1 //===------- llvm/CodeGen/ScheduleDAG.h - Common Base Class------*- C++ -*-===//
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 file implements the ScheduleDAG class, which is used as the common
11 // base class for instruction schedulers.
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_CODEGEN_SCHEDULEDAG_H
16 #define LLVM_CODEGEN_SCHEDULEDAG_H
18 #include "llvm/CodeGen/MachineBasicBlock.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/BitVector.h"
21 #include "llvm/ADT/GraphTraits.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/PointerIntPair.h"
28 class MachineConstantPool;
29 class MachineFunction;
30 class MachineRegisterInfo;
32 class TargetRegisterInfo;
35 class TargetInstrInfo;
36 class TargetInstrDesc;
38 class TargetRegisterClass;
39 template<class Graph> class GraphWriter;
41 /// SDep - Scheduling dependency. This represents one direction of an
42 /// edge in the scheduling DAG.
45 /// Kind - These are the different kinds of scheduling dependencies.
47 Data, ///< Regular data dependence (aka true-dependence).
48 Anti, ///< A register anti-dependedence (aka WAR).
49 Output, ///< A register output-dependence (aka WAW).
50 Order ///< Any other ordering dependency.
54 /// Dep - A pointer to the depending/depended-on SUnit, and an enum
55 /// indicating the kind of the dependency.
56 PointerIntPair<SUnit *, 2, Kind> Dep;
58 /// Contents - A union discriminated by the dependence kind.
60 /// Reg - For Data, Anti, and Output dependencies, the associated
61 /// register. For Data dependencies that don't currently have a register
62 /// assigned, this is set to zero.
65 /// Order - Additional information about Order dependencies.
67 /// isNormalMemory - True if both sides of the dependence
68 /// access memory in non-volatile and fully modeled ways.
69 bool isNormalMemory : 1;
71 /// isMustAlias - True if both sides of the dependence are known to
72 /// access the same memory.
75 /// isArtificial - True if this is an artificial dependency, meaning
76 /// it is not necessary for program correctness, and may be safely
77 /// deleted if necessary.
78 bool isArtificial : 1;
82 /// Latency - The time associated with this edge. Often this is just
83 /// the value of the Latency field of the predecessor, however advanced
84 /// models may provide additional information about specific edges.
88 /// SDep - Construct a null SDep. This is only for use by container
89 /// classes which require default constructors. SUnits may not
90 /// have null SDep edges.
91 SDep() : Dep(0, Data) {}
93 /// SDep - Construct an SDep with the specified values.
94 SDep(SUnit *S, Kind kind, unsigned latency = 1, unsigned Reg = 0,
95 bool isNormalMemory = false, bool isMustAlias = false,
96 bool isArtificial = false)
97 : Dep(S, kind), Contents(), Latency(latency) {
102 "SDep::Anti and SDep::Output must use a non-zero Reg!");
105 assert(!isMustAlias && "isMustAlias only applies with SDep::Order!");
106 assert(!isArtificial && "isArtificial only applies with SDep::Order!");
110 assert(Reg == 0 && "Reg given for non-register dependence!");
111 Contents.Order.isNormalMemory = isNormalMemory;
112 Contents.Order.isMustAlias = isMustAlias;
113 Contents.Order.isArtificial = isArtificial;
118 bool operator==(const SDep &Other) const {
119 if (Dep != Other.Dep || Latency != Other.Latency) return false;
120 switch (Dep.getInt()) {
124 return Contents.Reg == Other.Contents.Reg;
126 return Contents.Order.isNormalMemory ==
127 Other.Contents.Order.isNormalMemory &&
128 Contents.Order.isMustAlias == Other.Contents.Order.isMustAlias &&
129 Contents.Order.isArtificial == Other.Contents.Order.isArtificial;
131 assert(0 && "Invalid dependency kind!");
135 bool operator!=(const SDep &Other) const {
136 return !operator==(Other);
139 /// getLatency - Return the latency value for this edge, which roughly
140 /// means the minimum number of cycles that must elapse between the
141 /// predecessor and the successor, given that they have this edge
143 unsigned getLatency() const {
147 /// setLatency - Set the latency for this edge.
