ScheduleDAGSDNodes.cpp revision f55f61f8a105e225f3589cac07ae98c08c26b1f6 
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#define DEBUG_TYPE "pre-RA-sched"
16#include "ScheduleDAGSDNodes.h"
17#include "llvm/CodeGen/SelectionDAG.h"
18#include "llvm/Target/TargetMachine.h"
19#include "llvm/Target/TargetInstrInfo.h"
20#include "llvm/Target/TargetRegisterInfo.h"
21#include "llvm/Support/Debug.h"
22#include "llvm/Support/raw_ostream.h"
23using namespace llvm;
24
25ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
26  : ScheduleDAG(mf) {
27}
28
29/// Run - perform scheduling.
30///
31void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb,
32                             MachineBasicBlock::iterator insertPos) {
33  DAG = dag;
34  ScheduleDAG::Run(bb, insertPos);
35}
36
37SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) {
38  SUnit *SU = NewSUnit(Old->getNode());
39  SU->OrigNode = Old->OrigNode;
40  SU->Latency = Old->Latency;
41  SU->isTwoAddress = Old->isTwoAddress;
42  SU->isCommutable = Old->isCommutable;
43  SU->hasPhysRegDefs = Old->hasPhysRegDefs;
44  SU->hasPhysRegClobbers = Old->hasPhysRegClobbers;
45  Old->isCloned = true;
46  return SU;
47}
48
49/// CheckForPhysRegDependency - Check if the dependency between def and use of
50/// a specified operand is a physical register dependency. If so, returns the
51/// register and the cost of copying the register.
52static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op,
53                                      const TargetRegisterInfo *TRI,
54                                      const TargetInstrInfo *TII,
55                                      unsigned &PhysReg, int &Cost) {
56  if (Op != 2 || User->getOpcode() != ISD::CopyToReg)
57    return;
58
59  unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
60  if (TargetRegisterInfo::isVirtualRegister(Reg))
61    return;
62
63  unsigned ResNo = User->getOperand(2).getResNo();
64  if (Def->isMachineOpcode()) {
65    const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
66    if (ResNo >= II.getNumDefs() &&
67        II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
68      PhysReg = Reg;
69      const TargetRegisterClass *RC =
70        TRI->getPhysicalRegisterRegClass(Reg, Def->getValueType(ResNo));
71      Cost = RC->getCopyCost();
72    }
73  }
74}
75
76void ScheduleDAGSDNodes::BuildSchedUnits() {
77  // During scheduling, the NodeId field of SDNode is used to map SDNodes
78  // to their associated SUnits by holding SUnits table indices. A value
79  // of -1 means the SDNode does not yet have an associated SUnit.
80  unsigned NumNodes = 0;
81  for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
82       E = DAG->allnodes_end(); NI != E; ++NI) {
83    NI->setNodeId(-1);
84    ++NumNodes;
85  }
86
87  // Reserve entries in the vector for each of the SUnits we are creating.  This
88  // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
89  // invalidated.
90  // FIXME: Multiply by 2 because we may clone nodes during scheduling.
91  // This is a temporary workaround.
92  SUnits.reserve(NumNodes * 2);
93
94  // Check to see if the scheduler cares about latencies.
95  bool UnitLatencies = ForceUnitLatencies();
96
97  for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
98       E = DAG->allnodes_end(); NI != E; ++NI) {
99    if (isPassiveNode(NI))  // Leaf node, e.g. a TargetImmediate.
100      continue;
101
102    // If this node has already been processed, stop now.
103    if (NI->getNodeId() != -1) continue;
104
105    SUnit *NodeSUnit = NewSUnit(NI);
106
107    // See if anything is flagged to this node, if so, add them to flagged
108    // nodes.  Nodes can have at most one flag input and one flag output.  Flags
109    // are required to be the last operand and result of a node.
110
111    // Scan up to find flagged preds.
112    SDNode *N = NI;
113    while (N->getNumOperands() &&
114           N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
115      N = N->getOperand(N->getNumOperands()-1).getNode();
116      assert(N->getNodeId() == -1 && "Node already inserted!");
117      N->setNodeId(NodeSUnit->NodeNum);
118    }
119
120    // Scan down to find any flagged succs.
121    N = NI;
122    while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
123      SDValue FlagVal(N, N->getNumValues()-1);
124
125      // There are either zero or one users of the Flag result.
126      bool HasFlagUse = false;
127      for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
128           UI != E; ++UI)
129        if (FlagVal.isOperandOf(*UI)) {
130          HasFlagUse = true;
131          assert(N->getNodeId() == -1 && "Node already inserted!");
132          N->setNodeId(NodeSUnit->NodeNum);
133          N = *UI;
134          break;
135        }
136      if (!HasFlagUse) break;
137    }
138
139    // If there are flag operands involved, N is now the bottom-most node
140    // of the sequence of nodes that are flagged together.
141    // Update the SUnit.
142    NodeSUnit->setNode(N);
143    assert(N->getNodeId() == -1 && "Node already inserted!");
144    N->setNodeId(NodeSUnit->NodeNum);
145
146    // Assign the Latency field of NodeSUnit using target-provided information.
147    if (UnitLatencies)
148      NodeSUnit->Latency = 1;
149    else
150      ComputeLatency(NodeSUnit);
151  }
152}
153
154void ScheduleDAGSDNodes::AddSchedEdges() {
155  // Pass 2: add the preds, succs, etc.
