-//===-- ScheduleDAG.cpp - Implement a trivial DAG scheduler ---------------===//
+//===---- ScheduleDAG.cpp - Implement the ScheduleDAG class ---------------===//
//
// The LLVM Compiler Infrastructure
//
-// This file was developed by Chris Lattner and is distributed under the
-// University of Illinois Open Source License. See LICENSE.TXT for details.
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "sched"
+#define DEBUG_TYPE "pre-RA-sched"
+#include "llvm/Constants.h"
+#include "llvm/Type.h"
+#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFunction.h"
-#include "llvm/CodeGen/SelectionDAGISel.h"
-#include "llvm/CodeGen/SelectionDAG.h"
-#include "llvm/CodeGen/SSARegMap.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetLowering.h"
-#include "llvm/Support/CommandLine.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/Support/Debug.h"
-#include <iostream>
+#include "llvm/Support/MathExtras.h"
using namespace llvm;
-namespace {
- // Style of scheduling to use.
- enum ScheduleChoices {
- noScheduling,
- simpleScheduling,
- };
-} // namespace
-
-cl::opt<ScheduleChoices> ScheduleStyle("sched",
- cl::desc("Choose scheduling style"),
- cl::init(noScheduling),
- cl::values(
- clEnumValN(noScheduling, "none",
- "Trivial emission with no analysis"),
- clEnumValN(simpleScheduling, "simple",
- "Minimize critical path and maximize processor utilization"),
- clEnumValEnd));
+STATISTIC(NumCommutes, "Number of instructions commuted");
+ScheduleDAG::ScheduleDAG(SelectionDAG &dag, MachineBasicBlock *bb,
+ const TargetMachine &tm)
+ : DAG(dag), BB(bb), TM(tm), RegInfo(BB->getParent()->getRegInfo()) {
+ TII = TM.getInstrInfo();
+ MF = &DAG.getMachineFunction();
+ TRI = TM.getRegisterInfo();
+ ConstPool = BB->getParent()->getConstantPool();
+}
-#ifndef NDEBUG
-static cl::opt<bool>
-ViewDAGs("view-sched-dags", cl::Hidden,
- cl::desc("Pop up a window to show sched dags as they are processed"));
-#else
-static const bool ViewDAGs = 0;
-#endif
+/// CheckForPhysRegDependency - Check if the dependency between def and use of
+/// a specified operand is a physical register dependency. If so, returns the
+/// register and the cost of copying the register.
+static void CheckForPhysRegDependency(SDNode *Def, SDNode *Use, unsigned Op,
+ const TargetRegisterInfo *TRI,
+ const TargetInstrInfo *TII,
+ unsigned &PhysReg, int &Cost) {
+ if (Op != 2 || Use->getOpcode() != ISD::CopyToReg)
+ return;
-namespace {
-//===----------------------------------------------------------------------===//
-///
-/// BitsIterator - Provides iteration through individual bits in a bit vector.
-///
-template<class T>
-class BitsIterator {
-private:
- T Bits; // Bits left to iterate through
-
-public:
- /// Ctor.
- BitsIterator(T Initial) : Bits(Initial) {}
-
- /// Next - Returns the next bit set or zero if exhausted.
- inline T Next() {
- // Get the rightmost bit set
- T Result = Bits & -Bits;
- // Remove from rest
- Bits &= ~Result;
- // Return single bit or zero
- return Result;
+ unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
+ if (TargetRegisterInfo::isVirtualRegister(Reg))
+ return;
+
+ unsigned ResNo = Use->getOperand(2).ResNo;
+ if (Def->isTargetOpcode()) {
+ const TargetInstrDesc &II = TII->get(Def->getTargetOpcode());
+ if (ResNo >= II.getNumDefs() &&
+ II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
+ PhysReg = Reg;
+ const TargetRegisterClass *RC =
+ TRI->getPhysicalRegisterRegClass(Def->getValueType(ResNo), Reg);
+ Cost = RC->getCopyCost();
+ }
}
-};
-
-//===----------------------------------------------------------------------===//
+}
+SUnit *ScheduleDAG::Clone(SUnit *Old) {
+ SUnit *SU = NewSUnit(Old->Node);
+ for (unsigned i = 0, e = SU->FlaggedNodes.size(); i != e; ++i)
+ SU->FlaggedNodes.push_back(SU->FlaggedNodes[i]);
+ SU->InstanceNo = SUnitMap[Old->Node].size();
+ SU->Latency = Old->Latency;
+ SU->isTwoAddress = Old->isTwoAddress;
+ SU->isCommutable = Old->isCommutable;
+ SU->hasPhysRegDefs = Old->hasPhysRegDefs;
+ SUnitMap[Old->Node].push_back(SU);
+ return SU;
+}
-//===----------------------------------------------------------------------===//
-///
-/// ResourceTally - Manages the use of resources over time intervals. Each
-/// item (slot) in the tally vector represents the resources used at a given
-/// moment. A bit set to 1 indicates that a resource is in use, otherwise
-/// available. An assumption is made that the tally is large enough to schedule
-/// all current instructions (asserts otherwise.)
-///
-template<class T>
-class ResourceTally {
-private:
- std::vector<T> Tally; // Resources used per slot
- typedef typename std::vector<T>::iterator Iter;
- // Tally iterator
-
- /// AllInUse - Test to see if all of the resources in the slot are busy (set.)
- inline bool AllInUse(Iter Cursor, unsigned ResourceSet) {
- return (*Cursor & ResourceSet) == ResourceSet;
- }
- /// Skip - Skip over slots that use all of the specified resource (all are
- /// set.)
- Iter Skip(Iter Cursor, unsigned ResourceSet) {
- assert(ResourceSet && "At least one resource bit needs to bet set");
+/// BuildSchedUnits - Build SUnits from the selection dag that we are input.
+/// This SUnit graph is similar to the SelectionDAG, but represents flagged
+/// together nodes with a single SUnit.
+void ScheduleDAG::BuildSchedUnits() {
+ // Reserve entries in the vector for each of the SUnits we are creating. This
+ // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
+ // invalidated.
+ SUnits.reserve(std::distance(DAG.allnodes_begin(), DAG.allnodes_end()));
+
+ for (SelectionDAG::allnodes_iterator NI = DAG.allnodes_begin(),
+ E = DAG.allnodes_end(); NI != E; ++NI) {
+ if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
+ continue;
+
+ // If this node has already been processed, stop now.
+ if (SUnitMap[NI].size()) continue;
- // Continue to the end
- while (true) {
- // Break out if one of the resource bits is not set
- if (!AllInUse(Cursor, ResourceSet)) return Cursor;
- // Try next slot
- Cursor++;
- assert(Cursor < Tally.end() && "Tally is not large enough for schedule");
+ SUnit *NodeSUnit = NewSUnit(NI);
+
+ // See if anything is flagged to this node, if so, add them to flagged
+ // nodes. Nodes can have at most one flag input and one flag output. Flags
+ // are required the be the last operand and result of a node.
+
+ // Scan up, adding flagged preds to FlaggedNodes.
+ SDNode *N = NI;
+ if (N->getNumOperands() &&
+ N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
+ do {
+ N = N->getOperand(N->getNumOperands()-1).Val;
+ NodeSUnit->FlaggedNodes.push_back(N);
+ SUnitMap[N].push_back(NodeSUnit);
+ } while (N->getNumOperands() &&
+ N->getOperand(N->getNumOperands()-1).getValueType()== MVT::Flag);
+ std::reverse(NodeSUnit->FlaggedNodes.begin(),
+ NodeSUnit->FlaggedNodes.end());
}
- }
-
- /// FindSlots - Starting from Begin, locate N consecutive slots where at least
- /// one of the resource bits is available. Returns the address of first slot.
- Iter FindSlots(Iter Begin, unsigned N, unsigned ResourceSet,
- unsigned &Resource) {
- // Track position
- Iter Cursor = Begin;
- // Try all possible slots forward
- while (true) {
- // Skip full slots
- Cursor = Skip(Cursor, ResourceSet);
- // Determine end of interval
- Iter End = Cursor + N;
- assert(End <= Tally.end() && "Tally is not large enough for schedule");
+ // Scan down, adding this node and any flagged succs to FlaggedNodes if they
+ // have a user of the flag operand.
+ N = NI;
+ while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
+ SDOperand FlagVal(N, N->getNumValues()-1);
- // Iterate thru each resource
- BitsIterator<T> Resources(ResourceSet & ~*Cursor);
- while (unsigned Res = Resources.Next()) {
- // Check if resource is available for next N slots
- // Break out if resource is busy
- Iter Interval = Cursor;
- for (; Interval < End && !(*Interval & Res); Interval++) {}
-
- // If available for interval, return where and which resource
- if (Interval == End) {
- Resource = Res;
- return Cursor;
+ // There are either zero or one users of the Flag result.
+ bool HasFlagUse = false;
+ for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
+ UI != E; ++UI)
+ if (FlagVal.isOperandOf(*UI)) {
+ HasFlagUse = true;
+ NodeSUnit->FlaggedNodes.push_back(N);
+ SUnitMap[N].push_back(NodeSUnit);
+ N = *UI;
+ break;
}
- // Otherwise, check if worth checking other resources
- if (AllInUse(Interval, ResourceSet)) {
- // Start looking beyond interval
- Cursor = Interval;
+ if (!HasFlagUse) break;
+ }
+
+ // Now all flagged nodes are in FlaggedNodes and N is the bottom-most node.
