1 //===---- ScheduleDAG.cpp - Implement the ScheduleDAG class ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements a simple two pass scheduler. The first pass attempts to push
11 // backward any lengthy instructions and critical paths. The second pass packs
12 // instructions into semi-optimal time slots.
14 //===----------------------------------------------------------------------===//
16 #define DEBUG_TYPE "pre-RA-sched"
17 #include "llvm/Type.h"
18 #include "llvm/CodeGen/ScheduleDAG.h"
19 #include "llvm/CodeGen/MachineConstantPool.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/Target/TargetData.h"
24 #include "llvm/Target/TargetMachine.h"
25 #include "llvm/Target/TargetInstrInfo.h"
26 #include "llvm/Target/TargetLowering.h"
27 #include "llvm/ADT/Statistic.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/MathExtras.h"
33 STATISTIC(NumCommutes, "Number of instructions commuted");
37 SchedLiveInCopies("schedule-livein-copies",
38 cl::desc("Schedule copies of livein registers"),
42 ScheduleDAG::ScheduleDAG(SelectionDAG &dag, MachineBasicBlock *bb,
43 const TargetMachine &tm)
44 : DAG(dag), BB(bb), TM(tm), MRI(BB->getParent()->getRegInfo()) {
45 TII = TM.getInstrInfo();
46 MF = &DAG.getMachineFunction();
47 TRI = TM.getRegisterInfo();
48 TLI = &DAG.getTargetLoweringInfo();
49 ConstPool = BB->getParent()->getConstantPool();
52 /// CheckForPhysRegDependency - Check if the dependency between def and use of
53 /// a specified operand is a physical register dependency. If so, returns the
54 /// register and the cost of copying the register.
55 static void CheckForPhysRegDependency(SDNode *Def, SDNode *Use, unsigned Op,
56 const TargetRegisterInfo *TRI,
57 const TargetInstrInfo *TII,
58 unsigned &PhysReg, int &Cost) {
59 if (Op != 2 || Use->getOpcode() != ISD::CopyToReg)
62 unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
63 if (TargetRegisterInfo::isVirtualRegister(Reg))
66 unsigned ResNo = Use->getOperand(2).ResNo;
67 if (Def->isTargetOpcode()) {
68 const TargetInstrDesc &II = TII->get(Def->getTargetOpcode());
69 if (ResNo >= II.getNumDefs() &&
70 II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
72 const TargetRegisterClass *RC =
73 TRI->getPhysicalRegisterRegClass(Reg, Def->getValueType(ResNo));
74 Cost = RC->getCopyCost();
79 SUnit *ScheduleDAG::Clone(SUnit *Old) {
80 SUnit *SU = NewSUnit(Old->Node);
81 SU->OrigNode = Old->OrigNode;
82 SU->FlaggedNodes = Old->FlaggedNodes;
83 SU->Latency = Old->Latency;
84 SU->isTwoAddress = Old->isTwoAddress;
85 SU->isCommutable = Old->isCommutable;
86 SU->hasPhysRegDefs = Old->hasPhysRegDefs;
91 /// BuildSchedUnits - Build SUnits from the selection dag that we are input.
92 /// This SUnit graph is similar to the SelectionDAG, but represents flagged
93 /// together nodes with a single SUnit.
94 void ScheduleDAG::BuildSchedUnits() {
95 // Reserve entries in the vector for each of the SUnits we are creating. This
96 // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
98 SUnits.reserve(DAG.allnodes_size());
100 // During scheduling, the NodeId field of SDNode is used to map SDNodes
101 // to their associated SUnits by holding SUnits table indices. A value
102 // of -1 means the SDNode does not yet have an associated SUnit.
103 for (SelectionDAG::allnodes_iterator NI = DAG.allnodes_begin(),
104 E = DAG.allnodes_end(); NI != E; ++NI)
107 for (SelectionDAG::allnodes_iterator NI = DAG.allnodes_begin(),
108 E = DAG.allnodes_end(); NI != E; ++NI) {
109 if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
112 // If this node has already been processed, stop now.
113 if (NI->getNodeId() != -1) continue;
115 SUnit *NodeSUnit = NewSUnit(NI);
117 // See if anything is flagged to this node, if so, add them to flagged
118 // nodes. Nodes can have at most one flag input and one flag output. Flags
119 // are required the be the last operand and result of a node.
121 // Scan up, adding flagged preds to FlaggedNodes.
123 if (N->getNumOperands() &&
124 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
126 N = N->getOperand(N->getNumOperands()-1).Val;
127 NodeSUnit->FlaggedNodes.push_back(N);
128 assert(N->getNodeId() == -1 && "Node already inserted!");
129 N->setNodeId(NodeSUnit->NodeNum);
130 } while (N->getNumOperands() &&
131 N->getOperand(N->getNumOperands()-1).getValueType()== MVT::Flag);
132 std::reverse(NodeSUnit->FlaggedNodes.begin(),
133 NodeSUnit->FlaggedNodes.end());
136 // Scan down, adding this node and any flagged succs to FlaggedNodes if they
137 // have a user of the flag operand.
139 while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
140 SDOperand FlagVal(N, N->getNumValues()-1);
142 // There are either zero or one users of the Flag result.
143 bool HasFlagUse = false;
144 for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
146 if (FlagVal.isOperandOf(UI->getUser())) {
148 NodeSUnit->FlaggedNodes.push_back(N);
149 assert(N->getNodeId() == -1 && "Node already inserted!");
150 N->setNodeId(NodeSUnit->NodeNum);
154 if (!HasFlagUse) break;
157 // Now all flagged nodes are in FlaggedNodes and N is the bottom-most node.
