitNoop(); 662 continue; 663 } 664 665 BB->insert(InsertPos, SU->getInstr()); 666 for (unsigned i = 0, e = SU->DbgInstrList.size() ; i < e ; ++i) 667 BB->insert(InsertPos, SU->DbgInstrList[i]); 668 } 669 670 // Update the Begin iterator, as the first instruction in the block 671 // may have been scheduled later. 672 if (!DbgValueVec.empty()) { 673 for (int i = DbgValueVec.size()-1; i>=0; --i) 674 if (DbgValueVec[i]!=0) { 675 Begin = DbgValueVec[DbgValueVec.size()-1]; 676 break; 677 } 678 } else if (!Sequence.empty()) 679 Begin = Sequence[0]->getInstr(); 680 681 DbgValueVec.clear(); 682 return BB; 683} 684