156  for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
157    SUnit *SU = &SUnits[su];
158    SDNode *MainNode = SU->getNode();
159
160    if (MainNode->isMachineOpcode()) {
161      unsigned Opc = MainNode->getMachineOpcode();
162      const TargetInstrDesc &TID = TII->get(Opc);
163      for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
164        if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
165          SU->isTwoAddress = true;
166          break;
167        }
168      }
169      if (TID.isCommutable())
170        SU->isCommutable = true;
171    }
172
173    // Find all predecessors and successors of the group.
174    for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode()) {
175      if (N->isMachineOpcode() &&
176          TII->get(N->getMachineOpcode()).getImplicitDefs()) {
177        SU->hasPhysRegClobbers = true;
178        unsigned NumUsed = CountResults(N);
179        while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
180          --NumUsed;    // Skip over unused values at the end.
181        if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
182          SU->hasPhysRegDefs = true;
183      }
184
185      for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
186        SDNode *OpN = N->getOperand(i).getNode();
187        if (isPassiveNode(OpN)) continue;   // Not scheduled.
188        SUnit *OpSU = &SUnits[OpN->getNodeId()];
189        assert(OpSU && "Node has no SUnit!");
190        if (OpSU == SU) continue;           // In the same group.
191
192        MVT OpVT = N->getOperand(i).getValueType();
193        assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
194        bool isChain = OpVT == MVT::Other;
195
196        unsigned PhysReg = 0;
197        int Cost = 1;
198        // Determine if this is a physical register dependency.
199        CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
200        assert((PhysReg == 0 || !isChain) &&
201               "Chain dependence via physreg data?");
202        // FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
203        // emits a copy from the physical register to a virtual register unless
204        // it requires a cross class copy (cost < 0). That means we are only
205        // treating "expensive to copy" register dependency as physical register
206        // dependency. This may change in the future though.
207        if (Cost >= 0)
208          PhysReg = 0;
209        SU->addPred(SDep(OpSU, isChain ? SDep::Order : SDep::Data,
210                         OpSU->Latency, PhysReg));
211      }
212    }
213  }
214}
215
216/// BuildSchedGraph - Build the SUnit graph from the selection dag that we
217/// are input.  This SUnit graph is similar to the SelectionDAG, but
218/// excludes nodes that aren't interesting to scheduling, and represents
219/// flagged together nodes with a single SUnit.
220void ScheduleDAGSDNodes::BuildSchedGraph() {
221  // Populate the SUnits array.
222  BuildSchedUnits();
223  // Compute all the scheduling dependencies between nodes.
224  AddSchedEdges();
225}
226
227void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
228  const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
229
230  // Compute the latency for the node.  We use the sum of the latencies for
231  // all nodes flagged together into this SUnit.
232  SU->Latency = 0;
233  bool SawMachineOpcode = false;
234  for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode())
235    if (N->isMachineOpcode()) {
236      SawMachineOpcode = true;
237      SU->Latency +=
238        InstrItins.getLatency(TII->get(N->getMachineOpcode()).getSchedClass());
239    }
240}
241
242/// CountResults - The results of target nodes have register or immediate
243/// operands first, then an optional chain, and optional flag operands (which do
244/// not go into the resulting MachineInstr).
245unsigned ScheduleDAGSDNodes::CountResults(SDNode *Node) {
246  unsigned N = Node->getNumValues();
247  while (N && Node->getValueType(N - 1) == MVT::Flag)
248    --N;
249  if (N && Node->getValueType(N - 1) == MVT::Other)
250    --N;    // Skip over chain result.
251  return N;
252}
253
254/// CountOperands - The inputs to target nodes have any actual inputs first,
255/// followed by special operands that describe memory references, then an
256/// optional chain operand, then an optional flag operand.  Compute the number
257/// of actual operands that will go into the resulting MachineInstr.
258unsigned ScheduleDAGSDNodes::CountOperands(SDNode *Node) {
259  unsigned N = ComputeMemOperandsEnd(Node);
260  while (N && isa<MemOperandSDNode>(Node->getOperand(N - 1).getNode()))
261    --N; // Ignore MEMOPERAND nodes
262  return N;
263}
264
265/// ComputeMemOperandsEnd - Find the index one past the last MemOperandSDNode
266/// operand
267unsigned ScheduleDAGSDNodes::ComputeMemOperandsEnd(SDNode *Node) {
268  unsigned N = Node->getNumOperands();
269  while (N && Node->getOperand(N - 1).getValueType() == MVT::Flag)
270    --N;
271  if (N && Node->getOperand(N - 1).getValueType() == MVT::Other)
272    --N; // Ignore chain if it exists.
273  return N;
274}
275
276
277void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const {
278  if (!SU->getNode()) {
279    errs() << "PHYS REG COPY\n";
280    return;
281  }
282
283  SU->getNode()->dump(DAG);
284  errs() << "\n";
285  SmallVector<SDNode *, 4> FlaggedNodes;
286  for (SDNode *N = SU->getNode()->getFlaggedNode(); N; N = N->getFlaggedNode())
287    FlaggedNodes.push_back(N);
288  while (!FlaggedNodes.empty()) {
289    errs() << "    ";
290    FlaggedNodes.back()->dump(DAG);
291    errs() << "\n";
292    FlaggedNodes.pop_back();
293  }
294}
295