148 void setLatency(unsigned Lat) {
152 //// getSUnit - Return the SUnit to which this edge points.
153 SUnit *getSUnit() const {
154 return Dep.getPointer();
157 //// setSUnit - Assign the SUnit to which this edge points.
158 void setSUnit(SUnit *SU) {
162 /// getKind - Return an enum value representing the kind of the dependence.
163 Kind getKind() const {
167 /// isCtrl - Shorthand for getKind() != SDep::Data.
168 bool isCtrl() const {
169 return getKind() != Data;
172 /// isNormalMemory - Test if this is an Order dependence between two
173 /// memory accesses where both sides of the dependence access memory
174 /// in non-volatile and fully modeled ways.
175 bool isNormalMemory() const {
176 return getKind() == Order && Contents.Order.isNormalMemory;
179 /// isMustAlias - Test if this is an Order dependence that is marked
180 /// as "must alias", meaning that the SUnits at either end of the edge
181 /// have a memory dependence on a known memory location.
182 bool isMustAlias() const {
183 return getKind() == Order && Contents.Order.isMustAlias;
186 /// isArtificial - Test if this is an Order dependence that is marked
187 /// as "artificial", meaning it isn't necessary for correctness.
188 bool isArtificial() const {
189 return getKind() == Order && Contents.Order.isArtificial;
192 /// isAssignedRegDep - Test if this is a Data dependence that is
193 /// associated with a register.
194 bool isAssignedRegDep() const {
195 return getKind() == Data && Contents.Reg != 0;
198 /// getReg - Return the register associated with this edge. This is
199 /// only valid on Data, Anti, and Output edges. On Data edges, this
200 /// value may be zero, meaning there is no associated register.
201 unsigned getReg() const {
202 assert((getKind() == Data || getKind() == Anti || getKind() == Output) &&
203 "getReg called on non-register dependence edge!");
207 /// setReg - Assign the associated register for this edge. This is
208 /// only valid on Data, Anti, and Output edges. On Anti and Output
209 /// edges, this value must not be zero. On Data edges, the value may
210 /// be zero, which would mean that no specific register is associated
212 void setReg(unsigned Reg) {
213 assert((getKind() == Data || getKind() == Anti || getKind() == Output) &&
214 "setReg called on non-register dependence edge!");
215 assert((getKind() != Anti || Reg != 0) &&
216 "SDep::Anti edge cannot use the zero register!");
217 assert((getKind() != Output || Reg != 0) &&
218 "SDep::Output edge cannot use the zero register!");
223 /// SUnit - Scheduling unit. This is a node in the scheduling DAG.
226 SDNode *Node; // Representative node.
227 MachineInstr *Instr; // Alternatively, a MachineInstr.
229 SUnit *OrigNode; // If not this, the node from which
230 // this node was cloned.
232 // Preds/Succs - The SUnits before/after us in the graph. The boolean value
233 // is true if the edge is a token chain edge, false if it is a value edge.
234 SmallVector<SDep, 4> Preds; // All sunit predecessors.
235 SmallVector<SDep, 4> Succs; // All sunit successors.
237 typedef SmallVector<SDep, 4>::iterator pred_iterator;
238 typedef SmallVector<SDep, 4>::iterator succ_iterator;
239 typedef SmallVector<SDep, 4>::const_iterator const_pred_iterator;
240 typedef SmallVector<SDep, 4>::const_iterator const_succ_iterator;
242 unsigned NodeNum; // Entry # of node in the node vector.
243 unsigned NodeQueueId; // Queue id of node.
244 unsigned short Latency; // Node latency.
245 unsigned NumPreds; // # of SDep::Data preds.
246 unsigned NumSuccs; // # of SDep::Data sucss.
247 unsigned NumPredsLeft; // # of preds not scheduled.
248 unsigned NumSuccsLeft; // # of succs not scheduled.
249 bool isTwoAddress : 1; // Is a two-address instruction.