160 assert(N->getNodeId() == -1 && "Node already inserted!");
161 N->setNodeId(NodeSUnit->NodeNum);
163 ComputeLatency(NodeSUnit);
166 // Pass 2: add the preds, succs, etc.
167 for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
168 SUnit *SU = &SUnits[su];
169 SDNode *MainNode = SU->Node;
171 if (MainNode->isTargetOpcode()) {
172 unsigned Opc = MainNode->getTargetOpcode();
173 const TargetInstrDesc &TID = TII->get(Opc);
174 for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
175 if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
176 SU->isTwoAddress = true;
180 if (TID.isCommutable())
181 SU->isCommutable = true;
184 // Find all predecessors and successors of the group.
185 // Temporarily add N to make code simpler.
186 SU->FlaggedNodes.push_back(MainNode);
188 for (unsigned n = 0, e = SU->FlaggedNodes.size(); n != e; ++n) {
189 SDNode *N = SU->FlaggedNodes[n];
190 if (N->isTargetOpcode() &&
191 TII->get(N->getTargetOpcode()).getImplicitDefs() &&
192 CountResults(N) > TII->get(N->getTargetOpcode()).getNumDefs())
193 SU->hasPhysRegDefs = true;
195 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
196 SDNode *OpN = N->getOperand(i).Val;
197 if (isPassiveNode(OpN)) continue; // Not scheduled.
198 SUnit *OpSU = &SUnits[OpN->getNodeId()];
199 assert(OpSU && "Node has no SUnit!");
200 if (OpSU == SU) continue; // In the same group.
202 MVT OpVT = N->getOperand(i).getValueType();
203 assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
204 bool isChain = OpVT == MVT::Other;
206 unsigned PhysReg = 0;
208 // Determine if this is a physical register dependency.
209 CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
210 SU->addPred(OpSU, isChain, false, PhysReg, Cost);
214 // Remove MainNode from FlaggedNodes again.
215 SU->FlaggedNodes.pop_back();
221 void ScheduleDAG::ComputeLatency(SUnit *SU) {
222 const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
224 // Compute the latency for the node. We use the sum of the latencies for
225 // all nodes flagged together into this SUnit.
226 if (InstrItins.isEmpty()) {
227 // No latency information.
231 if (SU->Node->isTargetOpcode()) {
232 unsigned SchedClass =
233 TII->get(SU->Node->getTargetOpcode()).getSchedClass();
234 const InstrStage *S = InstrItins.begin(SchedClass);
235 const InstrStage *E = InstrItins.end(SchedClass);
237 SU->Latency += S->Cycles;
239 for (unsigned i = 0, e = SU->FlaggedNodes.size(); i != e; ++i) {
240 SDNode *FNode = SU->FlaggedNodes[i];
241 if (FNode->isTargetOpcode()) {
242 unsigned SchedClass =TII->get(FNode->getTargetOpcode()).getSchedClass();
243 const InstrStage *S = InstrItins.begin(SchedClass);
244 const InstrStage *E = InstrItins.end(SchedClass);
246 SU->Latency += S->Cycles;
252 /// CalculateDepths - compute depths using algorithms for the longest
254 void ScheduleDAG::CalculateDepths() {
255 unsigned DAGSize = SUnits.size();
256 std::vector<unsigned> InDegree(DAGSize);
257 std::vector<SUnit*> WorkList;
258 WorkList.reserve(DAGSize);
260 // Initialize the data structures
261 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
262 SUnit *SU = &SUnits[i];
263 int NodeNum = SU->NodeNum;
264 unsigned Degree = SU->Preds.size();
265 InDegree[NodeNum] = Degree;
268 // Is it a node without dependencies?
270 assert(SU->Preds.empty() && "SUnit should have no predecessors");
271 // Collect leaf nodes
272 WorkList.push_back(SU);
276 // Process nodes in the topological order
277 while (!WorkList.empty()) {
278 SUnit *SU = WorkList.back();
280 unsigned &SUDepth = SU->Depth;
282 // Use dynamic programming:
283 // When current node is being processed, all of its dependencies
284 // are already processed.
285 // So, just iterate over all predecessors and take the longest path
286 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
288 unsigned PredDepth = I->Dep->Depth;
289 if (PredDepth+1 > SUDepth) {
290 SUDepth = PredDepth + 1;
294 // Update InDegrees of all nodes depending on current SUnit
295 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
298 if (!--InDegree[SU->NodeNum])
299 // If all dependencies of the node are processed already,
300 // then the longest path for the node can be computed now
301 WorkList.push_back(SU);
306 /// CalculateHeights - compute heights using algorithms for the longest
308 void ScheduleDAG::CalculateHeights() {
309 unsigned DAGSize = SUnits.size();
310 std::vector<unsigned> InDegree(DAGSize);
311 std::vector<SUnit*> WorkList;
312 WorkList.reserve(DAGSize);
314 // Initialize the data structures
315 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
316 SUnit *SU = &SUnits[i];
317 int NodeNum = SU->NodeNum;
318 unsigned Degree = SU->Succs.size();
319 InDegree[NodeNum] = Degree;
322 // Is it a node without dependencies?
