1//===--- ScheduleDAGSDNodes.cpp - Implement the ScheduleDAGSDNodes class --===// 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 ScheduleDAG class, which is a base class used by 11// scheduling implementation classes. 12// 13//===----------------------------------------------------------------------===// 14 15#include "ScheduleDAGSDNodes.h" 16#include "InstrEmitter.h" 17#include "SDNodeDbgValue.h" 18#include "llvm/ADT/DenseMap.h" 19#include "llvm/ADT/SmallPtrSet.h" 20#include "llvm/ADT/SmallSet.h" 21#include "llvm/ADT/SmallVector.h" 22#include "llvm/ADT/Statistic.h" 23#include "llvm/CodeGen/MachineInstrBuilder.h" 24#include "llvm/CodeGen/MachineRegisterInfo.h" 25#include "llvm/CodeGen/SelectionDAG.h" 26#include "llvm/MC/MCInstrItineraries.h" 27#include "llvm/Support/CommandLine.h" 28#include "llvm/Support/Debug.h" 29#include "llvm/Support/raw_ostream.h" 30#include "llvm/Target/TargetInstrInfo.h" 31#include "llvm/Target/TargetLowering.h" 32#include "llvm/Target/TargetRegisterInfo.h" 33#include "llvm/Target/TargetSubtargetInfo.h" 34using namespace llvm; 35 36#define DEBUG_TYPE "pre-RA-sched" 37 38STATISTIC(LoadsClustered, "Number of loads clustered together"); 39 40// This allows the latency-based scheduler to notice high latency instructions 41// without a target itinerary. The choice of number here has more to do with 42// balancing scheduler heuristics than with the actual machine latency. 43static cl::opt<int> HighLatencyCycles( 44 "sched-high-latency-cycles", cl::Hidden, cl::init(10), 45 cl::desc("Roughly estimate the number of cycles that 'long latency'" 46 "instructions take for targets with no itinerary")); 47 48ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf) 49 : ScheduleDAG(mf), BB(nullptr), DAG(nullptr), 50 InstrItins(mf.getSubtarget().getInstrItineraryData()) {} 51 52/// Run - perform scheduling. 53/// 54void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb) { 55 BB = bb; 56 DAG = dag; 57 58 // Clear the scheduler's SUnit DAG. 59 ScheduleDAG::clearDAG(); 60 Sequence.clear(); 61 62 // Invoke the target's selection of scheduler. 63 Schedule(); 64} 65 66/// NewSUnit - Creates a new SUnit and return a ptr to it. 67/// 68SUnit *ScheduleDAGSDNodes::newSUnit(SDNode *N) { 69#ifndef NDEBUG 70 const SUnit *Addr = nullptr; 71 if (!SUnits.empty()) 72 Addr = &SUnits[0]; 73#endif 74 SUnits.emplace_back(N, (unsigned)SUnits.size()); 75 assert((Addr == nullptr || Addr == &SUnits[0]) && 76 "SUnits std::vector reallocated on the fly!"); 77 SUnits.back().OrigNode = &SUnits.back(); 78 SUnit *SU = &SUnits.back(); 79 const TargetLowering &TLI = DAG->getTargetLoweringInfo(); 80 if (!N || 81 (N->isMachineOpcode() && 82 N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF)) 83 SU->SchedulingPref = Sched::None; 84 else 85 SU->SchedulingPref = TLI.getSchedulingPreference(N); 86 return SU; 87} 88 89SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) { 90 SUnit *SU = newSUnit(Old->getNode()); 91 SU->OrigNode = Old->OrigNode; 92 SU->Latency = Old->Latency; 93 SU->isVRegCycle = Old->isVRegCycle; 94 SU->isCall = Old->isCall; 95 SU->isCallOp = Old->isCallOp; 96 SU->isTwoAddress = Old->isTwoAddress; 97 SU->isCommutable = Old->isCommutable; 98 SU->hasPhysRegDefs = Old->hasPhysRegDefs; 99 SU->hasPhysRegClobbers = Old->hasPhysRegClobbers; 100 SU->isScheduleHigh = Old->isScheduleHigh; 101 SU->isScheduleLow = Old->isScheduleLow; 102 SU->SchedulingPref = Old->SchedulingPref; 103 Old->isCloned = true; 104 return SU; 105} 106 107/// CheckForPhysRegDependency - Check if the dependency between def and use of 108/// a specified operand is a physical register dependency. If so, returns the 109/// register and the cost of copying the register. 