+ // Update the SUnit
+ NodeSUnit->Node = N;
+ SUnitMap[N].push_back(NodeSUnit);
+
+ ComputeLatency(NodeSUnit);
+ }
+
+ // Pass 2: add the preds, succs, etc.
+ for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
+ SUnit *SU = &SUnits[su];
+ SDNode *MainNode = SU->Node;
+
+ if (MainNode->isTargetOpcode()) {
+ unsigned Opc = MainNode->getTargetOpcode();
+ const TargetInstrDesc &TID = TII->get(Opc);
+ for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
+ if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
+ SU->isTwoAddress = true;
break;
}
}
- Cursor++;
+ if (TID.isCommutable())
+ SU->isCommutable = true;
}
+
+ // Find all predecessors and successors of the group.
+ // Temporarily add N to make code simpler.
+ SU->FlaggedNodes.push_back(MainNode);
+
+ for (unsigned n = 0, e = SU->FlaggedNodes.size(); n != e; ++n) {
+ SDNode *N = SU->FlaggedNodes[n];
+ if (N->isTargetOpcode() &&
+ TII->get(N->getTargetOpcode()).getImplicitDefs() &&
+ CountResults(N) > TII->get(N->getTargetOpcode()).getNumDefs())
+ SU->hasPhysRegDefs = true;
+
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
+ SDNode *OpN = N->getOperand(i).Val;
+ if (isPassiveNode(OpN)) continue; // Not scheduled.
+ SUnit *OpSU = SUnitMap[OpN].front();
+ assert(OpSU && "Node has no SUnit!");
+ if (OpSU == SU) continue; // In the same group.
+
+ MVT::ValueType OpVT = N->getOperand(i).getValueType();
+ assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
+ bool isChain = OpVT == MVT::Other;
+
+ unsigned PhysReg = 0;
+ int Cost = 1;
+ // Determine if this is a physical register dependency.
+ CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
+ SU->addPred(OpSU, isChain, false, PhysReg, Cost);
+ }
+ }
+
+ // Remove MainNode from FlaggedNodes again.
+ SU->FlaggedNodes.pop_back();
}
- /// Reserve - Mark busy (set) the specified N slots.
- void Reserve(Iter Begin, unsigned N, unsigned Resource) {
- // Determine end of interval
- Iter End = Begin + N;
- assert(End <= Tally.end() && "Tally is not large enough for schedule");
-
- // Set resource bit in each slot
- for (; Begin < End; Begin++)
- *Begin |= Resource;
- }
+ return;
+}
-public:
- /// Initialize - Resize and zero the tally to the specified number of time
- /// slots.
- inline void Initialize(unsigned N) {
- Tally.assign(N, 0); // Initialize tally to all zeros.
- }
+void ScheduleDAG::ComputeLatency(SUnit *SU) {
+ const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
- // FindAndReserve - Locate and mark busy (set) N bits started at slot I, using
- // ResourceSet for choices.
- unsigned FindAndReserve(unsigned I, unsigned N, unsigned ResourceSet) {
- // Which resource used
- unsigned Resource;
- // Find slots for instruction.
- Iter Where = FindSlots(Tally.begin() + I, N, ResourceSet, Resource);
- // Reserve the slots
- Reserve(Where, N, Resource);
- // Return time slot (index)
- return Where - Tally.begin();
+ // Compute the latency for the node. We use the sum of the latencies for
+ // all nodes flagged together into this SUnit.
+ if (InstrItins.isEmpty()) {
+ // No latency information.
+ SU->Latency = 1;
+ } else {
+ SU->Latency = 0;
+ if (SU->Node->isTargetOpcode()) {
+ unsigned SchedClass =
+ TII->get(SU->Node->getTargetOpcode()).getSchedClass();
+ InstrStage *S = InstrItins.begin(SchedClass);
+ InstrStage *E = InstrItins.end(SchedClass);
+ for (; S != E; ++S)
+ SU->Latency += S->Cycles;
+ }
+ for (unsigned i = 0, e = SU->FlaggedNodes.size(); i != e; ++i) {
+ SDNode *FNode = SU->FlaggedNodes[i];
+ if (FNode->isTargetOpcode()) {
+ unsigned SchedClass =TII->get(FNode->getTargetOpcode()).getSchedClass();
+ InstrStage *S = InstrItins.begin(SchedClass);
+ InstrStage *E = InstrItins.end(SchedClass);
+ for (; S != E; ++S)
+ SU->Latency += S->Cycles;
+ }
+ }
}
+}
-};
-//===----------------------------------------------------------------------===//
+/// CalculateDepths - compute depths using algorithms for the longest
+/// paths in the DAG
+void ScheduleDAG::CalculateDepths() {
+ unsigned DAGSize = SUnits.size();
+ std::vector<unsigned> InDegree(DAGSize);
+ std::vector<SUnit*> WorkList;
+ WorkList.reserve(DAGSize);
+ // Initialize the data structures
+ for (unsigned i = 0, e = DAGSize; i != e; ++i) {
+ SUnit *SU = &SUnits[i];
+ int NodeNum = SU->NodeNum;
+ unsigned Degree = SU->Preds.size();
+ InDegree[NodeNum] = Degree;
+ SU->Depth = 0;
-//===----------------------------------------------------------------------===//
-// This struct tracks information used to schedule the a node.
-struct ScheduleInfo {
- SDOperand Op; // Operand information
- unsigned Latency; // Cycles to complete instruction
- unsigned ResourceSet; // Bit vector of usable resources
- bool IsBoundary; // Do not shift passed this instruction.
- unsigned Slot; // Operand's time slot
-
- // Ctor.
- ScheduleInfo(SDOperand op)
- : Op(op)
- , Latency(0)
- , ResourceSet(0)
- , IsBoundary(false)
- , Slot(0)
- {}
-};
-//===----------------------------------------------------------------------===//
+ // Is it a node without dependencies?
+ if (Degree == 0) {
+ assert(SU->Preds.empty() && "SUnit should have no predecessors");
+ // Collect leaf nodes
+ WorkList.push_back(SU);
+ }
+ }
+ // Process nodes in the topological order
+ while (!WorkList.empty()) {
+ SUnit *SU = WorkList.back();
+ WorkList.pop_back();
+ unsigned &SUDepth = SU->Depth;
-//===----------------------------------------------------------------------===//
-class SimpleSched {
-private:
- // TODO - get ResourceSet from TII
- enum {
- RSInteger = 0x3, // Two integer units
- RSFloat = 0xC, // Two float units
- RSLoadStore = 0x30, // Two load store units
- RSOther = 0 // Processing unit independent
- };
-
- MachineBasicBlock *BB; // Current basic block
- SelectionDAG &DAG; // DAG of the current basic block
- const TargetMachine &TM; // Target processor
- const TargetInstrInfo &TII; // Target instruction information
- const MRegisterInfo &MRI; // Target processor register information
- SSARegMap *RegMap; // Virtual/real register map
- MachineConstantPool *ConstPool; // Target constant pool
- std::vector<ScheduleInfo> Operands; // All operands to be scheduled
- std::vector<ScheduleInfo*> Ordering; // Emit ordering of operands
- std::map<SDNode *, int> Visited; // Operands that have been visited
- ResourceTally<unsigned> Tally; // Resource usage tally
- unsigned NSlots; // Total latency
- std::map<SDNode *, unsigned>VRMap; // Operand to VR map
- static const unsigned NotFound = ~0U; // Search marker
-
-public:
-
- // Ctor.
- SimpleSched(SelectionDAG &D, MachineBasicBlock *bb)
- : BB(bb), DAG(D), TM(D.getTarget()), TII(*TM.getInstrInfo()),
- MRI(*TM.getRegisterInfo()), RegMap(BB->getParent()->getSSARegMap()),
- ConstPool(BB->getParent()->getConstantPool()),
- NSlots(0) {
- assert(&TII && "Target doesn't provide instr info?");
- assert(&MRI && "Target doesn't provide register info?");
- }
-
- // Run - perform scheduling.
- MachineBasicBlock *Run() {
- Schedule();
- return BB;
- }
-
-private:
- static bool isFlagDefiner(SDOperand Op) { return isFlagDefiner(Op.Val); }
- static bool isFlagUser(SDOperand Op) { return isFlagUser(Op.Val); }
- static bool isFlagDefiner(SDNode *A);
- static bool isFlagUser(SDNode *A);
- static bool isDefiner(SDNode *A, SDNode *B);
- static bool isPassiveOperand(SDOperand Op);
- void IncludeOperand(SDOperand Op);
- void VisitAll();
- void Schedule();
- void GatherOperandInfo();
- bool isStrongDependency(SDOperand A, SDOperand B) {
- return isStrongDependency(A.Val, B.Val);
- }
- bool isWeakDependency(SDOperand A, SDOperand B) {
- return isWeakDependency(A.Val, B.Val);
+ // Use dynamic programming:
+ // When current node is being processed, all of its dependencies
+ // are already processed.
