It turns out most of the thumb2 instructions are not allowed to touch SP. The semanti...
[llvm/avr.git] / lib / CodeGen / ScheduleDAG.cpp
blobbc48dc21ab17c6d3c5c8300f8828122a0b56d7c5
1 //===---- ScheduleDAG.cpp - Implement the ScheduleDAG class ---------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This implements the ScheduleDAG class, which is a base class used by
11 // scheduling implementation classes.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "pre-RA-sched"
16 #include "llvm/CodeGen/ScheduleDAG.h"
17 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
18 #include "llvm/Target/TargetMachine.h"
19 #include "llvm/Target/TargetInstrInfo.h"
20 #include "llvm/Target/TargetRegisterInfo.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Support/raw_ostream.h"
23 #include <climits>
24 using namespace llvm;
26 ScheduleDAG::ScheduleDAG(MachineFunction &mf)
27 : TM(mf.getTarget()),
28 TII(TM.getInstrInfo()),
29 TRI(TM.getRegisterInfo()),
30 TLI(TM.getTargetLowering()),
31 MF(mf), MRI(mf.getRegInfo()),
32 ConstPool(MF.getConstantPool()),
33 EntrySU(), ExitSU() {
36 ScheduleDAG::~ScheduleDAG() {}
38 /// dump - dump the schedule.
39 void ScheduleDAG::dumpSchedule() const {
40 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
41 if (SUnit *SU = Sequence[i])
42 SU->dump(this);
43 else
44 errs() << "**** NOOP ****\n";
49 /// Run - perform scheduling.
50 ///
51 void ScheduleDAG::Run(MachineBasicBlock *bb,
52 MachineBasicBlock::iterator insertPos) {
53 BB = bb;
54 InsertPos = insertPos;
56 SUnits.clear();
57 Sequence.clear();
58 EntrySU = SUnit();
59 ExitSU = SUnit();
61 Schedule();
63 DOUT << "*** Final schedule ***\n";
64 DEBUG(dumpSchedule());
65 DOUT << "\n";
68 /// addPred - This adds the specified edge as a pred of the current node if
69 /// not already. It also adds the current node as a successor of the
70 /// specified node.
71 void SUnit::addPred(const SDep &D) {
72 // If this node already has this depenence, don't add a redundant one.
73 for (SmallVector<SDep, 4>::const_iterator I = Preds.begin(), E = Preds.end();
74 I != E; ++I)
75 if (*I == D)
76 return;
77 // Now add a corresponding succ to N.
78 SDep P = D;
79 P.setSUnit(this);
80 SUnit *N = D.getSUnit();
81 // Update the bookkeeping.
82 if (D.getKind() == SDep::Data) {
83 ++NumPreds;
84 ++N->NumSuccs;
86 if (!N->isScheduled)
87 ++NumPredsLeft;
88 if (!isScheduled)
89 ++N->NumSuccsLeft;
90 Preds.push_back(D);
91 N->Succs.push_back(P);
92 if (P.getLatency() != 0) {
93 this->setDepthDirty();
94 N->setHeightDirty();
98 /// removePred - This removes the specified edge as a pred of the current
99 /// node if it exists. It also removes the current node as a successor of
100 /// the specified node.
101 void SUnit::removePred(const SDep &D) {
102 // Find the matching predecessor.
103 for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end();
104 I != E; ++I)
105 if (*I == D) {
106 bool FoundSucc = false;
107 // Find the corresponding successor in N.
108 SDep P = D;
109 P.setSUnit(this);
110 SUnit *N = D.getSUnit();
111 for (SmallVector<SDep, 4>::iterator II = N->Succs.begin(),
112 EE = N->Succs.end(); II != EE; ++II)
113 if (*II == P) {
114 FoundSucc = true;
115 N->Succs.erase(II);
116 break;
118 assert(FoundSucc && "Mismatching preds / succs lists!");
119 Preds.erase(I);
120 // Update the bookkeeping.
121 if (P.getKind() == SDep::Data) {
122 --NumPreds;
123 --N->NumSuccs;
125 if (!N->isScheduled)
126 --NumPredsLeft;
127 if (!isScheduled)
128 --N->NumSuccsLeft;
129 if (P.getLatency() != 0) {
130 this->setDepthDirty();
131 N->setHeightDirty();
133 return;
137 void SUnit::setDepthDirty() {
138 if (!isDepthCurrent) return;
139 SmallVector<SUnit*, 8> WorkList;
140 WorkList.push_back(this);
141 do {
142 SUnit *SU = WorkList.pop_back_val();
143 SU->isDepthCurrent = false;
144 for (SUnit::const_succ_iterator I = SU->Succs.begin(),
145 E = SU->Succs.end(); I != E; ++I) {
146 SUnit *SuccSU = I->getSUnit();
147 if (SuccSU->isDepthCurrent)
148 WorkList.push_back(SuccSU);
150 } while (!WorkList.empty());
153 void SUnit::setHeightDirty() {
154 if (!isHeightCurrent) return;
155 SmallVector<SUnit*, 8> WorkList;
156 WorkList.push_back(this);
157 do {
158 SUnit *SU = WorkList.pop_back_val();
159 SU->isHeightCurrent = false;
160 for (SUnit::const_pred_iterator I = SU->Preds.begin(),
161 E = SU->Preds.end(); I != E; ++I) {
162 SUnit *PredSU = I->getSUnit();
163 if (PredSU->isHeightCurrent)
164 WorkList.push_back(PredSU);
166 } while (!WorkList.empty());
169 /// setDepthToAtLeast - Update this node's successors to reflect the
170 /// fact that this node's depth just increased.