110static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op, 111 const TargetRegisterInfo *TRI, 112 const TargetInstrInfo *TII, 113 unsigned &PhysReg, int &Cost) { 114 if (Op != 2 || User->getOpcode() != ISD::CopyToReg) 115 return; 116 117 unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg(); 118 if (TargetRegisterInfo::isVirtualRegister(Reg)) 119 return; 120 121 unsigned ResNo = User->getOperand(2).getResNo(); 122 if (Def->getOpcode() == ISD::CopyFromReg && 123 cast<RegisterSDNode>(Def->getOperand(1))->getReg() == Reg) { 124 PhysReg = Reg; 125 } else if (Def->isMachineOpcode()) { 126 const MCInstrDesc &II = TII->get(Def->getMachineOpcode()); 127 if (ResNo >= II.getNumDefs() && 128 II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) 129 PhysReg = Reg; 130 } 131 132 if (PhysReg != 0) { 133 const TargetRegisterClass *RC = 134 TRI->getMinimalPhysRegClass(Reg, Def->getSimpleValueType(ResNo)); 135 Cost = RC->getCopyCost(); 136 } 137} 138 139// Helper for AddGlue to clone node operands. 140static void CloneNodeWithValues(SDNode *N, SelectionDAG *DAG, ArrayRef<EVT> VTs, 141 SDValue ExtraOper = SDValue()) { 142 SmallVector<SDValue, 8> Ops(N->op_begin(), N->op_end()); 143 if (ExtraOper.getNode()) 144 Ops.push_back(ExtraOper); 145 146 SDVTList VTList = DAG->getVTList(VTs); 147 MachineSDNode::mmo_iterator Begin = nullptr, End = nullptr; 148 MachineSDNode *MN = dyn_cast<MachineSDNode>(N); 149 150 // Store memory references. 151 if (MN) { 152 Begin = MN->memoperands_begin(); 153 End = MN->memoperands_end(); 154 } 155 156 DAG->MorphNodeTo(N, N->getOpcode(), VTList, Ops); 157 158 // Reset the memory references 159 if (MN) 160 MN->setMemRefs(Begin, End); 161} 162 163static bool AddGlue(SDNode *N, SDValue Glue, bool AddGlue, SelectionDAG *DAG) { 164 SDNode *GlueDestNode = Glue.getNode(); 165 166 // Don't add glue from a node to itself. 167 if (GlueDestNode == N) return false; 168 169 // Don't add a glue operand to something that already uses glue. 170 if (GlueDestNode && 171 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) { 172 return false; 173 } 174 // Don't add glue to something that already has a glue value. 175 if (N->getValueType(N->getNumValues() - 1) == MVT::Glue) return false; 176 177 SmallVector<EVT, 4> VTs(N->value_begin(), N->value_end()); 178 if (AddGlue) 179 VTs.push_back(MVT::Glue); 180 181 CloneNodeWithValues(N, DAG, VTs, Glue); 182 183 return true; 184} 185 186// Cleanup after unsuccessful AddGlue. Use the standard method of morphing the 187// node even though simply shrinking the value list is sufficient. 188static void RemoveUnusedGlue(SDNode *N, SelectionDAG *DAG) { 189 assert((N->getValueType(N->getNumValues() - 1) == MVT::Glue && 190 !N->hasAnyUseOfValue(N->getNumValues() - 1)) && 191 "expected an unused glue value"); 192 193 CloneNodeWithValues(N, DAG, 194 makeArrayRef(N->value_begin(), N->getNumValues() - 1)); 195} 196 197/// ClusterNeighboringLoads - Force nearby loads together by "gluing" them. 198/// This function finds loads of the same base and different offsets. If the 199/// offsets are not far apart (target specific), it add MVT::Glue inputs and 200/// outputs to ensure they are scheduled together and in order. This 201/// optimization may benefit some targets by improving cache locality. 202void ScheduleDAGSDNodes::ClusterNeighboringLoads(SDNode *Node) { 203 SDNode *Chain = nullptr; 204 unsigned NumOps = Node->getNumOperands(); 205 if (Node->getOperand(NumOps-1).getValueType() == MVT::Other) 206 Chain = Node->getOperand(NumOps-1).getNode(); 207 if (!Chain) 208 return; 209 210 // Look for other loads of the same chain. Find loads that are loading from 211 // the same base pointer and different offsets. 212 SmallPtrSet<SDNode*, 16> Visited; 213 SmallVector<int64_t, 4> Offsets; 214 DenseMap<long long, SDNode*> O2SMap; // Map from offset to SDNode. 215 bool Cluster = false; 216 SDNode *Base = Node; 217 // This algorithm requires a reasonably low use count before finding a match 218 // to avoid uselessly blowing up compile time in large blocks. 219 unsigned UseCount = 0; 220 for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end(); 221 I != E && UseCount < 100; ++I, ++UseCount) { 222 SDNode *User = *I; 223 if (User == Node || !Visited.insert(User).second) 224 continue; 225 int64_t Offset1, Offset2; 226 if (!TII->areLoadsFromSameBasePtr(Base, User, Offset1, Offset2) || 227 Offset1 == Offset2) 228 // FIXME: Should be ok if they addresses are identical. But earlier 229 // optimizations really should have eliminated one of the loads. 