+ // So, just iterate over all predecessors and take the longest path
+ for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ unsigned PredDepth = I->Dep->Depth;
+ if (PredDepth+1 > SUDepth) {
+ SUDepth = PredDepth + 1;
+ }
+ }
+
+ // Update InDegrees of all nodes depending on current SUnit
+ for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I) {
+ SUnit *SU = I->Dep;
+ if (!--InDegree[SU->NodeNum])
+ // If all dependencies of the node are processed already,
+ // then the longest path for the node can be computed now
+ WorkList.push_back(SU);
+ }
}
- static bool isStrongDependency(SDNode *A, SDNode *B);
- static bool isWeakDependency(SDNode *A, SDNode *B);
- void ScheduleBackward();
- void ScheduleForward();
- void EmitAll();
- void EmitFlagUsers(SDOperand Op);
- static unsigned CountResults(SDOperand Op);
- static unsigned CountOperands(SDOperand Op);
- unsigned CreateVirtualRegisters(SDOperand Op, MachineInstr *MI,
- unsigned NumResults,
- const TargetInstrDescriptor &II);
- unsigned Emit(SDOperand A);
-
- void printSI(std::ostream &O, ScheduleInfo *SI) const ;
- void print(std::ostream &O) const ;
- inline void dump(const char *tag) const { std::cerr << tag; dump(); }
- void dump() const;
-};
-//===----------------------------------------------------------------------===//
+}
+/// CalculateHeights - compute heights using algorithms for the longest
+/// paths in the DAG
+void ScheduleDAG::CalculateHeights() {
+ unsigned DAGSize = SUnits.size();
+ std::vector<unsigned> InDegree(DAGSize);
+ std::vector<SUnit*> WorkList;
+ WorkList.reserve(DAGSize);
-//===----------------------------------------------------------------------===//
-class FlagUserIterator {
-private:
- SDNode *Definer; // Node defining flag
- SDNode::use_iterator UI; // User node iterator
- SDNode::use_iterator E; // End of user nodes
- unsigned MinRes; // Minimum flag result
-
-public:
- // Ctor.
- FlagUserIterator(SDNode *D)
- : Definer(D)
- , UI(D->use_begin())
- , E(D->use_end())
- , MinRes(D->getNumValues()) {
- // Find minimum flag result.
- while (MinRes && D->getValueType(MinRes - 1) == MVT::Flag) --MinRes;
- }
-
- /// isFlagUser - Return true if node uses definer's flag.
- bool isFlagUser(SDNode *U) {
- // For each operand (in reverse to only look at flags)
- for (unsigned N = U->getNumOperands(); 0 < N--;) {
- // Get operand
- SDOperand Op = U->getOperand(N);
- // Not user if there are no flags
- if (Op.getValueType() != MVT::Flag) return false;
- // Return true if it is one of the flag results
- if (Op.Val == Definer && Op.ResNo >= MinRes) return true;
- }
- // Not a flag user
- return false;
- }
-
- SDNode *next() {
- // Continue to next user
- while (UI != E) {
- // Next user node
- SDNode *User = *UI++;
- // Return true if is a flag user
- if (isFlagUser(User)) return User;
+ // Initialize the data structures
+ for (unsigned i = 0, e = DAGSize; i != e; ++i) {
+ SUnit *SU = &SUnits[i];
+ int NodeNum = SU->NodeNum;
+ unsigned Degree = SU->Succs.size();
+ InDegree[NodeNum] = Degree;
+ SU->Height = 0;
+
+ // Is it a node without dependencies?
+ if (Degree == 0) {
+ assert(SU->Succs.empty() && "Something wrong");
+ assert(WorkList.empty() && "Should be empty");
+ // Collect leaf nodes
+ WorkList.push_back(SU);
}
-
- // No more user nodes
- return NULL;
}
-};
-} // namespace
+ // Process nodes in the topological order
+ while (!WorkList.empty()) {
+ SUnit *SU = WorkList.back();
+ WorkList.pop_back();
+ unsigned &SUHeight = SU->Height;
+ // Use dynamic programming:
+ // When current node is being processed, all of its dependencies
+ // are already processed.
+ // So, just iterate over all successors and take the longest path
+ for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
+ I != E; ++I) {
+ unsigned SuccHeight = I->Dep->Height;
+ if (SuccHeight+1 > SUHeight) {
+ SUHeight = SuccHeight + 1;
+ }
+ }
-//===----------------------------------------------------------------------===//
-/// isFlagDefiner - Returns true if the operand defines a flag result.
-bool SimpleSched::isFlagDefiner(SDNode *A) {
- unsigned N = A->getNumValues();
- return N && A->getValueType(N - 1) == MVT::Flag;
+ // Update InDegrees of all nodes depending on current SUnit
+ for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ SUnit *SU = I->Dep;
+ if (!--InDegree[SU->NodeNum])
+ // If all dependencies of the node are processed already,
+ // then the longest path for the node can be computed now
+ WorkList.push_back(SU);
+ }
+ }
}
-/// isFlagUser - Returns true if the operand uses a flag result.
-///
-bool SimpleSched::isFlagUser(SDNode *A) {
- unsigned N = A->getNumOperands();
- return N && A->getOperand(N - 1).getValueType() == MVT::Flag;
+/// CountResults - The results of target nodes have register or immediate
+/// operands first, then an optional chain, and optional flag operands (which do
+/// not go into the resulting MachineInstr).
+unsigned ScheduleDAG::CountResults(SDNode *Node) {
+ unsigned N = Node->getNumValues();
+ while (N && Node->getValueType(N - 1) == MVT::Flag)
+ --N;
+ if (N && Node->getValueType(N - 1) == MVT::Other)
+ --N; // Skip over chain result.
+ return N;
}
-/// isDefiner - Return true if Node A is a definder for B.
-///
-bool SimpleSched::isDefiner(SDNode *A, SDNode *B) {
- for (unsigned i = 0, N = B->getNumOperands(); i < N; i++) {
- if (B->getOperand(i).Val == A) return true;
- }
- return false;
+/// CountOperands - The inputs to target nodes have any actual inputs first,
+/// followed by special operands that describe memory references, then an
+/// optional chain operand, then flag operands. Compute the number of
+/// actual operands that will go into the resulting MachineInstr.
+unsigned ScheduleDAG::CountOperands(SDNode *Node) {
+ unsigned N = ComputeMemOperandsEnd(Node);
+ while (N && isa<MemOperandSDNode>(Node->getOperand(N - 1).Val))
+ --N; // Ignore MemOperand nodes
+ return N;
}
-/// isPassiveOperand - Return true if the operand is a non-scheduled leaf
-/// operand.
-bool SimpleSched::isPassiveOperand(SDOperand Op) {
- if (isa<ConstantSDNode>(Op)) return true;
- if (isa<RegisterSDNode>(Op)) return true;
- if (isa<GlobalAddressSDNode>(Op)) return true;
- if (isa<BasicBlockSDNode>(Op)) return true;
- if (isa<FrameIndexSDNode>(Op)) return true;
- if (isa<ConstantPoolSDNode>(Op)) return true;
- if (isa<ExternalSymbolSDNode>(Op)) return true;
- return false;
+/// ComputeMemOperandsEnd - Find the index one past the last MemOperandSDNode
+/// operand
+unsigned ScheduleDAG::ComputeMemOperandsEnd(SDNode *Node) {
+ unsigned N = Node->getNumOperands();
+ while (N && Node->getOperand(N - 1).getValueType() == MVT::Flag)
+ --N;
+ if (N && Node->getOperand(N - 1).getValueType() == MVT::Other)
+ --N; // Ignore chain if it exists.
+ return N;
}
-/// IncludeOperand - Add operand to ScheduleInfo vector.
-///
-void SimpleSched::IncludeOperand(SDOperand Op) {
- // Ignore entry node
- if (Op.getOpcode() == ISD::EntryToken) return;
- // Check current count for operand
- int Count = Visited[Op.Val];
- // If the operand is already in list
- if (Count < 0) return;
- // If this the first time then get count
- if (!Count) Count = Op.Val->use_size();
- // Decrement count to indicate a visit
- Count--;
- // If count has gone to zero then add operand to list
- if (!Count) {
- // Add operand
- Operands.push_back(ScheduleInfo(Op));
- // indicate operand has been added
- Count--;
+static const TargetRegisterClass *getInstrOperandRegClass(
+ const TargetRegisterInfo *TRI,
+ const TargetInstrInfo *TII,
+ const TargetInstrDesc &II,
+ unsigned Op) {
+ if (Op >= II.getNumOperands()) {
+ assert(II.isVariadic() && "Invalid operand # of instruction");
+ return NULL;
}
- // Mark as visited with new count
- Visited[Op.Val] = Count;
+ if (II.OpInfo[Op].isLookupPtrRegClass())
+ return TII->getPointerRegClass();
+ return TRI->getRegClass(II.OpInfo[Op].RegClass);
}
-/// VisitAll - Visit each operand breadth-wise to produce an initial ordering.
-/// Note that the ordering in the Operands vector is reversed.
-void SimpleSched::VisitAll() {
- // Add first element to list
- Operands.push_back(DAG.getRoot());
- for (unsigned i = 0; i < Operands.size(); i++) { // note: size() varies
- // Get next operand. Need copy because Operands vector is growing and
- // addresses can be ScheduleInfo changing.
- SDOperand Op = Operands[i].Op;
- // Get the number of real operands
- unsigned NodeOperands = CountOperands(Op);
- // Get the total number of operands
- unsigned NumOperands = Op.getNumOperands();
-
- // Visit all operands skipping the Other operand if present
- for (unsigned i = NumOperands; 0 < i--;) {
- SDOperand OpI = Op.getOperand(i);
- // Ignore passive operands
- if (isPassiveOperand(OpI)) continue;
- // Check out operand
- IncludeOperand(OpI);
- }
+void ScheduleDAG::EmitCopyFromReg(SDNode *Node, unsigned ResNo,
+ unsigned InstanceNo, unsigned SrcReg,
+ DenseMap<SDOperand, unsigned> &VRBaseMap) {
+ unsigned VRBase = 0;
+ if (TargetRegisterInfo::isVirtualRegister(SrcReg)) {
+ // Just use the input register directly!