324 assert(SU->Succs.empty() && "Something wrong");
325 assert(WorkList.empty() && "Should be empty");
326 // Collect leaf nodes
327 WorkList.push_back(SU);
331 // Process nodes in the topological order
332 while (!WorkList.empty()) {
333 SUnit *SU = WorkList.back();
335 unsigned &SUHeight = SU->Height;
337 // Use dynamic programming:
338 // When current node is being processed, all of its dependencies
339 // are already processed.
340 // So, just iterate over all successors and take the longest path
341 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
343 unsigned SuccHeight = I->Dep->Height;
344 if (SuccHeight+1 > SUHeight) {
345 SUHeight = SuccHeight + 1;
349 // Update InDegrees of all nodes depending on current SUnit
350 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
353 if (!--InDegree[SU->NodeNum])
354 // If all dependencies of the node are processed already,
355 // then the longest path for the node can be computed now
356 WorkList.push_back(SU);
361 /// CountResults - The results of target nodes have register or immediate
362 /// operands first, then an optional chain, and optional flag operands (which do
363 /// not go into the resulting MachineInstr).
364 unsigned ScheduleDAG::CountResults(SDNode *Node) {
365 unsigned N = Node->getNumValues();
366 while (N && Node->getValueType(N - 1) == MVT::Flag)
368 if (N && Node->getValueType(N - 1) == MVT::Other)
369 --N; // Skip over chain result.
373 /// CountOperands - The inputs to target nodes have any actual inputs first,
374 /// followed by special operands that describe memory references, then an
375 /// optional chain operand, then flag operands. Compute the number of
376 /// actual operands that will go into the resulting MachineInstr.
377 unsigned ScheduleDAG::CountOperands(SDNode *Node) {
378 unsigned N = ComputeMemOperandsEnd(Node);
379 while (N && isa<MemOperandSDNode>(Node->getOperand(N - 1).Val))
380 --N; // Ignore MEMOPERAND nodes
384 /// ComputeMemOperandsEnd - Find the index one past the last MemOperandSDNode
386 unsigned ScheduleDAG::ComputeMemOperandsEnd(SDNode *Node) {
387 unsigned N = Node->getNumOperands();
388 while (N && Node->getOperand(N - 1).getValueType() == MVT::Flag)
390 if (N && Node->getOperand(N - 1).getValueType() == MVT::Other)
391 --N; // Ignore chain if it exists.
395 static const TargetRegisterClass *getInstrOperandRegClass(
396 const TargetRegisterInfo *TRI,
397 const TargetInstrInfo *TII,
398 const TargetInstrDesc &II,
400 if (Op >= II.getNumOperands()) {
401 assert(II.isVariadic() && "Invalid operand # of instruction");
404 if (II.OpInfo[Op].isLookupPtrRegClass())
405 return TII->getPointerRegClass();
406 return TRI->getRegClass(II.OpInfo[Op].RegClass);
409 void ScheduleDAG::EmitCopyFromReg(SDNode *Node, unsigned ResNo,
410 bool IsClone, unsigned SrcReg,
411 DenseMap<SDOperand, unsigned> &VRBaseMap) {
413 if (TargetRegisterInfo::isVirtualRegister(SrcReg)) {
414 // Just use the input register directly!
416 VRBaseMap.erase(SDOperand(Node, ResNo));
417 bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,ResNo),SrcReg));
418 isNew = isNew; // Silence compiler warning.
419 assert(isNew && "Node emitted out of order - early");
423 // If the node is only used by a CopyToReg and the dest reg is a vreg, use
424 // the CopyToReg'd destination register instead of creating a new vreg.
425 bool MatchReg = true;
426 for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
428 SDNode *Use = UI->getUser();
430 if (Use->getOpcode() == ISD::CopyToReg &&
431 Use->getOperand(2).Val == Node &&
432 Use->getOperand(2).ResNo == ResNo) {
433 unsigned DestReg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
434 if (TargetRegisterInfo::isVirtualRegister(DestReg)) {
437 } else if (DestReg != SrcReg)
440 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
441 SDOperand Op = Use->getOperand(i);
442 if (Op.Val != Node || Op.ResNo != ResNo)
444 MVT VT = Node->getValueType(Op.ResNo);
445 if (VT != MVT::Other && VT != MVT::Flag)
454 const TargetRegisterClass *SrcRC = 0, *DstRC = 0;
455 SrcRC = TRI->getPhysicalRegisterRegClass(SrcReg, Node->getValueType(ResNo));
457 // Figure out the register class to create for the destreg.
459 DstRC = MRI.getRegClass(VRBase);
461 DstRC = TLI->getRegClassFor(Node->getValueType(ResNo));
464 // If all uses are reading from the src physical register and copying the
465 // register is either impossible or very expensive, then don't create a copy.
466 if (MatchReg && SrcRC->getCopyCost() < 0) {
469 // Create the reg, emit the copy.
470 VRBase = MRI.createVirtualRegister(DstRC);
471 TII->copyRegToReg(*BB, BB->end(), VRBase, SrcReg, DstRC, SrcRC);
475 VRBaseMap.erase(SDOperand(Node, ResNo));
476 bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,ResNo), VRBase));
477 isNew = isNew; // Silence compiler warning.
478 assert(isNew && "Node emitted out of order - early");
481 /// getDstOfCopyToRegUse - If the only use of the specified result number of
482 /// node is a CopyToReg, return its destination register. Return 0 otherwise.