172 void SUnit::setDepthToAtLeast(unsigned NewDepth) {
173 if (NewDepth <= getDepth())
174 return;
175 setDepthDirty();
176 Depth = NewDepth;
177 isDepthCurrent = true;
180 /// setHeightToAtLeast - Update this node's predecessors to reflect the
181 /// fact that this node's height just increased.
183 void SUnit::setHeightToAtLeast(unsigned NewHeight) {
184 if (NewHeight <= getHeight())
185 return;
186 setHeightDirty();
187 Height = NewHeight;
188 isHeightCurrent = true;
191 /// ComputeDepth - Calculate the maximal path from the node to the exit.
193 void SUnit::ComputeDepth() {
194 SmallVector<SUnit*, 8> WorkList;
195 WorkList.push_back(this);
196 do {
197 SUnit *Cur = WorkList.back();
199 bool Done = true;
200 unsigned MaxPredDepth = 0;
201 for (SUnit::const_pred_iterator I = Cur->Preds.begin(),
202 E = Cur->Preds.end(); I != E; ++I) {
203 SUnit *PredSU = I->getSUnit();
204 if (PredSU->isDepthCurrent)
205 MaxPredDepth = std::max(MaxPredDepth,
206 PredSU->Depth + I->getLatency());
207 else {
208 Done = false;
209 WorkList.push_back(PredSU);
213 if (Done) {
214 WorkList.pop_back();
215 if (MaxPredDepth != Cur->Depth) {
216 Cur->setDepthDirty();
217 Cur->Depth = MaxPredDepth;
219 Cur->isDepthCurrent = true;
221 } while (!WorkList.empty());
224 /// ComputeHeight - Calculate the maximal path from the node to the entry.
226 void SUnit::ComputeHeight() {
227 SmallVector<SUnit*, 8> WorkList;
228 WorkList.push_back(this);
229 do {
230 SUnit *Cur = WorkList.back();
232 bool Done = true;
233 unsigned MaxSuccHeight = 0;
234 for (SUnit::const_succ_iterator I = Cur->Succs.begin(),
235 E = Cur->Succs.end(); I != E; ++I) {
236 SUnit *SuccSU = I->getSUnit();
237 if (SuccSU->isHeightCurrent)
238 MaxSuccHeight = std::max(MaxSuccHeight,
239 SuccSU->Height + I->getLatency());
240 else {
241 Done = false;
242 WorkList.push_back(SuccSU);
246 if (Done) {
247 WorkList.pop_back();
248 if (MaxSuccHeight != Cur->Height) {
249 Cur->setHeightDirty();
250 Cur->Height = MaxSuccHeight;
252 Cur->isHeightCurrent = true;
254 } while (!WorkList.empty());
257 /// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
258 /// a group of nodes flagged together.
259 void SUnit::dump(const ScheduleDAG *G) const {
260 errs() << "SU(" << NodeNum << "): ";
261 G->dumpNode(this);
264 void SUnit::dumpAll(const ScheduleDAG *G) const {
265 dump(G);
267 errs() << " # preds left : " << NumPredsLeft << "\n";
268 errs() << " # succs left : " << NumSuccsLeft << "\n";
269 errs() << " Latency : " << Latency << "\n";
270 errs() << " Depth : " << Depth << "\n";
271 errs() << " Height : " << Height << "\n";
273 if (Preds.size() != 0) {
274 errs() << " Predecessors:\n";
275 for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
276 I != E; ++I) {
277 errs() << " ";
278 switch (I->getKind()) {
279 case SDep::Data: errs() << "val "; break;
280 case SDep::Anti: errs() << "anti"; break;
281 case SDep::Output: errs() << "out "; break;
282 case SDep::Order: errs() << "ch "; break;
284 errs() << "#";
285 errs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
286 if (I->isArtificial())
287 errs() << " *";
288 errs() << "\n";
291 if (Succs.size() != 0) {
292 errs() << " Successors:\n";
293 for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
294 I != E; ++I) {
295 errs() << " ";
296 switch (I->getKind()) {
297 case SDep::Data: errs() << "val "; break;
298 case SDep::Anti: errs() << "anti"; break;
299 case SDep::Output: errs() << "out "; break;
300 case SDep::Order: errs() << "ch "; break;
302 errs() << "#";
303 errs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
304 if (I->isArtificial())
305 errs() << " *";
306 errs() << "\n";
309 errs() << "\n";
312 #ifndef NDEBUG
313 /// VerifySchedule - Verify that all SUnits were scheduled and that
314 /// their state is consistent.