230 continue; 231 if (O2SMap.insert(std::make_pair(Offset1, Base)).second) 232 Offsets.push_back(Offset1); 233 O2SMap.insert(std::make_pair(Offset2, User)); 234 Offsets.push_back(Offset2); 235 if (Offset2 < Offset1) 236 Base = User; 237 Cluster = true; 238 // Reset UseCount to allow more matches. 239 UseCount = 0; 240 } 241 242 if (!Cluster) 243 return; 244 245 // Sort them in increasing order. 246 std::sort(Offsets.begin(), Offsets.end()); 247 248 // Check if the loads are close enough. 249 SmallVector<SDNode*, 4> Loads; 250 unsigned NumLoads = 0; 251 int64_t BaseOff = Offsets[0]; 252 SDNode *BaseLoad = O2SMap[BaseOff]; 253 Loads.push_back(BaseLoad); 254 for (unsigned i = 1, e = Offsets.size(); i != e; ++i) { 255 int64_t Offset = Offsets[i]; 256 SDNode *Load = O2SMap[Offset]; 257 if (!TII->shouldScheduleLoadsNear(BaseLoad, Load, BaseOff, Offset,NumLoads)) 258 break; // Stop right here. Ignore loads that are further away. 259 Loads.push_back(Load); 260 ++NumLoads; 261 } 262 263 if (NumLoads == 0) 264 return; 265 266 // Cluster loads by adding MVT::Glue outputs and inputs. This also 267 // ensure they are scheduled in order of increasing addresses. 268 SDNode *Lead = Loads[0]; 269 SDValue InGlue = SDValue(nullptr, 0); 270 if (AddGlue(Lead, InGlue, true, DAG)) 271 InGlue = SDValue(Lead, Lead->getNumValues() - 1); 272 for (unsigned I = 1, E = Loads.size(); I != E; ++I) { 273 bool OutGlue = I < E - 1; 274 SDNode *Load = Loads[I]; 275 276 // If AddGlue fails, we could leave an unsused glue value. This should not 277 // cause any 278 if (AddGlue(Load, InGlue, OutGlue, DAG)) { 279 if (OutGlue) 280 InGlue = SDValue(Load, Load->getNumValues() - 1); 281 282 ++LoadsClustered; 283 } 284 else if (!OutGlue && InGlue.getNode()) 285 RemoveUnusedGlue(InGlue.getNode(), DAG); 286 } 287} 288 289/// ClusterNodes - Cluster certain nodes which should be scheduled together. 290/// 291void ScheduleDAGSDNodes::ClusterNodes() { 292 for (SDNode &NI : DAG->allnodes()) { 293 SDNode *Node = &NI; 294 if (!Node || !Node->isMachineOpcode()) 295 continue; 296 297 unsigned Opc = Node->getMachineOpcode(); 298 const MCInstrDesc &MCID = TII->get(Opc); 299 if (MCID.mayLoad()) 300 // Cluster loads from "near" addresses into combined SUnits. 301 ClusterNeighboringLoads(Node); 302 } 303} 304 305void ScheduleDAGSDNodes::BuildSchedUnits() { 306 // During scheduling, the NodeId field of SDNode is used to map SDNodes 307 // to their associated SUnits by holding SUnits table indices. A value 308 // of -1 means the SDNode does not yet have an associated SUnit. 309 unsigned NumNodes = 0; 310 for (SDNode &NI : DAG->allnodes()) { 311 NI.setNodeId(-1); 312 ++NumNodes; 313 } 314 315 // Reserve entries in the vector for each of the SUnits we are creating. This 316 // ensure that reallocation of the vector won't happen, so SUnit*'s won't get 317 // invalidated. 318 // FIXME: Multiply by 2 because we may clone nodes during scheduling. 319 // This is a temporary workaround. 320 SUnits.reserve(NumNodes * 2); 321 322 // Add all nodes in depth first order. 323 SmallVector<SDNode*, 64> Worklist; 324 SmallPtrSet<SDNode*, 64> Visited; 325 Worklist.push_back(DAG->getRoot().getNode()); 326 Visited.insert(DAG->getRoot().getNode()); 327 328 SmallVector<SUnit*, 8> CallSUnits; 329 while (!Worklist.empty()) { 330 SDNode *NI = Worklist.pop_back_val(); 331 332 // Add all operands to the worklist unless they've already been added. 333 for (const SDValue &Op : NI->op_values()) 334 if (Visited.insert(Op.getNode()).second) 335 Worklist.push_back(Op.getNode()); 336 337 if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate. 338 continue; 339 340 // If this node has already been processed, stop now. 341 if (NI->getNodeId() != -1) continue; 342 343 SUnit *NodeSUnit = newSUnit(NI); 344 345 // See if anything is glued to this node, if so, add them to glued 346 // nodes. Nodes can have at most one glue input and one glue output. Glue 347 // is required to be the last operand and result of a node. 348 349 // Scan up to find glued preds. 