483 unsigned ScheduleDAG::getDstOfOnlyCopyToRegUse(SDNode *Node,
484 unsigned ResNo) const {
485 if (!Node->hasOneUse())
488 SDNode *Use = Node->use_begin()->getUser();
489 if (Use->getOpcode() == ISD::CopyToReg &&
490 Use->getOperand(2).Val == Node &&
491 Use->getOperand(2).ResNo == ResNo) {
492 unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
493 if (TargetRegisterInfo::isVirtualRegister(Reg))
499 void ScheduleDAG::CreateVirtualRegisters(SDNode *Node, MachineInstr *MI,
500 const TargetInstrDesc &II,
501 DenseMap<SDOperand, unsigned> &VRBaseMap) {
502 assert(Node->getTargetOpcode() != TargetInstrInfo::IMPLICIT_DEF &&
503 "IMPLICIT_DEF should have been handled as a special case elsewhere!");
505 for (unsigned i = 0; i < II.getNumDefs(); ++i) {
506 // If the specific node value is only used by a CopyToReg and the dest reg
507 // is a vreg, use the CopyToReg'd destination register instead of creating
510 for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
512 SDNode *Use = UI->getUser();
513 if (Use->getOpcode() == ISD::CopyToReg &&
514 Use->getOperand(2).Val == Node &&
515 Use->getOperand(2).ResNo == i) {
516 unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
517 if (TargetRegisterInfo::isVirtualRegister(Reg)) {
519 MI->addOperand(MachineOperand::CreateReg(Reg, true));
525 // Create the result registers for this node and add the result regs to
526 // the machine instruction.
528 const TargetRegisterClass *RC = getInstrOperandRegClass(TRI, TII, II, i);
529 assert(RC && "Isn't a register operand!");
530 VRBase = MRI.createVirtualRegister(RC);
531 MI->addOperand(MachineOperand::CreateReg(VRBase, true));
534 bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,i), VRBase));
535 isNew = isNew; // Silence compiler warning.
536 assert(isNew && "Node emitted out of order - early");
540 /// getVR - Return the virtual register corresponding to the specified result
541 /// of the specified node.
542 unsigned ScheduleDAG::getVR(SDOperand Op,
543 DenseMap<SDOperand, unsigned> &VRBaseMap) {
544 if (Op.isTargetOpcode() &&
545 Op.getTargetOpcode() == TargetInstrInfo::IMPLICIT_DEF) {
546 // Add an IMPLICIT_DEF instruction before every use.
547 unsigned VReg = getDstOfOnlyCopyToRegUse(Op.Val, Op.ResNo);
548 // IMPLICIT_DEF can produce any type of result so its TargetInstrDesc
549 // does not include operand register class info.
551 const TargetRegisterClass *RC = TLI->getRegClassFor(Op.getValueType());
552 VReg = MRI.createVirtualRegister(RC);
554 BuildMI(BB, TII->get(TargetInstrInfo::IMPLICIT_DEF), VReg);
558 DenseMap<SDOperand, unsigned>::iterator I = VRBaseMap.find(Op);
559 assert(I != VRBaseMap.end() && "Node emitted out of order - late");
564 /// AddOperand - Add the specified operand to the specified machine instr. II
565 /// specifies the instruction information for the node, and IIOpNum is the
566 /// operand number (in the II) that we are adding. IIOpNum and II are used for
568 void ScheduleDAG::AddOperand(MachineInstr *MI, SDOperand Op,
570 const TargetInstrDesc *II,
571 DenseMap<SDOperand, unsigned> &VRBaseMap) {
572 if (Op.isTargetOpcode()) {
573 // Note that this case is redundant with the final else block, but we
574 // include it because it is the most common and it makes the logic
576 assert(Op.getValueType() != MVT::Other &&
577 Op.getValueType() != MVT::Flag &&
578 "Chain and flag operands should occur at end of operand list!");
579 // Get/emit the operand.
580 unsigned VReg = getVR(Op, VRBaseMap);
581 const TargetInstrDesc &TID = MI->getDesc();
582 bool isOptDef = IIOpNum < TID.getNumOperands() &&
583 TID.OpInfo[IIOpNum].isOptionalDef();
584 MI->addOperand(MachineOperand::CreateReg(VReg, isOptDef));
586 // Verify that it is right.
587 assert(TargetRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?");
590 // There may be no register class for this operand if it is a variadic
591 // argument (RC will be NULL in this case). In this case, we just assume
592 // the regclass is ok.
593 const TargetRegisterClass *RC =
594 getInstrOperandRegClass(TRI, TII, *II, IIOpNum);
595 assert((RC || II->isVariadic()) && "Expected reg class info!");
596 const TargetRegisterClass *VRC = MRI.getRegClass(VReg);
597 if (RC && VRC != RC) {
598 cerr << "Register class of operand and regclass of use don't agree!\n";
599 cerr << "Operand = " << IIOpNum << "\n";
600 cerr << "Op->Val = "; Op.Val->dump(&DAG); cerr << "\n";
601 cerr << "MI = "; MI->print(cerr);
602 cerr << "VReg = " << VReg << "\n";
603 cerr << "VReg RegClass size = " << VRC->getSize()
604 << ", align = " << VRC->getAlignment() << "\n";
605 cerr << "Expected RegClass size = " << RC->getSize()
606 << ", align = " << RC->getAlignment() << "\n";
607 cerr << "Fatal error, aborting.\n";
612 } else if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
613 MI->addOperand(MachineOperand::CreateImm(C->getValue()));
614 } else if (ConstantFPSDNode *F = dyn_cast<ConstantFPSDNode>(Op)) {
615 ConstantFP *CFP = ConstantFP::get(F->getValueAPF());
616 MI->addOperand(MachineOperand::CreateFPImm(CFP));
617 } else if (RegisterSDNode *R = dyn_cast<RegisterSDNode>(Op)) {
618 MI->addOperand(MachineOperand::CreateReg(R->getReg(), false));
619 } else if (GlobalAddressSDNode *TGA = dyn_cast<GlobalAddressSDNode>(Op)) {
620 MI->addOperand(MachineOperand::CreateGA(TGA->getGlobal(),TGA->getOffset()));
621 } else if (BasicBlockSDNode *BB = dyn_cast<BasicBlockSDNode>(Op)) {
622 MI->addOperand(MachineOperand::CreateMBB(BB->getBasicBlock()));
623 } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Op)) {
624 MI->addOperand(MachineOperand::CreateFI(FI->getIndex()));
625 } else if (JumpTableSDNode *JT = dyn_cast<JumpTableSDNode>(Op)) {
626 MI->addOperand(MachineOperand::CreateJTI(JT->getIndex()));
627 } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op)) {
628 int Offset = CP->getOffset();
629 unsigned Align = CP->getAlignment();
630 const Type *Type = CP->getType();
631 // MachineConstantPool wants an explicit alignment.