+ if (InstanceNo > 0)
+ VRBaseMap.erase(SDOperand(Node, ResNo));
+ bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,ResNo),SrcReg));
+ assert(isNew && "Node emitted out of order - early");
+ return;
}
- // Add entry node last (IncludeOperand filters entry nodes)
- if (DAG.getEntryNode().Val != DAG.getRoot().Val)
- Operands.push_back(DAG.getEntryNode());
-}
-
-/// GatherOperandInfo - Get latency and resource information about each operand.
-///
-void SimpleSched::GatherOperandInfo() {
- // Add addresses of operand info to ordering vector
- // Get number of operands
- unsigned N = Operands.size();
- // FIXME: This is an ugly (but temporary!) hack to test the scheduler before
- // we have real target info.
-
- // For each operand being scheduled
- for (unsigned i = 0; i < N; i++) {
- ScheduleInfo* SI = &Operands[N - i - 1];
- SDOperand Op = SI->Op;
- MVT::ValueType VT = Op.Val->getValueType(0);
- if (Op.isTargetOpcode()) {
- MachineOpCode TOpc = Op.getTargetOpcode();
- // FIXME SI->Latency = std::max(1, TII.maxLatency(TOpc));
- // FIXME SI->ResourceSet = TII.resources(TOpc);
- // There is a cost for keeping values across a call.
- SI->IsBoundary = TII.isCall(TOpc);
-
- if (TII.isLoad(TOpc)) {
- SI->ResourceSet = RSLoadStore;
- SI->Latency = 5;
- } else if (TII.isStore(TOpc)) {
- SI->ResourceSet = RSLoadStore;
- SI->Latency = 2;
- } else if (MVT::isInteger(VT)) {
- SI->ResourceSet = RSInteger;
- SI->Latency = 2;
- } else if (MVT::isFloatingPoint(VT)) {
- SI->ResourceSet = RSFloat;
- SI->Latency = 3;
- } else {
- SI->ResourceSet = RSOther;
- SI->Latency = 0;
- }
+ // If the node is only used by a CopyToReg and the dest reg is a vreg, use
+ // the CopyToReg'd destination register instead of creating a new vreg.
+ bool MatchReg = true;
+ for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
+ UI != E; ++UI) {
+ SDNode *Use = *UI;
+ bool Match = true;
+ if (Use->getOpcode() == ISD::CopyToReg &&
+ Use->getOperand(2).Val == Node &&
+ Use->getOperand(2).ResNo == ResNo) {
+ unsigned DestReg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
+ if (TargetRegisterInfo::isVirtualRegister(DestReg)) {
+ VRBase = DestReg;
+ Match = false;
+ } else if (DestReg != SrcReg)
+ Match = false;
} else {
- if (MVT::isInteger(VT)) {
- SI->ResourceSet = RSInteger;
- SI->Latency = 2;
- } else if (MVT::isFloatingPoint(VT)) {
- SI->ResourceSet = RSFloat;
- SI->Latency = 3;
- } else {
- SI->ResourceSet = RSOther;
- SI->Latency = 0;
+ for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) {
+ SDOperand Op = Use->getOperand(i);
+ if (Op.Val != Node || Op.ResNo != ResNo)
+ continue;
+ MVT::ValueType VT = Node->getValueType(Op.ResNo);
+ if (VT != MVT::Other && VT != MVT::Flag)
+ Match = false;
}
}
+ MatchReg &= Match;
+ if (VRBase)
+ break;
+ }
+
+ const TargetRegisterClass *SrcRC = 0, *DstRC = 0;
+ SrcRC = TRI->getPhysicalRegisterRegClass(Node->getValueType(ResNo), SrcReg);
+
+ // Figure out the register class to create for the destreg.
+ if (VRBase) {
+ DstRC = RegInfo.getRegClass(VRBase);
+ } else {
+ DstRC = DAG.getTargetLoweringInfo()
+ .getRegClassFor(Node->getValueType(ResNo));
+ }
- // Add one slot for the instruction itself
- SI->Latency++;
-
- // Sum up all the latencies for max tally size
- NSlots += SI->Latency;
-
- // Place in initial sorted order
- // FIXME - PUNT - ignore flag users
- if (!isFlagUser(Op)) Ordering.push_back(SI);
+ // If all uses are reading from the src physical register and copying the
+ // register is either impossible or very expensive, then don't create a copy.
+ if (MatchReg && SrcRC->getCopyCost() < 0) {
+ VRBase = SrcReg;
+ } else {
+ // Create the reg, emit the copy.
+ VRBase = RegInfo.createVirtualRegister(DstRC);
+ TII->copyRegToReg(*BB, BB->end(), VRBase, SrcReg, DstRC, SrcRC);
}
+
+ if (InstanceNo > 0)
+ VRBaseMap.erase(SDOperand(Node, ResNo));
+ bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,ResNo), VRBase));
+ assert(isNew && "Node emitted out of order - early");
}
-/// isStrongDependency - Return true if operand A has results used by operand B.
-/// I.E., B must wait for latency of A.
-bool SimpleSched::isStrongDependency(SDNode *A, SDNode *B) {
- // If A defines for B then it's a strong dependency
- if (isDefiner(A, B)) return true;
- // If A defines a flag then it's users are part of the dependency
- if (isFlagDefiner(A)) {
- // Check each flag user
- FlagUserIterator FI(A);
- while (SDNode *User = FI.next()) {
- // If flag user has strong dependency so does B
- if (isStrongDependency(User, B)) return true;
+void ScheduleDAG::CreateVirtualRegisters(SDNode *Node,
+ MachineInstr *MI,
+ const TargetInstrDesc &II,
+ DenseMap<SDOperand, unsigned> &VRBaseMap) {
+ for (unsigned i = 0; i < II.getNumDefs(); ++i) {
+ // If the specific node value is only used by a CopyToReg and the dest reg
+ // is a vreg, use the CopyToReg'd destination register instead of creating
+ // a new vreg.
+ unsigned VRBase = 0;
+ for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
+ UI != E; ++UI) {
+ SDNode *Use = *UI;
+ if (Use->getOpcode() == ISD::CopyToReg &&
+ Use->getOperand(2).Val == Node &&
+ Use->getOperand(2).ResNo == i) {
+ unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
+ if (TargetRegisterInfo::isVirtualRegister(Reg)) {
+ VRBase = Reg;
+ MI->addOperand(MachineOperand::CreateReg(Reg, true));
+ break;
+ }
+ }
}
- }
- // If B defines a flag then it's users are part of the dependency
- if (isFlagDefiner(B)) {
- // Check each flag user
- FlagUserIterator FI(B);
- while (SDNode *User = FI.next()) {
- // If flag user has strong dependency so does B
- if (isStrongDependency(A, User)) return true;
+
+ // Create the result registers for this node and add the result regs to
+ // the machine instruction.
+ if (VRBase == 0) {
+ const TargetRegisterClass *RC = getInstrOperandRegClass(TRI, TII, II, i);
+ assert(RC && "Isn't a register operand!");
+ VRBase = RegInfo.createVirtualRegister(RC);
+ MI->addOperand(MachineOperand::CreateReg(VRBase, true));
}
+
+ bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,i), VRBase));
+ assert(isNew && "Node emitted out of order - early");
}
- return false;
}
-/// isWeakDependency Return true if operand A produces a result that will
-/// conflict with operands of B.
-bool SimpleSched::isWeakDependency(SDNode *A, SDNode *B) {
- // TODO check for conflicting real registers and aliases
- return A->getOpcode() == ISD::EntryToken || isStrongDependency(B, A);
+/// getVR - Return the virtual register corresponding to the specified result
+/// of the specified node.
+static unsigned getVR(SDOperand Op, DenseMap<SDOperand, unsigned> &VRBaseMap) {
+ DenseMap<SDOperand, unsigned>::iterator I = VRBaseMap.find(Op);
+ assert(I != VRBaseMap.end() && "Node emitted out of order - late");
+ return I->second;
}
-/// ScheduleBackward - Schedule instructions so that any long latency
-/// instructions and the critical path get pushed back in time. Time is run in
-/// reverse to allow code reuse of the Tally and eliminate the overhead of
-/// biasing every slot indices against NSlots.
-void SimpleSched::ScheduleBackward() {
- // Size and clear the resource tally
- Tally.Initialize(NSlots);
- // Get number of operands to schedule
- unsigned N = Ordering.size();
-
- // For each operand being scheduled
- for (unsigned i = N; 0 < i--;) {
- ScheduleInfo *SI = Ordering[i];
- // Track insertion
- unsigned Slot = NotFound;
+
+/// AddOperand - Add the specified operand to the specified machine instr. II
+/// specifies the instruction information for the node, and IIOpNum is the
+/// operand number (in the II) that we are adding. IIOpNum and II are used for
+/// assertions only.
+void ScheduleDAG::AddOperand(MachineInstr *MI, SDOperand Op,
+ unsigned IIOpNum,
+ const TargetInstrDesc *II,
+ DenseMap<SDOperand, unsigned> &VRBaseMap) {
+ if (Op.isTargetOpcode()) {
+ // Note that this case is redundant with the final else block, but we
+ // include it because it is the most common and it makes the logic
+ // simpler here.