350 SDNode *N = NI; 351 while (N->getNumOperands() && 352 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) { 353 N = N->getOperand(N->getNumOperands()-1).getNode(); 354 assert(N->getNodeId() == -1 && "Node already inserted!"); 355 N->setNodeId(NodeSUnit->NodeNum); 356 if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall()) 357 NodeSUnit->isCall = true; 358 } 359 360 // Scan down to find any glued succs. 361 N = NI; 362 while (N->getValueType(N->getNumValues()-1) == MVT::Glue) { 363 SDValue GlueVal(N, N->getNumValues()-1); 364 365 // There are either zero or one users of the Glue result. 366 bool HasGlueUse = false; 367 for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end(); 368 UI != E; ++UI) 369 if (GlueVal.isOperandOf(*UI)) { 370 HasGlueUse = true; 371 assert(N->getNodeId() == -1 && "Node already inserted!"); 372 N->setNodeId(NodeSUnit->NodeNum); 373 N = *UI; 374 if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall()) 375 NodeSUnit->isCall = true; 376 break; 377 } 378 if (!HasGlueUse) break; 379 } 380 381 if (NodeSUnit->isCall) 382 CallSUnits.push_back(NodeSUnit); 383 384 // Schedule zero-latency TokenFactor below any nodes that may increase the 385 // schedule height. Otherwise, ancestors of the TokenFactor may appear to 386 // have false stalls. 387 if (NI->getOpcode() == ISD::TokenFactor) 388 NodeSUnit->isScheduleLow = true; 389 390 // If there are glue operands involved, N is now the bottom-most node 391 // of the sequence of nodes that are glued together. 392 // Update the SUnit. 393 NodeSUnit->setNode(N); 394 assert(N->getNodeId() == -1 && "Node already inserted!"); 395 N->setNodeId(NodeSUnit->NodeNum); 396 397 // Compute NumRegDefsLeft. This must be done before AddSchedEdges. 398 InitNumRegDefsLeft(NodeSUnit); 399 400 // Assign the Latency field of NodeSUnit using target-provided information. 401 computeLatency(NodeSUnit); 402 } 403 404 // Find all call operands. 405 while (!CallSUnits.empty()) { 406 SUnit *SU = CallSUnits.pop_back_val(); 407 for (const SDNode *SUNode = SU->getNode(); SUNode; 408 SUNode = SUNode->getGluedNode()) { 409 if (SUNode->getOpcode() != ISD::CopyToReg) 410 continue; 411 SDNode *SrcN = SUNode->getOperand(2).getNode(); 412 if (isPassiveNode(SrcN)) continue; // Not scheduled. 413 SUnit *SrcSU = &SUnits[SrcN->getNodeId()]; 414 SrcSU->isCallOp = true; 415 } 416 } 417} 418 419void ScheduleDAGSDNodes::AddSchedEdges() { 420 const TargetSubtargetInfo &ST = MF.getSubtarget(); 421 422 // Check to see if the scheduler cares about latencies. 423 bool UnitLatencies = forceUnitLatencies(); 424 425 // Pass 2: add the preds, succs, etc. 426 for (unsigned su = 0, e = SUnits.size(); su != e; ++su) { 427 SUnit *SU = &SUnits[su]; 428 SDNode *MainNode = SU->getNode(); 429 430 if (MainNode->isMachineOpcode()) { 431 unsigned Opc = MainNode->getMachineOpcode(); 432 const MCInstrDesc &MCID = TII->get(Opc); 433 for (unsigned i = 0; i != MCID.getNumOperands(); ++i) { 434 if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) { 435 SU->isTwoAddress = true; 436 break; 437 } 438 } 439 if (MCID.isCommutable()) 440 SU->isCommutable = true; 441 } 442 443 // Find all predecessors and successors of the group. 444 for (SDNode *N = SU->getNode(); N; N = N->getGluedNode()) { 445 if (N->isMachineOpcode() && 446 TII->get(N->getMachineOpcode()).getImplicitDefs()) { 447 SU->hasPhysRegClobbers = true; 448 unsigned NumUsed = InstrEmitter::CountResults(N); 449 while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1)) 450 --NumUsed; // Skip over unused values at the end. 451 if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs()) 452 SU->hasPhysRegDefs = true; 453 } 454 455 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 456 SDNode *OpN = N->getOperand(i).getNode(); 457 if (isPassiveNode(OpN)) continue; // Not scheduled. 458 SUnit *OpSU = &SUnits[OpN->getNodeId()]; 459 assert(OpSU && "Node has no SUnit!"); 460 if (OpSU == SU) continue; // In the same group. 