633 Align = TM.getTargetData()->getPreferredTypeAlignmentShift(Type);
635 // Alignment of vector types. FIXME!
636 Align = TM.getTargetData()->getABITypeSize(Type);
637 Align = Log2_64(Align);
642 if (CP->isMachineConstantPoolEntry())
643 Idx = ConstPool->getConstantPoolIndex(CP->getMachineCPVal(), Align);
645 Idx = ConstPool->getConstantPoolIndex(CP->getConstVal(), Align);
646 MI->addOperand(MachineOperand::CreateCPI(Idx, Offset));
647 } else if (ExternalSymbolSDNode *ES = dyn_cast<ExternalSymbolSDNode>(Op)) {
648 MI->addOperand(MachineOperand::CreateES(ES->getSymbol()));
650 assert(Op.getValueType() != MVT::Other &&
651 Op.getValueType() != MVT::Flag &&
652 "Chain and flag operands should occur at end of operand list!");
653 unsigned VReg = getVR(Op, VRBaseMap);
654 MI->addOperand(MachineOperand::CreateReg(VReg, false));
656 // Verify that it is right. Note that the reg class of the physreg and the
657 // vreg don't necessarily need to match, but the target copy insertion has
658 // to be able to handle it. This handles things like copies from ST(0) to
659 // an FP vreg on x86.
660 assert(TargetRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?");
661 if (II && !II->isVariadic()) {
662 assert(getInstrOperandRegClass(TRI, TII, *II, IIOpNum) &&
663 "Don't have operand info for this instruction!");
669 void ScheduleDAG::AddMemOperand(MachineInstr *MI, const MachineMemOperand &MO) {
670 MI->addMemOperand(MO);
673 // Returns the Register Class of a subregister
674 static const TargetRegisterClass *getSubRegisterRegClass(
675 const TargetRegisterClass *TRC,
677 // Pick the register class of the subregister
678 TargetRegisterInfo::regclass_iterator I =
679 TRC->subregclasses_begin() + SubIdx-1;
680 assert(I < TRC->subregclasses_end() &&
681 "Invalid subregister index for register class");
685 static const TargetRegisterClass *getSuperregRegisterClass(
686 const TargetRegisterClass *TRC,
689 // Pick the register class of the superegister for this type
690 for (TargetRegisterInfo::regclass_iterator I = TRC->superregclasses_begin(),
691 E = TRC->superregclasses_end(); I != E; ++I)
692 if ((*I)->hasType(VT) && getSubRegisterRegClass(*I, SubIdx) == TRC)
694 assert(false && "Couldn't find the register class");
698 /// EmitSubregNode - Generate machine code for subreg nodes.
700 void ScheduleDAG::EmitSubregNode(SDNode *Node,
701 DenseMap<SDOperand, unsigned> &VRBaseMap) {
703 unsigned Opc = Node->getTargetOpcode();
705 // If the node is only used by a CopyToReg and the dest reg is a vreg, use
706 // the CopyToReg'd destination register instead of creating a new vreg.
707 for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
709 SDNode *Use = UI->getUser();
710 if (Use->getOpcode() == ISD::CopyToReg &&
711 Use->getOperand(2).Val == Node) {
712 unsigned DestReg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
713 if (TargetRegisterInfo::isVirtualRegister(DestReg)) {
720 if (Opc == TargetInstrInfo::EXTRACT_SUBREG) {
721 unsigned SubIdx = cast<ConstantSDNode>(Node->getOperand(1))->getValue();
723 // Create the extract_subreg machine instruction.
724 MachineInstr *MI = BuildMI(TII->get(TargetInstrInfo::EXTRACT_SUBREG));
726 // Figure out the register class to create for the destreg.