250 bool isCommutable : 1; // Is a commutable instruction.
251 bool hasPhysRegDefs : 1; // Has physreg defs that are being used.
252 bool hasPhysRegClobbers : 1; // Has any physreg defs, used or not.
253 bool isPending : 1; // True once pending.
254 bool isAvailable : 1; // True once available.
255 bool isScheduled : 1; // True once scheduled.
256 bool isScheduleHigh : 1; // True if preferable to schedule high.
257 bool isCloned : 1; // True if this node has been cloned.
259 SmallVector<MachineInstr*, 4> DbgInstrList; // dbg_values referencing this.
261 bool isDepthCurrent : 1; // True if Depth is current.
262 bool isHeightCurrent : 1; // True if Height is current.
263 unsigned Depth; // Node depth.
264 unsigned Height; // Node height.
266 const TargetRegisterClass *CopyDstRC; // Is a special copy node if not null.
267 const TargetRegisterClass *CopySrcRC;
269 /// SUnit - Construct an SUnit for pre-regalloc scheduling to represent
270 /// an SDNode and any nodes flagged to it.
271 SUnit(SDNode *node, unsigned nodenum)
272 : Node(node), Instr(0), OrigNode(0), NodeNum(nodenum),
273 NodeQueueId(0), Latency(0), NumPreds(0), NumSuccs(0), NumPredsLeft(0),
274 NumSuccsLeft(0), isTwoAddress(false), isCommutable(false),
275 hasPhysRegDefs(false), hasPhysRegClobbers(false),
276 isPending(false), isAvailable(false), isScheduled(false),
277 isScheduleHigh(false), isCloned(false),
278 isDepthCurrent(false), isHeightCurrent(false), Depth(0), Height(0),
279 CopyDstRC(NULL), CopySrcRC(NULL) {}
281 /// SUnit - Construct an SUnit for post-regalloc scheduling to represent
283 SUnit(MachineInstr *instr, unsigned nodenum)
284 : Node(0), Instr(instr), OrigNode(0), NodeNum(nodenum),
285 NodeQueueId(0), Latency(0), NumPreds(0), NumSuccs(0), NumPredsLeft(0),
286 NumSuccsLeft(0), isTwoAddress(false), isCommutable(false),
287 hasPhysRegDefs(false), hasPhysRegClobbers(false),
288 isPending(false), isAvailable(false), isScheduled(false),
289 isScheduleHigh(false), isCloned(false),
290 isDepthCurrent(false), isHeightCurrent(false), Depth(0), Height(0),
291 CopyDstRC(NULL), CopySrcRC(NULL) {}
293 /// SUnit - Construct a placeholder SUnit.
295 : Node(0), Instr(0), OrigNode(0), NodeNum(~0u),
296 NodeQueueId(0), Latency(0), NumPreds(0), NumSuccs(0), NumPredsLeft(0),
297 NumSuccsLeft(0), isTwoAddress(false), isCommutable(false),
298 hasPhysRegDefs(false), hasPhysRegClobbers(false),
299 isPending(false), isAvailable(false), isScheduled(false),
300 isScheduleHigh(false), isCloned(false),
301 isDepthCurrent(false), isHeightCurrent(false), Depth(0), Height(0),
302 CopyDstRC(NULL), CopySrcRC(NULL) {}
304 /// setNode - Assign the representative SDNode for this SUnit.
305 /// This may be used during pre-regalloc scheduling.
306 void setNode(SDNode *N) {
307 assert(!Instr && "Setting SDNode of SUnit with MachineInstr!");
311 /// getNode - Return the representative SDNode for this SUnit.
312 /// This may be used during pre-regalloc scheduling.
313 SDNode *getNode() const {
314 assert(!Instr && "Reading SDNode of SUnit with MachineInstr!");
318 /// setInstr - Assign the instruction for the SUnit.
319 /// This may be used during post-regalloc scheduling.
320 void setInstr(MachineInstr *MI) {
321 assert(!Node && "Setting MachineInstr of SUnit with SDNode!");
325 /// getInstr - Return the representative MachineInstr for this SUnit.