+ assert(Op.getValueType() != MVT::Other &&
+ Op.getValueType() != MVT::Flag &&
+ "Chain and flag operands should occur at end of operand list!");
- // Compare against those previously scheduled operands
- for (unsigned j = i + 1; j < N; j++) {
- // Get following instruction
- ScheduleInfo *Other = Ordering[j];
-
- // Check dependency against previously inserted operands
- if (isStrongDependency(SI->Op, Other->Op)) {
- Slot = Other->Slot + Other->Latency;
- break;
- } else if (SI->IsBoundary || Other->IsBoundary ||
- isWeakDependency(SI->Op, Other->Op)) {
- Slot = Other->Slot;
- break;
+ // Get/emit the operand.
+ unsigned VReg = getVR(Op, VRBaseMap);
+ const TargetInstrDesc &TID = MI->getDesc();
+ bool isOptDef = (IIOpNum < TID.getNumOperands())
+ ? (TID.OpInfo[IIOpNum].isOptionalDef()) : false;
+ MI->addOperand(MachineOperand::CreateReg(VReg, isOptDef));
+
+ // Verify that it is right.
+ assert(TargetRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?");
+ if (II) {
+ const TargetRegisterClass *RC =
+ getInstrOperandRegClass(TRI, TII, *II, IIOpNum);
+ assert(RC && "Don't have operand info for this instruction!");
+ const TargetRegisterClass *VRC = RegInfo.getRegClass(VReg);
+ if (VRC != RC) {
+ cerr << "Register class of operand and regclass of use don't agree!\n";
+#ifndef NDEBUG
+ cerr << "Operand = " << IIOpNum << "\n";
+ cerr << "Op->Val = "; Op.Val->dump(&DAG); cerr << "\n";
+ cerr << "MI = "; MI->print(cerr);
+ cerr << "VReg = " << VReg << "\n";
+ cerr << "VReg RegClass size = " << VRC->getSize()
+ << ", align = " << VRC->getAlignment() << "\n";
+ cerr << "Expected RegClass size = " << RC->getSize()
+ << ", align = " << RC->getAlignment() << "\n";
+#endif
+ cerr << "Fatal error, aborting.\n";
+ abort();
}
}
-
- // If independent of others (or first entry)
- if (Slot == NotFound) Slot = 0;
-
- // Find a slot where the needed resources are available
- if (SI->ResourceSet)
- Slot = Tally.FindAndReserve(Slot, SI->Latency, SI->ResourceSet);
-
- // Set operand slot
- SI->Slot = Slot;
-
- // Insert sort based on slot
- unsigned j = i + 1;
- for (; j < N; j++) {
- // Get following instruction
- ScheduleInfo *Other = Ordering[j];
- // Should we look further
- if (Slot >= Other->Slot) break;
- // Shuffle other into ordering
- Ordering[j - 1] = Other;
- }
- // Insert operand in proper slot
- if (j != i + 1) Ordering[j - 1] = SI;
- }
-}
-
-/// ScheduleForward - Schedule instructions to maximize packing.
-///
-void SimpleSched::ScheduleForward() {
- // Size and clear the resource tally
- Tally.Initialize(NSlots);
- // Get number of operands to schedule
- unsigned N = Ordering.size();
-
- // For each operand being scheduled
- for (unsigned i = 0; i < N; i++) {
- ScheduleInfo *SI = Ordering[i];
- // Track insertion
- unsigned Slot = NotFound;
-
- // Compare against those previously scheduled operands
- for (unsigned j = i; 0 < j--;) {
- // Get following instruction
- ScheduleInfo *Other = Ordering[j];
-
- // Check dependency against previously inserted operands
- if (isStrongDependency(Other->Op, SI->Op)) {
- Slot = Other->Slot + Other->Latency;
- break;
- } else if (SI->IsBoundary || Other->IsBoundary ||
- isWeakDependency(Other->Op, SI->Op)) {
- Slot = Other->Slot;
- break;
+ } else if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
+ MI->addOperand(MachineOperand::CreateImm(C->getValue()));
+ } else if (ConstantFPSDNode *F = dyn_cast<ConstantFPSDNode>(Op)) {
+ const Type *FType = MVT::getTypeForValueType(Op.getValueType());
+ ConstantFP *CFP = ConstantFP::get(FType, F->getValueAPF());
+ MI->addOperand(MachineOperand::CreateFPImm(CFP));
+ } else if (RegisterSDNode *R = dyn_cast<RegisterSDNode>(Op)) {
+ MI->addOperand(MachineOperand::CreateReg(R->getReg(), false));
+ } else if (GlobalAddressSDNode *TGA = dyn_cast<GlobalAddressSDNode>(Op)) {
+ MI->addOperand(MachineOperand::CreateGA(TGA->getGlobal(),TGA->getOffset()));
+ } else if (BasicBlockSDNode *BB = dyn_cast<BasicBlockSDNode>(Op)) {
+ MI->addOperand(MachineOperand::CreateMBB(BB->getBasicBlock()));
+ } else if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Op)) {
+ MI->addOperand(MachineOperand::CreateFI(FI->getIndex()));
+ } else if (JumpTableSDNode *JT = dyn_cast<JumpTableSDNode>(Op)) {
+ MI->addOperand(MachineOperand::CreateJTI(JT->getIndex()));
+ } else if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op)) {
+ int Offset = CP->getOffset();
+ unsigned Align = CP->getAlignment();
+ const Type *Type = CP->getType();
+ // MachineConstantPool wants an explicit alignment.
+ if (Align == 0) {
+ Align = TM.getTargetData()->getPreferredTypeAlignmentShift(Type);
+ if (Align == 0) {
+ // Alignment of vector types. FIXME!
+ Align = TM.getTargetData()->getABITypeSize(Type);
+ Align = Log2_64(Align);
}
}
- // If independent of others (or first entry)
- if (Slot == NotFound) Slot = 0;
-
- // Find a slot where the needed resources are available
- if (SI->ResourceSet)
- Slot = Tally.FindAndReserve(Slot, SI->Latency, SI->ResourceSet);
-
- // Set operand slot
- SI->Slot = Slot;
+ unsigned Idx;
+ if (CP->isMachineConstantPoolEntry())
+ Idx = ConstPool->getConstantPoolIndex(CP->getMachineCPVal(), Align);
+ else
+ Idx = ConstPool->getConstantPoolIndex(CP->getConstVal(), Align);
+ MI->addOperand(MachineOperand::CreateCPI(Idx, Offset));
+ } else if (ExternalSymbolSDNode *ES = dyn_cast<ExternalSymbolSDNode>(Op)) {
+ MI->addOperand(MachineOperand::CreateES(ES->getSymbol()));
+ } else {
+ assert(Op.getValueType() != MVT::Other &&
+ Op.getValueType() != MVT::Flag &&
+ "Chain and flag operands should occur at end of operand list!");
+ unsigned VReg = getVR(Op, VRBaseMap);
+ MI->addOperand(MachineOperand::CreateReg(VReg, false));
- // Insert sort based on slot
- unsigned j = i;
- for (; 0 < j--;) {
- // Get following instruction
- ScheduleInfo *Other = Ordering[j];
- // Should we look further
- if (Slot >= Other->Slot) break;
- // Shuffle other into ordering
- Ordering[j + 1] = Other;
- }
- // Insert operand in proper slot
- if (j != i) Ordering[j + 1] = SI;
+ // Verify that it is right. Note that the reg class of the physreg and the
+ // vreg don't necessarily need to match, but the target copy insertion has
+ // to be able to handle it. This handles things like copies from ST(0) to
+ // an FP vreg on x86.
+ assert(TargetRegisterInfo::isVirtualRegister(VReg) && "Not a vreg?");
+ if (II) {
+ assert(getInstrOperandRegClass(TRI, TII, *II, IIOpNum) &&
+ "Don't have operand info for this instruction!");
+ }
}
+
}
-/// EmitAll - Emit all operands in schedule sorted order.
-///
-void SimpleSched::EmitAll() {
- // For each operand in the ordering
- for (unsigned i = 0, N = Ordering.size(); i < N; i++) {
- // Get the scheduling info
- ScheduleInfo *SI = Ordering[i];
- // Get the operand
- SDOperand Op = SI->Op;
- // Emit the operand
- Emit(Op);
- // FIXME - PUNT - If Op defines a flag then it's users need to be emitted now
- if (isFlagDefiner(Op)) EmitFlagUsers(Op);
- }
+void ScheduleDAG::AddMemOperand(MachineInstr *MI, const MemOperand &MO) {
+ MI->addMemOperand(MO);
}
-/// EmitFlagUsers - Emit users of operands flag.
-///
-void SimpleSched::EmitFlagUsers(SDOperand Op) {
- // Check each flag user
- FlagUserIterator FI(Op.Val);
- while (SDNode *User = FI.next()) {
- // Construct user node as operand
- SDOperand OpU(User, 0);
- // Emit user node
- Emit(OpU);
- // If user defines a flag then it's users need to be emitted now
- if (isFlagDefiner(User)) EmitFlagUsers(OpU);
- }
+// Returns the Register Class of a subregister
+static const TargetRegisterClass *getSubRegisterRegClass(
+ const TargetRegisterClass *TRC,
+ unsigned SubIdx) {
+ // Pick the register class of the subregister
+ TargetRegisterInfo::regclass_iterator I =
+ TRC->subregclasses_begin() + SubIdx-1;
+ assert(I < TRC->subregclasses_end() &&
+ "Invalid subregister index for register class");
+ return *I;
}
-/// CountResults - The results of target nodes have register or immediate
-/// operands first, then an optional chain, and optional flag operands (which do
-/// not go into the machine instrs.)
-unsigned SimpleSched::CountResults(SDOperand Op) {
- unsigned N = Op.Val->getNumValues();
- while (N && Op.Val->getValueType(N - 1) == MVT::Flag)
- --N;
- if (N && Op.Val->getValueType(N - 1) == MVT::Other)
- --N; // Skip over chain result.