727 unsigned VReg = getVR(Node->getOperand(0), VRBaseMap);
728 const TargetRegisterClass *TRC = MRI.getRegClass(VReg);
729 const TargetRegisterClass *SRC = getSubRegisterRegClass(TRC, SubIdx);
732 // Grab the destination register
734 const TargetRegisterClass *DRC = MRI.getRegClass(VRBase);
735 assert(SRC && DRC && SRC == DRC &&
736 "Source subregister and destination must have the same class");
740 assert(SRC && "Couldn't find source register class");
741 VRBase = MRI.createVirtualRegister(SRC);
744 // Add def, source, and subreg index
745 MI->addOperand(MachineOperand::CreateReg(VRBase, true));
746 AddOperand(MI, Node->getOperand(0), 0, 0, VRBaseMap);
747 MI->addOperand(MachineOperand::CreateImm(SubIdx));
749 } else if (Opc == TargetInstrInfo::INSERT_SUBREG ||
750 Opc == TargetInstrInfo::SUBREG_TO_REG) {
751 SDOperand N0 = Node->getOperand(0);
752 SDOperand N1 = Node->getOperand(1);
753 SDOperand N2 = Node->getOperand(2);
754 unsigned SubReg = getVR(N1, VRBaseMap);
755 unsigned SubIdx = cast<ConstantSDNode>(N2)->getValue();
758 // Figure out the register class to create for the destreg.
759 const TargetRegisterClass *TRC = 0;
761 TRC = MRI.getRegClass(VRBase);
763 TRC = getSuperregRegisterClass(MRI.getRegClass(SubReg), SubIdx,
764 Node->getValueType(0));
765 assert(TRC && "Couldn't determine register class for insert_subreg");
766 VRBase = MRI.createVirtualRegister(TRC); // Create the reg
769 // Create the insert_subreg or subreg_to_reg machine instruction.
770 MachineInstr *MI = BuildMI(TII->get(Opc));
771 MI->addOperand(MachineOperand::CreateReg(VRBase, true));
773 // If creating a subreg_to_reg, then the first input operand
774 // is an implicit value immediate, otherwise it's a register
775 if (Opc == TargetInstrInfo::SUBREG_TO_REG) {
776 const ConstantSDNode *SD = cast<ConstantSDNode>(N0);
777 MI->addOperand(MachineOperand::CreateImm(SD->getValue()));
779 AddOperand(MI, N0, 0, 0, VRBaseMap);
780 // Add the subregster being inserted
781 AddOperand(MI, N1, 0, 0, VRBaseMap);
782 MI->addOperand(MachineOperand::CreateImm(SubIdx));
785 assert(0 && "Node is not insert_subreg, extract_subreg, or subreg_to_reg");
787 bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,0), VRBase));
788 isNew = isNew; // Silence compiler warning.
789 assert(isNew && "Node emitted out of order - early");
792 /// EmitNode - Generate machine code for an node and needed dependencies.
794 void ScheduleDAG::EmitNode(SDNode *Node, bool IsClone,
795 DenseMap<SDOperand, unsigned> &VRBaseMap) {
796 // If machine instruction
797 if (Node->isTargetOpcode()) {
798 unsigned Opc = Node->getTargetOpcode();
800 // Handle subreg insert/extract specially
801 if (Opc == TargetInstrInfo::EXTRACT_SUBREG ||
802 Opc == TargetInstrInfo::INSERT_SUBREG ||
803 Opc == TargetInstrInfo::SUBREG_TO_REG) {
804 EmitSubregNode(Node, VRBaseMap);
808 if (Opc == TargetInstrInfo::IMPLICIT_DEF)
809 // We want a unique VR for each IMPLICIT_DEF use.
812 const TargetInstrDesc &II = TII->get(Opc);
813 unsigned NumResults = CountResults(Node);
814 unsigned NodeOperands = CountOperands(Node);
815 unsigned MemOperandsEnd = ComputeMemOperandsEnd(Node);
816 bool HasPhysRegOuts = (NumResults > II.getNumDefs()) &&
817 II.getImplicitDefs() != 0;
819 unsigned NumMIOperands = NodeOperands + NumResults;
820 assert((II.getNumOperands() == NumMIOperands ||
821 HasPhysRegOuts || II.isVariadic()) &&
822 "#operands for dag node doesn't match .td file!");
825 // Create the new machine instruction.
826 MachineInstr *MI = BuildMI(II);
828 // Add result register values for things that are defined by this
831 CreateVirtualRegisters(Node, MI, II, VRBaseMap);
833 // Emit all of the actual operands of this instruction, adding them to the
834 // instruction as appropriate.
835 for (unsigned i = 0; i != NodeOperands; ++i)
836 AddOperand(MI, Node->getOperand(i), i+II.getNumDefs(), &II, VRBaseMap);
838 // Emit all of the memory operands of this instruction
839 for (unsigned i = NodeOperands; i != MemOperandsEnd; ++i)
840 AddMemOperand(MI, cast<MemOperandSDNode>(Node->getOperand(i))->MO);
842 // Commute node if it has been determined to be profitable.
843 if (CommuteSet.count(Node)) {
844 MachineInstr *NewMI = TII->commuteInstruction(MI);
846 DOUT << "Sched: COMMUTING FAILED!\n";
848 DOUT << "Sched: COMMUTED TO: " << *NewMI;
857 if (II.usesCustomDAGSchedInsertionHook())
858 // Insert this instruction into the basic block using a target
859 // specific inserter which may returns a new basic block.
860 BB = TLI->EmitInstrWithCustomInserter(MI, BB);
864 // Additional results must be an physical register def.