326 /// This may be used during post-regalloc scheduling.
327 MachineInstr *getInstr() const {
328 assert(!Node && "Reading MachineInstr of SUnit with SDNode!");
332 /// addPred - This adds the specified edge as a pred of the current node if
333 /// not already. It also adds the current node as a successor of the
335 void addPred(const SDep &D);
337 /// removePred - This removes the specified edge as a pred of the current
338 /// node if it exists. It also removes the current node as a successor of
339 /// the specified node.
340 void removePred(const SDep &D);
342 /// getDepth - Return the depth of this node, which is the length of the
343 /// maximum path up to any node with has no predecessors.
344 unsigned getDepth() const {
346 const_cast<SUnit *>(this)->ComputeDepth();
350 /// getHeight - Return the height of this node, which is the length of the
351 /// maximum path down to any node with has no successors.
352 unsigned getHeight() const {
353 if (!isHeightCurrent)
354 const_cast<SUnit *>(this)->ComputeHeight();
358 /// setDepthToAtLeast - If NewDepth is greater than this node's
359 /// depth value, set it to be the new depth value. This also
360 /// recursively marks successor nodes dirty.
361 void setDepthToAtLeast(unsigned NewDepth);
363 /// setDepthToAtLeast - If NewDepth is greater than this node's
364 /// depth value, set it to be the new height value. This also
365 /// recursively marks predecessor nodes dirty.
366 void setHeightToAtLeast(unsigned NewHeight);
368 /// setDepthDirty - Set a flag in this node to indicate that its
369 /// stored Depth value will require recomputation the next time
370 /// getDepth() is called.
371 void setDepthDirty();
373 /// setHeightDirty - Set a flag in this node to indicate that its
374 /// stored Height value will require recomputation the next time
375 /// getHeight() is called.
376 void setHeightDirty();
378 /// isPred - Test if node N is a predecessor of this node.
379 bool isPred(SUnit *N) {
380 for (unsigned i = 0, e = (unsigned)Preds.size(); i != e; ++i)
381 if (Preds[i].getSUnit() == N)
386 /// isSucc - Test if node N is a successor of this node.
387 bool isSucc(SUnit *N) {
388 for (unsigned i = 0, e = (unsigned)Succs.size(); i != e; ++i)
389 if (Succs[i].getSUnit() == N)
394 void dump(const ScheduleDAG *G) const;
395 void dumpAll(const ScheduleDAG *G) const;
396 void print(raw_ostream &O, const ScheduleDAG *G) const;
400 void ComputeHeight();
403 //===--------------------------------------------------------------------===//
404 /// SchedulingPriorityQueue - This interface is used to plug different
405 /// priorities computation algorithms into the list scheduler. It implements
406 /// the interface of a standard priority queue, where nodes are inserted in
407 /// arbitrary order and returned in priority order. The computation of the
408 /// priority and the representation of the queue are totally up to the
409 /// implementation to decide.
411 class SchedulingPriorityQueue {
414 SchedulingPriorityQueue() : CurCycle(0) {}
415 virtual ~SchedulingPriorityQueue() {}
417 virtual void initNodes(std::vector<SUnit> &SUnits) = 0;
418 virtual void addNode(const SUnit *SU) = 0;
419 virtual void updateNode(const SUnit *SU) = 0;
420 virtual void releaseState() = 0;
422 virtual unsigned size() const = 0;
423 virtual bool empty() const = 0;
424 virtual void push(SUnit *U) = 0;
426 virtual void push_all(const std::vector<SUnit *> &Nodes) = 0;
427 virtual SUnit *pop() = 0;
429 virtual void remove(SUnit *SU) = 0;
431 /// ScheduledNode - As each node is scheduled, this method is invoked. This
432 /// allows the priority function to adjust the priority of related
433 /// unscheduled nodes, for example.