- return N;
+static const TargetRegisterClass *getSuperregRegisterClass(
+ const TargetRegisterClass *TRC,
+ unsigned SubIdx,
+ MVT::ValueType VT) {
+ // Pick the register class of the superegister for this type
+ for (TargetRegisterInfo::regclass_iterator I = TRC->superregclasses_begin(),
+ E = TRC->superregclasses_end(); I != E; ++I)
+ if ((*I)->hasType(VT) && getSubRegisterRegClass(*I, SubIdx) == TRC)
+ return *I;
+ assert(false && "Couldn't find the register class");
+ return 0;
}
-/// CountOperands The inputs to target nodes have any actual inputs first,
-/// followed by an optional chain operand, then flag operands. Compute the
-/// number of actual operands that will go into the machine instr.
-unsigned SimpleSched::CountOperands(SDOperand Op) {
- unsigned N = Op.getNumOperands();
- while (N && Op.getOperand(N - 1).getValueType() == MVT::Flag)
- --N;
- if (N && Op.getOperand(N - 1).getValueType() == MVT::Other)
- --N; // Ignore chain if it exists.
- return N;
-}
+/// EmitSubregNode - Generate machine code for subreg nodes.
+///
+void ScheduleDAG::EmitSubregNode(SDNode *Node,
+ DenseMap<SDOperand, unsigned> &VRBaseMap) {
+ unsigned VRBase = 0;
+ unsigned Opc = Node->getTargetOpcode();
+ if (Opc == TargetInstrInfo::EXTRACT_SUBREG) {
+ // If the node is only used by a CopyToReg and the dest reg is a vreg, use
+ // the CopyToReg'd destination register instead of creating a new vreg.
+ for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
+ UI != E; ++UI) {
+ SDNode *Use = *UI;
+ if (Use->getOpcode() == ISD::CopyToReg &&
+ Use->getOperand(2).Val == Node) {
+ unsigned DestReg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
+ if (TargetRegisterInfo::isVirtualRegister(DestReg)) {
+ VRBase = DestReg;
+ break;
+ }
+ }
+ }
+
+ unsigned SubIdx = cast<ConstantSDNode>(Node->getOperand(1))->getValue();
+
+ // TODO: If the node is a use of a CopyFromReg from a physical register
+ // fold the extract into the copy now
-/// CreateVirtualRegisters - Add result register values for things that are
-/// defined by this instruction.
-unsigned SimpleSched::CreateVirtualRegisters(SDOperand Op, MachineInstr *MI,
- unsigned NumResults,
- const TargetInstrDescriptor &II) {
- // Create the result registers for this node and add the result regs to
- // the machine instruction.
- const TargetOperandInfo *OpInfo = II.OpInfo;
- unsigned ResultReg = RegMap->createVirtualRegister(OpInfo[0].RegClass);
- MI->addRegOperand(ResultReg, MachineOperand::Def);
- for (unsigned i = 1; i != NumResults; ++i) {
- assert(OpInfo[i].RegClass && "Isn't a register operand!");
- MI->addRegOperand(RegMap->createVirtualRegister(OpInfo[0].RegClass),
- MachineOperand::Def);
- }
- return ResultReg;
+ // Create the extract_subreg machine instruction.
+ MachineInstr *MI =
+ new MachineInstr(BB, TII->get(TargetInstrInfo::EXTRACT_SUBREG));
+
+ // Figure out the register class to create for the destreg.
+ unsigned VReg = getVR(Node->getOperand(0), VRBaseMap);
+ const TargetRegisterClass *TRC = RegInfo.getRegClass(VReg);
+ const TargetRegisterClass *SRC = getSubRegisterRegClass(TRC, SubIdx);
+
+ if (VRBase) {
+ // Grab the destination register
+ const TargetRegisterClass *DRC = RegInfo.getRegClass(VRBase);
+ assert(SRC && DRC && SRC == DRC &&
+ "Source subregister and destination must have the same class");
+ } else {
+ // Create the reg
+ assert(SRC && "Couldn't find source register class");
+ VRBase = RegInfo.createVirtualRegister(SRC);
+ }
+
+ // Add def, source, and subreg index
+ MI->addOperand(MachineOperand::CreateReg(VRBase, true));
+ AddOperand(MI, Node->getOperand(0), 0, 0, VRBaseMap);
+ MI->addOperand(MachineOperand::CreateImm(SubIdx));
+
+ } else if (Opc == TargetInstrInfo::INSERT_SUBREG) {
+ assert((Node->getNumOperands() == 2 || Node->getNumOperands() == 3) &&
+ "Malformed insert_subreg node");
+ bool isUndefInput = (Node->getNumOperands() == 2);
+ unsigned SubReg = 0;
+ unsigned SubIdx = 0;
+
+ if (isUndefInput) {
+ SubReg = getVR(Node->getOperand(0), VRBaseMap);
+ SubIdx = cast<ConstantSDNode>(Node->getOperand(1))->getValue();
+ } else {
+ SubReg = getVR(Node->getOperand(1), VRBaseMap);
+ SubIdx = cast<ConstantSDNode>(Node->getOperand(2))->getValue();
+ }
+
+ // TODO: Add tracking info to MachineRegisterInfo of which vregs are subregs
+ // to allow coalescing in the allocator
+
+ // If the node is only used by a CopyToReg and the dest reg is a vreg, use
+ // the CopyToReg'd destination register instead of creating a new vreg.
+ // If the CopyToReg'd destination register is physical, then fold the
+ // insert into the copy
+ for (SDNode::use_iterator UI = Node->use_begin(), E = Node->use_end();
+ UI != E; ++UI) {
+ SDNode *Use = *UI;
+ if (Use->getOpcode() == ISD::CopyToReg &&
+ Use->getOperand(2).Val == Node) {
+ unsigned DestReg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
+ if (TargetRegisterInfo::isVirtualRegister(DestReg)) {
+ VRBase = DestReg;
+ break;
+ }
+ }
+ }
+
+ // Create the insert_subreg machine instruction.
+ MachineInstr *MI =
+ new MachineInstr(BB, TII->get(TargetInstrInfo::INSERT_SUBREG));
+
+ // Figure out the register class to create for the destreg.
+ const TargetRegisterClass *TRC = 0;
+ if (VRBase) {
+ TRC = RegInfo.getRegClass(VRBase);
+ } else {
+ TRC = getSuperregRegisterClass(RegInfo.getRegClass(SubReg), SubIdx,
+ Node->getValueType(0));
+ assert(TRC && "Couldn't determine register class for insert_subreg");
+ VRBase = RegInfo.createVirtualRegister(TRC); // Create the reg
+ }
+
+ MI->addOperand(MachineOperand::CreateReg(VRBase, true));
+ AddOperand(MI, Node->getOperand(0), 0, 0, VRBaseMap);
+ if (!isUndefInput)
+ AddOperand(MI, Node->getOperand(1), 0, 0, VRBaseMap);
+ MI->addOperand(MachineOperand::CreateImm(SubIdx));
+ } else
+ assert(0 && "Node is not a subreg insert or extract");
+
+ bool isNew = VRBaseMap.insert(std::make_pair(SDOperand(Node,0), VRBase));
+ assert(isNew && "Node emitted out of order - early");
}
-/// Emit - Generate machine code for an operand and needed dependencies.
+/// EmitNode - Generate machine code for an node and needed dependencies.
///
-unsigned SimpleSched::Emit(SDOperand Op) {
- std::map<SDNode *, unsigned>::iterator OpI = VRMap.lower_bound(Op.Val);
- if (OpI != VRMap.end() && OpI->first == Op.Val)
- return OpI->second + Op.ResNo;
- unsigned &OpSlot = VRMap.insert(OpI, std::make_pair(Op.Val, 0))->second;
-
- unsigned ResultReg = 0;
- if (Op.isTargetOpcode()) {
- unsigned Opc = Op.getTargetOpcode();
- const TargetInstrDescriptor &II = TII.get(Opc);
+void ScheduleDAG::EmitNode(SDNode *Node, unsigned InstanceNo,
+ DenseMap<SDOperand, unsigned> &VRBaseMap) {
+ // If machine instruction
+ if (Node->isTargetOpcode()) {
+ unsigned Opc = Node->getTargetOpcode();
+
+ // Handle subreg insert/extract specially
+ if (Opc == TargetInstrInfo::EXTRACT_SUBREG ||
+ Opc == TargetInstrInfo::INSERT_SUBREG) {
+ EmitSubregNode(Node, VRBaseMap);
+ return;
+ }
+
+ const TargetInstrDesc &II = TII->get(Opc);
- unsigned NumResults = CountResults(Op);
- unsigned NodeOperands = CountOperands(Op);
+ unsigned NumResults = CountResults(Node);
+ unsigned NodeOperands = CountOperands(Node);
+ unsigned MemOperandsEnd = ComputeMemOperandsEnd(Node);
unsigned NumMIOperands = NodeOperands + NumResults;
+ bool HasPhysRegOuts = (NumResults > II.getNumDefs()) &&
+ II.getImplicitDefs() != 0;
#ifndef NDEBUG
- assert((unsigned(II.numOperands) == NumMIOperands || II.numOperands == -1)&&
+ assert((II.getNumOperands() == NumMIOperands ||
+ HasPhysRegOuts || II.isVariadic()) &&
"#operands for dag node doesn't match .td file!");
#endif
// Create the new machine instruction.