865 if (HasPhysRegOuts) {
866 for (unsigned i = II.getNumDefs(); i < NumResults; ++i) {
867 unsigned Reg = II.getImplicitDefs()[i - II.getNumDefs()];
868 if (Node->hasAnyUseOfValue(i))
869 EmitCopyFromReg(Node, i, IsClone, Reg, VRBaseMap);
873 switch (Node->getOpcode()) {
878 assert(0 && "This target-independent node should have been selected!");
880 case ISD::EntryToken:
881 assert(0 && "EntryToken should have been excluded from the schedule!");
883 case ISD::TokenFactor: // fall thru
888 case ISD::CopyToReg: {
890 SDOperand SrcVal = Node->getOperand(2);
891 if (RegisterSDNode *R = dyn_cast<RegisterSDNode>(SrcVal))
892 SrcReg = R->getReg();
894 SrcReg = getVR(SrcVal, VRBaseMap);
896 unsigned DestReg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
897 if (SrcReg == DestReg) // Coalesced away the copy? Ignore.
900 const TargetRegisterClass *SrcTRC = 0, *DstTRC = 0;
901 // Get the register classes of the src/dst.
902 if (TargetRegisterInfo::isVirtualRegister(SrcReg))
903 SrcTRC = MRI.getRegClass(SrcReg);
905 SrcTRC = TRI->getPhysicalRegisterRegClass(SrcReg,SrcVal.getValueType());
907 if (TargetRegisterInfo::isVirtualRegister(DestReg))
908 DstTRC = MRI.getRegClass(DestReg);
910 DstTRC = TRI->getPhysicalRegisterRegClass(DestReg,
911 Node->getOperand(1).getValueType());
912 TII->copyRegToReg(*BB, BB->end(), DestReg, SrcReg, DstTRC, SrcTRC);
915 case ISD::CopyFromReg: {
916 unsigned SrcReg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
917 EmitCopyFromReg(Node, 0, IsClone, SrcReg, VRBaseMap);
920 case ISD::INLINEASM: {
921 unsigned NumOps = Node->getNumOperands();
922 if (Node->getOperand(NumOps-1).getValueType() == MVT::Flag)
923 --NumOps; // Ignore the flag operand.
925 // Create the inline asm machine instruction.
926 MachineInstr *MI = BuildMI(TII->get(TargetInstrInfo::INLINEASM));
928 // Add the asm string as an external symbol operand.
930 cast<ExternalSymbolSDNode>(Node->getOperand(1))->getSymbol();
931 MI->addOperand(MachineOperand::CreateES(AsmStr));
933 // Add all of the operand registers to the instruction.
934 for (unsigned i = 2; i != NumOps;) {
935 unsigned Flags = cast<ConstantSDNode>(Node->getOperand(i))->getValue();
936 unsigned NumVals = Flags >> 3;
938 MI->addOperand(MachineOperand::CreateImm(Flags));
939 ++i; // Skip the ID value.
942 default: assert(0 && "Bad flags!");
943 case 1: // Use of register.
944 for (; NumVals; --NumVals, ++i) {
945 unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
946 MI->addOperand(MachineOperand::CreateReg(Reg, false));
949 case 2: // Def of register.
950 for (; NumVals; --NumVals, ++i) {
951 unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
952 MI->addOperand(MachineOperand::CreateReg(Reg, true));
955 case 3: { // Immediate.
956 for (; NumVals; --NumVals, ++i) {
957 if (ConstantSDNode *CS =
958 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
959 MI->addOperand(MachineOperand::CreateImm(CS->getValue()));
960 } else if (GlobalAddressSDNode *GA =
961 dyn_cast<GlobalAddressSDNode>(Node->getOperand(i))) {
962 MI->addOperand(MachineOperand::CreateGA(GA->getGlobal(),
965 BasicBlockSDNode *BB =cast<BasicBlockSDNode>(Node->getOperand(i));
966 MI->addOperand(MachineOperand::CreateMBB(BB->getBasicBlock()));
971 case 4: // Addressing mode.
972 // The addressing mode has been selected, just add all of the
973 // operands to the machine instruction.
974 for (; NumVals; --NumVals, ++i)
975 AddOperand(MI, Node->getOperand(i), 0, 0, VRBaseMap);
986 void ScheduleDAG::EmitNoop() {
987 TII->insertNoop(*BB, BB->end());
990 void ScheduleDAG::EmitCrossRCCopy(SUnit *SU,
991 DenseMap<SUnit*, unsigned> &VRBaseMap) {
992 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
994 if (I->isCtrl) continue; // ignore chain preds
996 // Copy to physical register.
997 DenseMap<SUnit*, unsigned>::iterator VRI = VRBaseMap.find(I->Dep);
998 assert(VRI != VRBaseMap.end() && "Node emitted out of order - late");
999 // Find the destination physical register.
1001 for (SUnit::const_succ_iterator II = SU->Succs.begin(),
1002 EE = SU->Succs.end(); II != EE; ++II) {
1008 assert(I->Reg && "Unknown physical register!");
1009 TII->copyRegToReg(*BB, BB->end(), Reg, VRI->second,
1010 SU->CopyDstRC, SU->CopySrcRC);
1012 // Copy from physical register.
1013 assert(I->Reg && "Unknown physical register!");
1014 unsigned VRBase = MRI.createVirtualRegister(SU->CopyDstRC);
1015 bool isNew = VRBaseMap.insert(std::make_pair(SU, VRBase));
1016 isNew = isNew; // Silence compiler warning.