461 462 EVT OpVT = N->getOperand(i).getValueType(); 463 assert(OpVT != MVT::Glue && "Glued nodes should be in same sunit!"); 464 bool isChain = OpVT == MVT::Other; 465 466 unsigned PhysReg = 0; 467 int Cost = 1; 468 // Determine if this is a physical register dependency. 469 CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost); 470 assert((PhysReg == 0 || !isChain) && 471 "Chain dependence via physreg data?"); 472 // FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler 473 // emits a copy from the physical register to a virtual register unless 474 // it requires a cross class copy (cost < 0). That means we are only 475 // treating "expensive to copy" register dependency as physical register 476 // dependency. This may change in the future though. 477 if (Cost >= 0 && !StressSched) 478 PhysReg = 0; 479 480 // If this is a ctrl dep, latency is 1. 481 unsigned OpLatency = isChain ? 1 : OpSU->Latency; 482 // Special-case TokenFactor chains as zero-latency. 483 if(isChain && OpN->getOpcode() == ISD::TokenFactor) 484 OpLatency = 0; 485 486 SDep Dep = isChain ? SDep(OpSU, SDep::Barrier) 487 : SDep(OpSU, SDep::Data, PhysReg); 488 Dep.setLatency(OpLatency); 489 if (!isChain && !UnitLatencies) { 490 computeOperandLatency(OpN, N, i, Dep); 491 ST.adjustSchedDependency(OpSU, SU, Dep); 492 } 493 494 if (!SU->addPred(Dep) && !Dep.isCtrl() && OpSU->NumRegDefsLeft > 1) { 495 // Multiple register uses are combined in the same SUnit. For example, 496 // we could have a set of glued nodes with all their defs consumed by 497 // another set of glued nodes. Register pressure tracking sees this as 498 // a single use, so to keep pressure balanced we reduce the defs. 499 // 500 // We can't tell (without more book-keeping) if this results from 501 // glued nodes or duplicate operands. As long as we don't reduce 502 // NumRegDefsLeft to zero, we handle the common cases well. 503 --OpSU->NumRegDefsLeft; 504 } 505 } 506 } 507 } 508} 509 510/// BuildSchedGraph - Build the SUnit graph from the selection dag that we 511/// are input. This SUnit graph is similar to the SelectionDAG, but 512/// excludes nodes that aren't interesting to scheduling, and represents 513/// glued together nodes with a single SUnit. 514void ScheduleDAGSDNodes::BuildSchedGraph(AliasAnalysis *AA) { 515 // Cluster certain nodes which should be scheduled together. 516 ClusterNodes(); 517 // Populate the SUnits array. 518 BuildSchedUnits(); 519 // Compute all the scheduling dependencies between nodes. 520 AddSchedEdges(); 521} 522 523// Initialize NumNodeDefs for the current Node's opcode. 524void ScheduleDAGSDNodes::RegDefIter::InitNodeNumDefs() { 525 // Check for phys reg copy. 526 if (!Node) 527 return; 528 529 if (!Node->isMachineOpcode()) { 530 if (Node->getOpcode() == ISD::CopyFromReg) 531 NodeNumDefs = 1; 532 else 533 NodeNumDefs = 0; 534 return; 535 } 536 unsigned POpc = Node->getMachineOpcode(); 537 if (POpc == TargetOpcode::IMPLICIT_DEF) { 538 // No register need be allocated for this. 539 NodeNumDefs = 0; 540 return; 541 } 542 if (POpc == TargetOpcode::PATCHPOINT && 543 Node->getValueType(0) == MVT::Other) { 544 // PATCHPOINT is defined to have one result, but it might really have none 545 // if we're not using CallingConv::AnyReg. Don't mistake the chain for a 546 // real definition. 547 NodeNumDefs = 0; 548 return; 549 } 550 unsigned NRegDefs = SchedDAG->TII->get(Node->getMachineOpcode()).getNumDefs(); 551 // Some instructions define regs that are not represented in the selection DAG 552 // (e.g. unused flags). See tMOVi8. Make sure we don't access past NumValues. 553 NodeNumDefs = std::min(Node->getNumValues(), NRegDefs); 554 DefIdx = 0; 555} 556 557// Construct a RegDefIter for this SUnit and find the first valid value. 558ScheduleDAGSDNodes::RegDefIter::RegDefIter(const SUnit *SU, 559 const ScheduleDAGSDNodes *SD) 560 : SchedDAG(SD), Node(SU->getNode()), DefIdx(0), NodeNumDefs(0) { 561 InitNodeNumDefs(); 562 Advance(); 563} 564 565// Advance to the next valid value defined by the SUnit. 566void ScheduleDAGSDNodes::RegDefIter::Advance() { 567 for (;Node;) { // Visit all glued nodes. 