- MachineInstr *MI = new MachineInstr(Opc, NumMIOperands, true, true);
+ MachineInstr *MI = new MachineInstr(II);
// Add result register values for things that are defined by this
// instruction.
- if (NumResults) ResultReg = CreateVirtualRegisters(Op, MI, NumResults, II);
-
- // If there is a token chain operand, emit it first, as a hack to get avoid
- // really bad cases.
- if (Op.getNumOperands() > NodeOperands &&
- Op.getOperand(NodeOperands).getValueType() == MVT::Other) {
- Emit(Op.getOperand(NodeOperands));
- }
+ if (NumResults)
+ CreateVirtualRegisters(Node, MI, II, VRBaseMap);
// Emit all of the actual operands of this instruction, adding them to the
// instruction as appropriate.
- for (unsigned i = 0; i != NodeOperands; ++i) {
- if (Op.getOperand(i).isTargetOpcode()) {
- // Note that this case is redundant with the final else block, but we
- // include it because it is the most common and it makes the logic
- // simpler here.
- assert(Op.getOperand(i).getValueType() != MVT::Other &&
- Op.getOperand(i).getValueType() != MVT::Flag &&
- "Chain and flag operands should occur at end of operand list!");
-
- MI->addRegOperand(Emit(Op.getOperand(i)), MachineOperand::Use);
- } else if (ConstantSDNode *C =
- dyn_cast<ConstantSDNode>(Op.getOperand(i))) {
- MI->addZeroExtImm64Operand(C->getValue());
- } else if (RegisterSDNode*R =dyn_cast<RegisterSDNode>(Op.getOperand(i))) {
- MI->addRegOperand(R->getReg(), MachineOperand::Use);
- } else if (GlobalAddressSDNode *TGA =
- dyn_cast<GlobalAddressSDNode>(Op.getOperand(i))) {
- MI->addGlobalAddressOperand(TGA->getGlobal(), false, 0);
- } else if (BasicBlockSDNode *BB =
- dyn_cast<BasicBlockSDNode>(Op.getOperand(i))) {
- MI->addMachineBasicBlockOperand(BB->getBasicBlock());
- } else if (FrameIndexSDNode *FI =
- dyn_cast<FrameIndexSDNode>(Op.getOperand(i))) {
- MI->addFrameIndexOperand(FI->getIndex());
- } else if (ConstantPoolSDNode *CP =
- dyn_cast<ConstantPoolSDNode>(Op.getOperand(i))) {
- unsigned Idx = ConstPool->getConstantPoolIndex(CP->get());
- MI->addConstantPoolIndexOperand(Idx);
- } else if (ExternalSymbolSDNode *ES =
- dyn_cast<ExternalSymbolSDNode>(Op.getOperand(i))) {
- MI->addExternalSymbolOperand(ES->getSymbol(), false);
- } else {
- assert(Op.getOperand(i).getValueType() != MVT::Other &&
- Op.getOperand(i).getValueType() != MVT::Flag &&
- "Chain and flag operands should occur at end of operand list!");
- MI->addRegOperand(Emit(Op.getOperand(i)), MachineOperand::Use);
+ for (unsigned i = 0; i != NodeOperands; ++i)
+ AddOperand(MI, Node->getOperand(i), i+II.getNumDefs(), &II, VRBaseMap);
+
+ // Emit all of the memory operands of this instruction
+ for (unsigned i = NodeOperands; i != MemOperandsEnd; ++i)
+ AddMemOperand(MI, cast<MemOperandSDNode>(Node->getOperand(i))->MO);
+
+ // Commute node if it has been determined to be profitable.
+ if (CommuteSet.count(Node)) {
+ MachineInstr *NewMI = TII->commuteInstruction(MI);
+ if (NewMI == 0)
+ DOUT << "Sched: COMMUTING FAILED!\n";
+ else {
+ DOUT << "Sched: COMMUTED TO: " << *NewMI;
+ if (MI != NewMI) {
+ delete MI;
+ MI = NewMI;
+ }
+ ++NumCommutes;
}
}
- // Finally, if this node has any flag operands, we *must* emit them last, to
- // avoid emitting operations that might clobber the flags.
- if (Op.getNumOperands() > NodeOperands) {
- unsigned i = NodeOperands;
- if (Op.getOperand(i).getValueType() == MVT::Other)
- ++i; // the chain is already selected.
- for (unsigned N = Op.getNumOperands(); i < N; i++) {
- assert(Op.getOperand(i).getValueType() == MVT::Flag &&
- "Must be flag operands!");
- Emit(Op.getOperand(i));
+ if (II.usesCustomDAGSchedInsertionHook())
+ // Insert this instruction into the basic block using a target
+ // specific inserter which may returns a new basic block.
+ BB = DAG.getTargetLoweringInfo().EmitInstrWithCustomInserter(MI, BB);
+ else
+ BB->push_back(MI);
+
+ // Additional results must be an physical register def.
+ if (HasPhysRegOuts) {
+ for (unsigned i = II.getNumDefs(); i < NumResults; ++i) {
+ unsigned Reg = II.getImplicitDefs()[i - II.getNumDefs()];
+ if (Node->hasAnyUseOfValue(i))
+ EmitCopyFromReg(Node, i, InstanceNo, Reg, VRBaseMap);
}
}
-
- // Now that we have emitted all operands, emit this instruction itself.
- if ((II.Flags & M_USES_CUSTOM_DAG_SCHED_INSERTION) == 0) {
- BB->insert(BB->end(), MI);
- } else {
- // Insert this instruction into the end of the basic block, potentially
- // taking some custom action.
- BB = DAG.getTargetLoweringInfo().InsertAtEndOfBasicBlock(MI, BB);
- }
} else {
- switch (Op.getOpcode()) {
+ switch (Node->getOpcode()) {
default:
- Op.Val->dump();
+#ifndef NDEBUG
+ Node->dump(&DAG);
+#endif
assert(0 && "This target-independent node should have been selected!");
- case ISD::EntryToken: break;
+ case ISD::EntryToken: // fall thru
case ISD::TokenFactor:
- for (unsigned i = 0, N = Op.getNumOperands(); i < N; i++) {
- Emit(Op.getOperand(i));
- }
+ case ISD::LABEL:
+ case ISD::DECLARE:
+ case ISD::SRCVALUE:
break;
case ISD::CopyToReg: {
- SDOperand FlagOp;
- if (Op.getNumOperands() == 4) {
- FlagOp = Op.getOperand(3);
+ unsigned InReg;
+ if (RegisterSDNode *R = dyn_cast<RegisterSDNode>(Node->getOperand(2)))
+ InReg = R->getReg();
+ else
+ InReg = getVR(Node->getOperand(2), VRBaseMap);
+ unsigned DestReg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
+ if (InReg != DestReg) {// Coalesced away the copy?
+ const TargetRegisterClass *TRC = 0;
+ // Get the target register class
+ if (TargetRegisterInfo::isVirtualRegister(InReg))
+ TRC = RegInfo.getRegClass(InReg);
+ else
+ TRC =
+ TRI->getPhysicalRegisterRegClass(Node->getOperand(2).getValueType(),
+ InReg);
+ TII->copyRegToReg(*BB, BB->end(), DestReg, InReg, TRC, TRC);
}
- if (Op.getOperand(0).Val != FlagOp.Val) {
- Emit(Op.getOperand(0)); // Emit the chain.
- }
- unsigned Val = Emit(Op.getOperand(2));
- if (FlagOp.Val) {
- Emit(FlagOp);
- }
- MRI.copyRegToReg(*BB, BB->end(),
- cast<RegisterSDNode>(Op.getOperand(1))->getReg(), Val,
- RegMap->getRegClass(Val));
break;
}
case ISD::CopyFromReg: {
- Emit(Op.getOperand(0)); // Emit the chain.
- unsigned SrcReg = cast<RegisterSDNode>(Op.getOperand(1))->getReg();
+ unsigned SrcReg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
+ EmitCopyFromReg(Node, 0, InstanceNo, SrcReg, VRBaseMap);
+ break;
+ }
+ case ISD::INLINEASM: {
+ unsigned NumOps = Node->getNumOperands();
+ if (Node->getOperand(NumOps-1).getValueType() == MVT::Flag)
+ --NumOps; // Ignore the flag operand.
- // Figure out the register class to create for the destreg.
- const TargetRegisterClass *TRC = 0;
- if (MRegisterInfo::isVirtualRegister(SrcReg)) {
- TRC = RegMap->getRegClass(SrcReg);
- } else {
- // FIXME: we don't know what register class to generate this for. Do
- // a brute force search and pick the first match. :(
- for (MRegisterInfo::regclass_iterator I = MRI.regclass_begin(),
- E = MRI.regclass_end(); I != E; ++I)
- if ((*I)->contains(SrcReg)) {
- TRC = *I;
- break;
+ // Create the inline asm machine instruction.
+ MachineInstr *MI =
+ new MachineInstr(BB, TII->get(TargetInstrInfo::INLINEASM));
+
+ // Add the asm string as an external symbol operand.
+ const char *AsmStr =
+ cast<ExternalSymbolSDNode>(Node->getOperand(1))->getSymbol();
+ MI->addOperand(MachineOperand::CreateES(AsmStr));
+
+ // Add all of the operand registers to the instruction.
+ for (unsigned i = 2; i != NumOps;) {
+ unsigned Flags = cast<ConstantSDNode>(Node->getOperand(i))->getValue();
+ unsigned NumVals = Flags >> 3;
+
+ MI->addOperand(MachineOperand::CreateImm(Flags));
+ ++i; // Skip the ID value.