435 virtual void ScheduledNode(SUnit *) {}
437 virtual void UnscheduledNode(SUnit *) {}
439 void setCurCycle(unsigned Cycle) {
443 unsigned getCurCycle() const {
450 MachineBasicBlock *BB; // The block in which to insert instructions
451 MachineBasicBlock::iterator InsertPos;// The position to insert instructions
452 const TargetMachine &TM; // Target processor
453 const TargetInstrInfo *TII; // Target instruction information
454 const TargetRegisterInfo *TRI; // Target processor register info
455 MachineFunction &MF; // Machine function
456 MachineRegisterInfo &MRI; // Virtual/real register map
457 std::vector<SUnit*> Sequence; // The schedule. Null SUnit*'s
458 // represent noop instructions.
459 std::vector<SUnit> SUnits; // The scheduling units.
460 SUnit EntrySU; // Special node for the region entry.
461 SUnit ExitSU; // Special node for the region exit.
463 explicit ScheduleDAG(MachineFunction &mf);
465 virtual ~ScheduleDAG();
467 /// viewGraph - Pop up a GraphViz/gv window with the ScheduleDAG rendered
472 /// EmitSchedule - Insert MachineInstrs into the MachineBasicBlock
473 /// according to the order specified in Sequence.
475 virtual MachineBasicBlock *EmitSchedule() = 0;
477 void dumpSchedule() const;
479 virtual void dumpNode(const SUnit *SU) const = 0;
481 /// getGraphNodeLabel - Return a label for an SUnit node in a visualization
482 /// of the ScheduleDAG.
483 virtual std::string getGraphNodeLabel(const SUnit *SU) const = 0;
485 /// addCustomGraphFeatures - Add custom features for a visualization of
487 virtual void addCustomGraphFeatures(GraphWriter<ScheduleDAG*> &) const {}
490 /// VerifySchedule - Verify that all SUnits were scheduled and that
491 /// their state is consistent.
492 void VerifySchedule(bool isBottomUp);
496 /// Run - perform scheduling.
498 void Run(MachineBasicBlock *bb, MachineBasicBlock::iterator insertPos);
500 /// BuildSchedGraph - Build SUnits and set up their Preds and Succs
501 /// to form the scheduling dependency graph.
503 virtual void BuildSchedGraph(AliasAnalysis *AA) = 0;
505 /// ComputeLatency - Compute node latency.
507 virtual void ComputeLatency(SUnit *SU) = 0;
509 /// ComputeOperandLatency - Override dependence edge latency using
510 /// operand use/def information
512 virtual void ComputeOperandLatency(SUnit *, SUnit *,
515 /// Schedule - Order nodes according to selected style, filling
516 /// in the Sequence member.
518 virtual void Schedule() = 0;
520 /// ForceUnitLatencies - Return true if all scheduling edges should be given
521 /// a latency value of one. The default is to return false; schedulers may
522 /// override this as needed.
523 virtual bool ForceUnitLatencies() const { return false; }
525 /// EmitNoop - Emit a noop instruction.
529 void EmitPhysRegCopy(SUnit *SU, DenseMap<SUnit*, unsigned> &VRBaseMap);
532 class SUnitIterator : public std::iterator<std::forward_iterator_tag,
537 SUnitIterator(SUnit *N, unsigned Op) : Node(N), Operand(Op) {}
539 bool operator==(const SUnitIterator& x) const {
540 return Operand == x.Operand;
542 bool operator!=(const SUnitIterator& x) const { return !operator==(x); }
544 const SUnitIterator &operator=(const SUnitIterator &I) {
545 assert(I.Node==Node && "Cannot assign iterators to two different nodes!");
550 pointer operator*() const {
551 return Node->Preds[Operand].getSUnit();
553 pointer operator->() const { return operator*(); }
555 SUnitIterator& operator++() { // Preincrement
559 SUnitIterator operator++(int) { // Postincrement
560 SUnitIterator tmp = *this; ++*this; return tmp;
563 static SUnitIterator begin(SUnit *N) { return SUnitIterator(N, 0); }
564 static SUnitIterator end (SUnit *N) {
565 return SUnitIterator(N, (unsigned)N->Preds.size());
568 unsigned getOperand() const { return Operand; }
569 const SUnit *getNode() const { return Node; }
570 /// isCtrlDep - Test if this is not an SDep::Data dependence.