568 for (;DefIdx < NodeNumDefs; ++DefIdx) { 569 if (!Node->hasAnyUseOfValue(DefIdx)) 570 continue; 571 ValueType = Node->getSimpleValueType(DefIdx); 572 ++DefIdx; 573 return; // Found a normal regdef. 574 } 575 Node = Node->getGluedNode(); 576 if (!Node) { 577 return; // No values left to visit. 578 } 579 InitNodeNumDefs(); 580 } 581} 582 583void ScheduleDAGSDNodes::InitNumRegDefsLeft(SUnit *SU) { 584 assert(SU->NumRegDefsLeft == 0 && "expect a new node"); 585 for (RegDefIter I(SU, this); I.IsValid(); I.Advance()) { 586 assert(SU->NumRegDefsLeft < USHRT_MAX && "overflow is ok but unexpected"); 587 ++SU->NumRegDefsLeft; 588 } 589} 590 591void ScheduleDAGSDNodes::computeLatency(SUnit *SU) { 592 SDNode *N = SU->getNode(); 593 594 // TokenFactor operands are considered zero latency, and some schedulers 595 // (e.g. Top-Down list) may rely on the fact that operand latency is nonzero 596 // whenever node latency is nonzero. 597 if (N && N->getOpcode() == ISD::TokenFactor) { 598 SU->Latency = 0; 599 return; 600 } 601 602 // Check to see if the scheduler cares about latencies. 603 if (forceUnitLatencies()) { 604 SU->Latency = 1; 605 return; 606 } 607 608 if (!InstrItins || InstrItins->isEmpty()) { 609 if (N && N->isMachineOpcode() && 610 TII->isHighLatencyDef(N->getMachineOpcode())) 611 SU->Latency = HighLatencyCycles; 612 else 613 SU->Latency = 1; 614 return; 615 } 616 617 // Compute the latency for the node. We use the sum of the latencies for 618 // all nodes glued together into this SUnit. 619 SU->Latency = 0; 620 for (SDNode *N = SU->getNode(); N; N = N->getGluedNode()) 621 if (N->isMachineOpcode()) 622 SU->Latency += TII->getInstrLatency(InstrItins, N); 623} 624 625void ScheduleDAGSDNodes::computeOperandLatency(SDNode *Def, SDNode *Use, 626 unsigned OpIdx, SDep& dep) const{ 627 // Check to see if the scheduler cares about latencies. 628 if (forceUnitLatencies()) 629 return; 630 631 if (dep.getKind() != SDep::Data) 632 return; 633 634 unsigned DefIdx = Use->getOperand(OpIdx).getResNo(); 635 if (Use->isMachineOpcode()) 636 // Adjust the use operand index by num of defs. 637 OpIdx += TII->get(Use->getMachineOpcode()).getNumDefs(); 638 int Latency = TII->getOperandLatency(InstrItins, Def, DefIdx, Use, OpIdx); 639 if (Latency > 1 && Use->getOpcode() == ISD::CopyToReg && 640 !BB->succ_empty()) { 641 unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg(); 642 if (TargetRegisterInfo::isVirtualRegister(Reg)) 643 // This copy is a liveout value. It is likely coalesced, so reduce the 644 // latency so not to penalize the def. 645 // FIXME: need target specific adjustment here? 646 Latency = (Latency > 1) ? Latency - 1 : 1; 647 } 648 if (Latency >= 0) 649 dep.setLatency(Latency); 650} 651 652void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const { 653#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 654 if (!SU->getNode()) { 655 dbgs() << "PHYS REG COPY\n"; 656 return; 657 } 658 659 SU->getNode()->dump(DAG); 660 dbgs() << "\n"; 661 SmallVector<SDNode *, 4> GluedNodes; 662 for (SDNode *N = SU->getNode()->getGluedNode(); N; N = N->getGluedNode()) 663 GluedNodes.push_back(N); 664 while (!GluedNodes.empty()) { 665 dbgs() << " "; 666 GluedNodes.back()->dump(DAG); 667 dbgs() << "\n"; 668 GluedNodes.pop_back(); 669 } 670#endif 671} 672 673#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 674void ScheduleDAGSDNodes::dumpSchedule() const { 675 for (unsigned i = 0, e = Sequence.size(); i != e; i++) { 676 if (SUnit *SU = Sequence[i]) 677 SU->dump(this); 678 else 679 dbgs() << "**** NOOP ****\n"; 680 } 681} 682#endif 683 684#ifndef NDEBUG 685/// VerifyScheduledSequence - Verify that all SUnits were scheduled and that 686/// their state is consistent with the nodes listed in Sequence. 