1017 assert(isNew && "Node emitted out of order - early");
1018 TII->copyRegToReg(*BB, BB->end(), VRBase, I->Reg,
1019 SU->CopyDstRC, SU->CopySrcRC);
1025 /// EmitLiveInCopy - Emit a copy for a live in physical register. If the
1026 /// physical register has only a single copy use, then coalesced the copy
1028 void ScheduleDAG::EmitLiveInCopy(MachineBasicBlock *MBB,
1029 MachineBasicBlock::iterator &InsertPos,
1030 unsigned VirtReg, unsigned PhysReg,
1031 const TargetRegisterClass *RC,
1032 DenseMap<MachineInstr*, unsigned> &CopyRegMap){
1033 unsigned NumUses = 0;
1034 MachineInstr *UseMI = NULL;
1035 for (MachineRegisterInfo::use_iterator UI = MRI.use_begin(VirtReg),
1036 UE = MRI.use_end(); UI != UE; ++UI) {
1042 // If the number of uses is not one, or the use is not a move instruction,
1043 // don't coalesce. Also, only coalesce away a virtual register to virtual
1045 bool Coalesced = false;
1046 unsigned SrcReg, DstReg;
1048 TII->isMoveInstr(*UseMI, SrcReg, DstReg) &&
1049 TargetRegisterInfo::isVirtualRegister(DstReg)) {
1054 // Now find an ideal location to insert the copy.
1055 MachineBasicBlock::iterator Pos = InsertPos;
1056 while (Pos != MBB->begin()) {
1057 MachineInstr *PrevMI = prior(Pos);
1058 DenseMap<MachineInstr*, unsigned>::iterator RI = CopyRegMap.find(PrevMI);
1059 // copyRegToReg might emit multiple instructions to do a copy.
1060 unsigned CopyDstReg = (RI == CopyRegMap.end()) ? 0 : RI->second;
1061 if (CopyDstReg && !TRI->regsOverlap(CopyDstReg, PhysReg))
1062 // This is what the BB looks like right now:
1067 // We want to insert "r1025 = mov r1". Inserting this copy below the
1068 // move to r1024 makes it impossible for that move to be coalesced.
1075 break; // Woot! Found a good location.
1079 TII->copyRegToReg(*MBB, Pos, VirtReg, PhysReg, RC, RC);
1080 CopyRegMap.insert(std::make_pair(prior(Pos), VirtReg));
1082 if (&*InsertPos == UseMI) ++InsertPos;
1087 /// EmitLiveInCopies - If this is the first basic block in the function,
1088 /// and if it has live ins that need to be copied into vregs, emit the
1089 /// copies into the top of the block.
1090 void ScheduleDAG::EmitLiveInCopies(MachineBasicBlock *MBB) {
1091 DenseMap<MachineInstr*, unsigned> CopyRegMap;
1092 MachineBasicBlock::iterator InsertPos = MBB->begin();
1093 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
1094 E = MRI.livein_end(); LI != E; ++LI)
1096 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
1097 EmitLiveInCopy(MBB, InsertPos, LI->second, LI->first, RC, CopyRegMap);
1101 /// EmitSchedule - Emit the machine code in scheduled order.
1102 void ScheduleDAG::EmitSchedule() {
1103 bool isEntryBB = &MF->front() == BB;
1105 if (isEntryBB && !SchedLiveInCopies) {
1106 // If this is the first basic block in the function, and if it has live ins
1107 // that need to be copied into vregs, emit the copies into the top of the
1108 // block before emitting the code for the block.
1109 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
1110 E = MRI.livein_end(); LI != E; ++LI)
1112 const TargetRegisterClass *RC = MRI.getRegClass(LI->second);
1113 TII->copyRegToReg(*MF->begin(), MF->begin()->end(), LI->second,
1118 // Finally, emit the code for all of the scheduled instructions.
1119 DenseMap<SDOperand, unsigned> VRBaseMap;
1120 DenseMap<SUnit*, unsigned> CopyVRBaseMap;
1121 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
1122 SUnit *SU = Sequence[i];
1124 // Null SUnit* is a noop.
1128 for (unsigned j = 0, ee = SU->FlaggedNodes.size(); j != ee; ++j)
1129 EmitNode(SU->FlaggedNodes[j], SU->OrigNode != SU, VRBaseMap);
1131 EmitCrossRCCopy(SU, CopyVRBaseMap);
1133 EmitNode(SU->Node, SU->OrigNode != SU, VRBaseMap);
1136 if (isEntryBB && SchedLiveInCopies)
1137 EmitLiveInCopies(MF->begin());
1140 /// dump - dump the schedule.
1141 void ScheduleDAG::dumpSchedule() const {
1142 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
1143 if (SUnit *SU = Sequence[i])
1146 cerr << "**** NOOP ****\n";
1151 /// Run - perform scheduling.
1153 MachineBasicBlock *ScheduleDAG::Run() {
1158 /// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
1159 /// a group of nodes flagged together.
1160 void SUnit::dump(const SelectionDAG *G) const {
1161 cerr << "SU(" << NodeNum << "): ";
1165 cerr << "CROSS RC COPY ";
1167 if (FlaggedNodes.size() != 0) {
1168 for (unsigned i = 0, e = FlaggedNodes.size(); i != e; i++) {
1170 FlaggedNodes[i]->dump(G);
1176 void SUnit::dumpAll(const SelectionDAG *G) const {
1179 cerr << " # preds left : " << NumPredsLeft << "\n";
1180 cerr << " # succs left : " << NumSuccsLeft << "\n";
1181 cerr << " Latency : " << Latency << "\n";
1182 cerr << " Depth : " << Depth << "\n";
1183 cerr << " Height : " << Height << "\n";
1185 if (Preds.size() != 0) {
1186 cerr << " Predecessors:\n";
1187 for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
1193 cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";
1199 if (Succs.size() != 0) {
1200 cerr << " Successors:\n";
1201 for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
1207 cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";