571 bool isCtrlDep() const {
572 return getSDep().isCtrl();
574 bool isArtificialDep() const {
575 return getSDep().isArtificial();
577 const SDep &getSDep() const {
578 return Node->Preds[Operand];
582 template <> struct GraphTraits<SUnit*> {
583 typedef SUnit NodeType;
584 typedef SUnitIterator ChildIteratorType;
585 static inline NodeType *getEntryNode(SUnit *N) { return N; }
586 static inline ChildIteratorType child_begin(NodeType *N) {
587 return SUnitIterator::begin(N);
589 static inline ChildIteratorType child_end(NodeType *N) {
590 return SUnitIterator::end(N);
594 template <> struct GraphTraits<ScheduleDAG*> : public GraphTraits<SUnit*> {
595 typedef std::vector<SUnit>::iterator nodes_iterator;
596 static nodes_iterator nodes_begin(ScheduleDAG *G) {
597 return G->SUnits.begin();
599 static nodes_iterator nodes_end(ScheduleDAG *G) {
600 return G->SUnits.end();
604 /// ScheduleDAGTopologicalSort is a class that computes a topological
605 /// ordering for SUnits and provides methods for dynamically updating
606 /// the ordering as new edges are added.
608 /// This allows a very fast implementation of IsReachable, for example.
610 class ScheduleDAGTopologicalSort {
611 /// SUnits - A reference to the ScheduleDAG's SUnits.
612 std::vector<SUnit> &SUnits;
614 /// Index2Node - Maps topological index to the node number.
615 std::vector<int> Index2Node;
616 /// Node2Index - Maps the node number to its topological index.
617 std::vector<int> Node2Index;
618 /// Visited - a set of nodes visited during a DFS traversal.
621 /// DFS - make a DFS traversal and mark all nodes affected by the
622 /// edge insertion. These nodes will later get new topological indexes
623 /// by means of the Shift method.
624 void DFS(const SUnit *SU, int UpperBound, bool& HasLoop);
626 /// Shift - reassign topological indexes for the nodes in the DAG
627 /// to preserve the topological ordering.
628 void Shift(BitVector& Visited, int LowerBound, int UpperBound);
630 /// Allocate - assign the topological index to the node n.
631 void Allocate(int n, int index);
634 explicit ScheduleDAGTopologicalSort(std::vector<SUnit> &SUnits);
636 /// InitDAGTopologicalSorting - create the initial topological
637 /// ordering from the DAG to be scheduled.
638 void InitDAGTopologicalSorting();
640 /// IsReachable - Checks if SU is reachable from TargetSU.
641 bool IsReachable(const SUnit *SU, const SUnit *TargetSU);
643 /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU
644 /// will create a cycle.
645 bool WillCreateCycle(SUnit *SU, SUnit *TargetSU);
647 /// AddPred - Updates the topological ordering to accomodate an edge
648 /// to be added from SUnit X to SUnit Y.
649 void AddPred(SUnit *Y, SUnit *X);
651 /// RemovePred - Updates the topological ordering to accomodate an
652 /// an edge to be removed from the specified node N from the predecessors
653 /// of the current node M.
654 void RemovePred(SUnit *M, SUnit *N);
656 typedef std::vector<int>::iterator iterator;
657 typedef std::vector<int>::const_iterator const_iterator;
658 iterator begin() { return Index2Node.begin(); }
659 const_iterator begin() const { return Index2Node.begin(); }
660 iterator end() { return Index2Node.end(); }
661 const_iterator end() const { return Index2Node.end(); }
663 typedef std::vector<int>::reverse_iterator reverse_iterator;
664 typedef std::vector<int>::const_reverse_iterator const_reverse_iterator;
665 reverse_iterator rbegin() { return Index2Node.rbegin(); }
666 const_reverse_iterator rbegin() const { return Index2Node.rbegin(); }
667 reverse_iterator rend() { return Index2Node.rend(); }
668 const_reverse_iterator rend() const { return Index2Node.rend(); }