687/// 688void ScheduleDAGSDNodes::VerifyScheduledSequence(bool isBottomUp) { 689 unsigned ScheduledNodes = ScheduleDAG::VerifyScheduledDAG(isBottomUp); 690 unsigned Noops = 0; 691 for (unsigned i = 0, e = Sequence.size(); i != e; ++i) 692 if (!Sequence[i]) 693 ++Noops; 694 assert(Sequence.size() - Noops == ScheduledNodes && 695 "The number of nodes scheduled doesn't match the expected number!"); 696} 697#endif // NDEBUG 698 699/// ProcessSDDbgValues - Process SDDbgValues associated with this node. 700static void 701ProcessSDDbgValues(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter, 702 SmallVectorImpl<std::pair<unsigned, MachineInstr*> > &Orders, 703 DenseMap<SDValue, unsigned> &VRBaseMap, unsigned Order) { 704 if (!N->getHasDebugValue()) 705 return; 706 707 // Opportunistically insert immediate dbg_value uses, i.e. those with source 708 // order number right after the N. 709 MachineBasicBlock *BB = Emitter.getBlock(); 710 MachineBasicBlock::iterator InsertPos = Emitter.getInsertPos(); 711 ArrayRef<SDDbgValue*> DVs = DAG->GetDbgValues(N); 712 for (unsigned i = 0, e = DVs.size(); i != e; ++i) { 713 if (DVs[i]->isInvalidated()) 714 continue; 715 unsigned DVOrder = DVs[i]->getOrder(); 716 if (!Order || DVOrder == ++Order) { 717 MachineInstr *DbgMI = Emitter.EmitDbgValue(DVs[i], VRBaseMap); 718 if (DbgMI) { 719 Orders.push_back(std::make_pair(DVOrder, DbgMI)); 720 BB->insert(InsertPos, DbgMI); 721 } 722 DVs[i]->setIsInvalidated(); 723 } 724 } 725} 726 727// ProcessSourceNode - Process nodes with source order numbers. These are added 728// to a vector which EmitSchedule uses to determine how to insert dbg_value 729// instructions in the right order. 730static void 731ProcessSourceNode(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter, 732 DenseMap<SDValue, unsigned> &VRBaseMap, 733 SmallVectorImpl<std::pair<unsigned, MachineInstr*> > &Orders, 734 SmallSet<unsigned, 8> &Seen) { 735 unsigned Order = N->getIROrder(); 736 if (!Order || !Seen.insert(Order).second) { 737 // Process any valid SDDbgValues even if node does not have any order 738 // assigned. 739 ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, 0); 740 return; 741 } 742 743 MachineBasicBlock *BB = Emitter.getBlock(); 744 if (Emitter.getInsertPos() == BB->begin() || BB->back().isPHI() || 745 // Fast-isel may have inserted some instructions, in which case the 746 // BB->back().isPHI() test will not fire when we want it to. 747 std::prev(Emitter.getInsertPos())->isPHI()) { 748 // Did not insert any instruction. 749 Orders.push_back(std::make_pair(Order, (MachineInstr*)nullptr)); 750 return; 751 } 752 753 Orders.push_back(std::make_pair(Order, std::prev(Emitter.getInsertPos()))); 754 ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, Order); 755} 756 757void ScheduleDAGSDNodes:: 758EmitPhysRegCopy(SUnit *SU, DenseMap<SUnit*, unsigned> &VRBaseMap, 759 MachineBasicBlock::iterator InsertPos) { 760 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end(); 761 I != E; ++I) { 762 if (I->isCtrl()) continue; // ignore chain preds 763 if (I->getSUnit()->CopyDstRC) { 764 // Copy to physical register. 765 DenseMap<SUnit*, unsigned>::iterator VRI = VRBaseMap.find(I->getSUnit()); 766 assert(VRI != VRBaseMap.end() && "Node emitted out of order - late"); 767 // Find the destination physical register. 768 unsigned Reg = 0; 769 for (SUnit::const_succ_iterator II = SU->Succs.begin(), 770 EE = SU->Succs.end(); II != EE; ++II) { 771 if (II->isCtrl()) continue; // ignore chain preds 772 if (II->getReg()) { 773 Reg = II->getReg(); 774 break; 775 } 776 } 777 BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), Reg) 778 .addReg(VRI->second); 779 } else { 780 // Copy from physical register. 781 assert(I->getReg() && "Unknown physical register!"); 782 unsigned VRBase = MRI.createVirtualRegister(SU->CopyDstRC); 783 bool isNew = VRBaseMap.insert(std::make_pair(SU, VRBase)).second; 784 (void)isNew; // Silence compiler warning. 785 assert(isNew && "Node emitted out of order - early"); 786 BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), VRBase) 787 .addReg(I->getReg()); 788 } 789 break; 790 } 791} 792 793/// EmitSchedule - Emit the machine code in scheduled order. Return the new 794/// InsertPos and MachineBasicBlock that contains this insertion 795/// point. ScheduleDAGSDNodes holds a BB pointer for convenience, but this does 796/// not necessarily refer to returned BB. The emitter may split blocks. 