+
+ switch (Flags & 7) {
+ default: assert(0 && "Bad flags!");
+ case 1: // Use of register.
+ for (; NumVals; --NumVals, ++i) {
+ unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
+ MI->addOperand(MachineOperand::CreateReg(Reg, false));
+ }
+ break;
+ case 2: // Def of register.
+ for (; NumVals; --NumVals, ++i) {
+ unsigned Reg = cast<RegisterSDNode>(Node->getOperand(i))->getReg();
+ MI->addOperand(MachineOperand::CreateReg(Reg, true));
+ }
+ break;
+ case 3: { // Immediate.
+ for (; NumVals; --NumVals, ++i) {
+ if (ConstantSDNode *CS =
+ dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
+ MI->addOperand(MachineOperand::CreateImm(CS->getValue()));
+ } else if (GlobalAddressSDNode *GA =
+ dyn_cast<GlobalAddressSDNode>(Node->getOperand(i))) {
+ MI->addOperand(MachineOperand::CreateGA(GA->getGlobal(),
+ GA->getOffset()));
+ } else {
+ BasicBlockSDNode *BB =cast<BasicBlockSDNode>(Node->getOperand(i));
+ MI->addOperand(MachineOperand::CreateMBB(BB->getBasicBlock()));
+ }
}
- assert(TRC && "Couldn't find register class for reg copy!");
+ break;
+ }
+ case 4: // Addressing mode.
+ // The addressing mode has been selected, just add all of the
+ // operands to the machine instruction.
+ for (; NumVals; --NumVals, ++i)
+ AddOperand(MI, Node->getOperand(i), 0, 0, VRBaseMap);
+ break;
+ }
}
-
- // Create the reg, emit the copy.
- ResultReg = RegMap->createVirtualRegister(TRC);
- MRI.copyRegToReg(*BB, BB->end(), ResultReg, SrcReg, TRC);
break;
}
}
}
+}
- OpSlot = ResultReg;
- return ResultReg+Op.ResNo;
+void ScheduleDAG::EmitNoop() {
+ TII->insertNoop(*BB, BB->end());
}
-/// Schedule - Order operands according to selected style.
-///
-void SimpleSched::Schedule() {
- switch (ScheduleStyle) {
- case simpleScheduling:
- // Breadth first walk of DAG
- VisitAll();
- // Get latency and resource requirements
- GatherOperandInfo();
- // Don't waste time if is only entry and return
- if (Operands.size() > 2) {
- DEBUG(dump("Pre-"));
- // Push back long instructions and critical path
- ScheduleBackward();
- DEBUG(dump("Mid-"));
- // Pack instructions to maximize resource utilization
- ScheduleForward();
- DEBUG(dump("Post-"));
- // Emit in scheduled order
- EmitAll();
- break;
- } // fall thru
- case noScheduling:
- // Emit instructions in using a DFS from the exit root
- Emit(DAG.getRoot());
+void ScheduleDAG::EmitCrossRCCopy(SUnit *SU,
+ DenseMap<SUnit*, unsigned> &VRBaseMap) {
+ for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl) continue; // ignore chain preds
+ if (!I->Dep->Node) {
+ // Copy to physical register.
+ DenseMap<SUnit*, unsigned>::iterator VRI = VRBaseMap.find(I->Dep);
+ assert(VRI != VRBaseMap.end() && "Node emitted out of order - late");
+ // Find the destination physical register.
+ unsigned Reg = 0;
+ for (SUnit::const_succ_iterator II = SU->Succs.begin(),
+ EE = SU->Succs.end(); II != EE; ++II) {
+ if (I->Reg) {
+ Reg = I->Reg;
+ break;
+ }
+ }
+ assert(I->Reg && "Unknown physical register!");
+ TII->copyRegToReg(*BB, BB->end(), Reg, VRI->second,
+ SU->CopyDstRC, SU->CopySrcRC);
+ } else {
+ // Copy from physical register.
+ assert(I->Reg && "Unknown physical register!");
+ unsigned VRBase = RegInfo.createVirtualRegister(SU->CopyDstRC);
+ bool isNew = VRBaseMap.insert(std::make_pair(SU, VRBase));
+ assert(isNew && "Node emitted out of order - early");
+ TII->copyRegToReg(*BB, BB->end(), VRBase, I->Reg,
+ SU->CopyDstRC, SU->CopySrcRC);
+ }
break;
}
}
-/// printSI - Print schedule info.
-///
-void SimpleSched::printSI(std::ostream &O, ScheduleInfo *SI) const {
-#ifndef NDEBUG
- using namespace std;
- SDOperand Op = SI->Op;
- O << " "
- << hex << Op.Val
- << ", RS=" << SI->ResourceSet
- << ", Lat=" << SI->Latency
- << ", Slot=" << SI->Slot
- << ", ARITY=(" << Op.getNumOperands() << ","
- << Op.Val->getNumValues() << ")"
- << " " << Op.Val->getOperationName(&DAG);
- if (isFlagDefiner(Op)) O << "<#";
- if (isFlagUser(Op)) O << ">#";
-#endif
+/// EmitSchedule - Emit the machine code in scheduled order.
+void ScheduleDAG::EmitSchedule() {
+ // If this is the first basic block in the function, and if it has live ins
+ // that need to be copied into vregs, emit the copies into the top of the
+ // block before emitting the code for the block.
+ if (&MF->front() == BB) {
+ for (MachineRegisterInfo::livein_iterator LI = RegInfo.livein_begin(),
+ E = RegInfo.livein_end(); LI != E; ++LI)
+ if (LI->second) {
+ const TargetRegisterClass *RC = RegInfo.getRegClass(LI->second);
+ TII->copyRegToReg(*MF->begin(), MF->begin()->end(), LI->second,
+ LI->first, RC, RC);
+ }
+ }
+
+
+ // Finally, emit the code for all of the scheduled instructions.
+ DenseMap<SDOperand, unsigned> VRBaseMap;
+ DenseMap<SUnit*, unsigned> CopyVRBaseMap;
+ for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
+ if (SUnit *SU = Sequence[i]) {
+ for (unsigned j = 0, ee = SU->FlaggedNodes.size(); j != ee; ++j)
+ EmitNode(SU->FlaggedNodes[j], SU->InstanceNo, VRBaseMap);
+ if (SU->Node)
+ EmitNode(SU->Node, SU->InstanceNo, VRBaseMap);
+ else
+ EmitCrossRCCopy(SU, CopyVRBaseMap);
+ } else {
+ // Null SUnit* is a noop.
+ EmitNoop();
+ }
+ }
}
-/// print - Print ordering to specified output stream.
-///
-void SimpleSched::print(std::ostream &O) const {
-#ifndef NDEBUG
- using namespace std;
- O << "Ordering\n";
- for (unsigned i = 0, N = Ordering.size(); i < N; i++) {
- printSI(O, Ordering[i]);
- O << "\n";
+/// dump - dump the schedule.
+void ScheduleDAG::dumpSchedule() const {
+ for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
+ if (SUnit *SU = Sequence[i])
+ SU->dump(&DAG);
+ else
+ cerr << "**** NOOP ****\n";
}
-#endif
}
-/// dump - Print ordering to std::cerr.
+
+/// Run - perform scheduling.
///
-void SimpleSched::dump() const {
- print(std::cerr);
+MachineBasicBlock *ScheduleDAG::Run() {
+ Schedule();
+ return BB;
}
-//===----------------------------------------------------------------------===//
+/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
+/// a group of nodes flagged together.
+void SUnit::dump(const SelectionDAG *G) const {
+ cerr << "SU(" << NodeNum << "): ";
+ if (Node)
+ Node->dump(G);
+ else
+ cerr << "CROSS RC COPY ";
+ cerr << "\n";
+ if (FlaggedNodes.size() != 0) {
+ for (unsigned i = 0, e = FlaggedNodes.size(); i != e; i++) {
+ cerr << " ";
+ FlaggedNodes[i]->dump(G);
+ cerr << "\n";
+ }
+ }
+}
-//===----------------------------------------------------------------------===//
-/// ScheduleAndEmitDAG - Pick a safe ordering and emit instructions for each
-/// target node in the graph.
-void SelectionDAGISel::ScheduleAndEmitDAG(SelectionDAG &SD) {
- if (ViewDAGs) SD.viewGraph();
- BB = SimpleSched(SD, BB).Run();
+void SUnit::dumpAll(const SelectionDAG *G) const {
+ dump(G);
+
+ cerr << " # preds left : " << NumPredsLeft << "\n";
+ cerr << " # succs left : " << NumSuccsLeft << "\n";
+ cerr << " Latency : " << Latency << "\n";
+ cerr << " Depth : " << Depth << "\n";
+ cerr << " Height : " << Height << "\n";
+
+ if (Preds.size() != 0) {
+ cerr << " Predecessors:\n";
+ for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
+ I != E; ++I) {
+ if (I->isCtrl)
+ cerr << " ch #";
+ else
+ cerr << " val #";
+ cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";
+ if (I->isSpecial)
+ cerr << " *";
+ cerr << "\n";
+ }
+ }
+ if (Succs.size() != 0) {
+ cerr << " Successors:\n";
+ for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
+ I != E; ++I) {
+ if (I->isCtrl)
+ cerr << " ch #";
+ else
+ cerr << " val #";
+ cerr << I->Dep << " - SU(" << I->Dep->NodeNum << ")";
+ if (I->isSpecial)
+ cerr << " *";
+ cerr << "\n";
+ }
+ }
+ cerr << "\n";
}