797MachineBasicBlock *ScheduleDAGSDNodes:: 798EmitSchedule(MachineBasicBlock::iterator &InsertPos) { 799 InstrEmitter Emitter(BB, InsertPos); 800 DenseMap<SDValue, unsigned> VRBaseMap; 801 DenseMap<SUnit*, unsigned> CopyVRBaseMap; 802 SmallVector<std::pair<unsigned, MachineInstr*>, 32> Orders; 803 SmallSet<unsigned, 8> Seen; 804 bool HasDbg = DAG->hasDebugValues(); 805 806 // If this is the first BB, emit byval parameter dbg_value's. 807 if (HasDbg && BB->getParent()->begin() == MachineFunction::iterator(BB)) { 808 SDDbgInfo::DbgIterator PDI = DAG->ByvalParmDbgBegin(); 809 SDDbgInfo::DbgIterator PDE = DAG->ByvalParmDbgEnd(); 810 for (; PDI != PDE; ++PDI) { 811 MachineInstr *DbgMI= Emitter.EmitDbgValue(*PDI, VRBaseMap); 812 if (DbgMI) 813 BB->insert(InsertPos, DbgMI); 814 } 815 } 816 817 for (unsigned i = 0, e = Sequence.size(); i != e; i++) { 818 SUnit *SU = Sequence[i]; 819 if (!SU) { 820 // Null SUnit* is a noop. 821 TII->insertNoop(*Emitter.getBlock(), InsertPos); 822 continue; 823 } 824 825 // For pre-regalloc scheduling, create instructions corresponding to the 826 // SDNode and any glued SDNodes and append them to the block. 827 if (!SU->getNode()) { 828 // Emit a copy. 829 EmitPhysRegCopy(SU, CopyVRBaseMap, InsertPos); 830 continue; 831 } 832 833 SmallVector<SDNode *, 4> GluedNodes; 834 for (SDNode *N = SU->getNode()->getGluedNode(); N; N = N->getGluedNode()) 835 GluedNodes.push_back(N); 836 while (!GluedNodes.empty()) { 837 SDNode *N = GluedNodes.back(); 838 Emitter.EmitNode(GluedNodes.back(), SU->OrigNode != SU, SU->isCloned, 839 VRBaseMap); 840 // Remember the source order of the inserted instruction. 841 if (HasDbg) 842 ProcessSourceNode(N, DAG, Emitter, VRBaseMap, Orders, Seen); 843 GluedNodes.pop_back(); 844 } 845 Emitter.EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned, 846 VRBaseMap); 847 // Remember the source order of the inserted instruction. 848 if (HasDbg) 849 ProcessSourceNode(SU->getNode(), DAG, Emitter, VRBaseMap, Orders, 850 Seen); 851 } 852 853 // Insert all the dbg_values which have not already been inserted in source 854 // order sequence. 855 if (HasDbg) { 856 MachineBasicBlock::iterator BBBegin = BB->getFirstNonPHI(); 857 858 // Sort the source order instructions and use the order to insert debug 859 // values. 860 std::sort(Orders.begin(), Orders.end(), less_first()); 861 862 SDDbgInfo::DbgIterator DI = DAG->DbgBegin(); 863 SDDbgInfo::DbgIterator DE = DAG->DbgEnd(); 864 // Now emit the rest according to source order. 865 unsigned LastOrder = 0; 866 for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) { 867 unsigned Order = Orders[i].first; 868 MachineInstr *MI = Orders[i].second; 869 // Insert all SDDbgValue's whose order(s) are before "Order". 870 if (!MI) 871 continue; 872 for (; DI != DE && 873 (*DI)->getOrder() >= LastOrder && (*DI)->getOrder() < Order; ++DI) { 874 if ((*DI)->isInvalidated()) 875 continue; 876 MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap); 877 if (DbgMI) { 878 if (!LastOrder) 879 // Insert to start of the BB (after PHIs). 880 BB->insert(BBBegin, DbgMI); 881 else { 882 // Insert at the instruction, which may be in a different 883 // block, if the block was split by a custom inserter. 884 MachineBasicBlock::iterator Pos = MI; 885 MI->getParent()->insert(Pos, DbgMI); 886 } 887 } 888 } 889 LastOrder = Order; 890 } 891 // Add trailing DbgValue's before the terminator. FIXME: May want to add 892 // some of them before one or more conditional branches? 893 SmallVector<MachineInstr*, 8> DbgMIs; 894 while (DI != DE) { 895 if (!(*DI)->isInvalidated()) 896 if (MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap)) 897 DbgMIs.push_back(DbgMI); 898 ++DI; 899 } 900 901 MachineBasicBlock *InsertBB = Emitter.getBlock(); 902 MachineBasicBlock::iterator Pos = InsertBB->getFirstTerminator(); 903 InsertBB->insert(Pos, DbgMIs.begin(), DbgMIs.end()); 904 } 905 906 InsertPos = Emitter.getInsertPos(); 907 return Emitter.getBlock(); 908} 909 910/// Return the basic block label. 911std::string ScheduleDAGSDNodes::getDAGName() const { 912 return "sunit-dag." + BB->getFullName(); 913} 914