316 void ScheduleDAG::VerifySchedule(bool isBottomUp) {
317 bool AnyNotSched = false;
318 unsigned DeadNodes = 0;
319 unsigned Noops = 0;
320 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
321 if (!SUnits[i].isScheduled) {
322 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
323 ++DeadNodes;
324 continue;
326 if (!AnyNotSched)
327 errs() << "*** Scheduling failed! ***\n";
328 SUnits[i].dump(this);
329 errs() << "has not been scheduled!\n";
330 AnyNotSched = true;
332 if (SUnits[i].isScheduled &&
333 (isBottomUp ? SUnits[i].getHeight() : SUnits[i].getHeight()) >
334 unsigned(INT_MAX)) {
335 if (!AnyNotSched)
336 errs() << "*** Scheduling failed! ***\n";
337 SUnits[i].dump(this);
338 errs() << "has an unexpected "
339 << (isBottomUp ? "Height" : "Depth") << " value!\n";
340 AnyNotSched = true;
342 if (isBottomUp) {
343 if (SUnits[i].NumSuccsLeft != 0) {
344 if (!AnyNotSched)
345 errs() << "*** Scheduling failed! ***\n";
346 SUnits[i].dump(this);
347 errs() << "has successors left!\n";
348 AnyNotSched = true;
350 } else {
351 if (SUnits[i].NumPredsLeft != 0) {
352 if (!AnyNotSched)
353 errs() << "*** Scheduling failed! ***\n";
354 SUnits[i].dump(this);
355 errs() << "has predecessors left!\n";
356 AnyNotSched = true;
360 for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
361 if (!Sequence[i])
362 ++Noops;
363 assert(!AnyNotSched);
364 assert(Sequence.size() + DeadNodes - Noops == SUnits.size() &&
365 "The number of nodes scheduled doesn't match the expected number!");
367 #endif
369 /// InitDAGTopologicalSorting - create the initial topological
370 /// ordering from the DAG to be scheduled.
372 /// The idea of the algorithm is taken from
373 /// "Online algorithms for managing the topological order of
374 /// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
375 /// This is the MNR algorithm, which was first introduced by
376 /// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
377 /// "Maintaining a topological order under edge insertions".
379 /// Short description of the algorithm:
381 /// Topological ordering, ord, of a DAG maps each node to a topological
382 /// index so that for all edges X->Y it is the case that ord(X) < ord(Y).
384 /// This means that if there is a path from the node X to the node Z,
385 /// then ord(X) < ord(Z).
387 /// This property can be used to check for reachability of nodes:
388 /// if Z is reachable from X, then an insertion of the edge Z->X would
389 /// create a cycle.
391 /// The algorithm first computes a topological ordering for the DAG by
392 /// initializing the Index2Node and Node2Index arrays and then tries to keep
393 /// the ordering up-to-date after edge insertions by reordering the DAG.
395 /// On insertion of the edge X->Y, the algorithm first marks by calling DFS
396 /// the nodes reachable from Y, and then shifts them using Shift to lie
397 /// immediately after X in Index2Node.
398 void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() {
399 unsigned DAGSize = SUnits.size();
400 std::vector<SUnit*> WorkList;
401 WorkList.reserve(DAGSize);
403 Index2Node.resize(DAGSize);
404 Node2Index.resize(DAGSize);
406 // Initialize the data structures.
407 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
408 SUnit *SU = &SUnits[i];
409 int NodeNum = SU->NodeNum;
410 unsigned Degree = SU->Succs.size();
411 // Temporarily use the Node2Index array as scratch space for degree counts.
412 Node2Index[NodeNum] = Degree;
414 // Is it a node without dependencies?
415 if (Degree == 0) {
416 assert(SU->Succs.empty() && "SUnit should have no successors");
417 // Collect leaf nodes.
418 WorkList.push_back(SU);
422 int Id = DAGSize;
423 while (!WorkList.empty()) {
424 SUnit *SU = WorkList.back();
425 WorkList.pop_back();
426 Allocate(SU->NodeNum, --Id);
427 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
428 I != E; ++I) {
429 SUnit *SU = I->getSUnit();
430 if (!--Node2Index[SU->NodeNum])
431 // If all dependencies of the node are processed already,
432 // then the node can be computed now.
433 WorkList.push_back(SU);
437 Visited.resize(DAGSize);
439 #ifndef NDEBUG
440 // Check correctness of the ordering
441 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
442 SUnit *SU = &SUnits[i];
443 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
444 I != E; ++I) {
445 assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] &&
446 "Wrong topological sorting");
449 #endif
452 /// AddPred - Updates the topological ordering to accomodate an edge
453 /// to be added from SUnit X to SUnit Y.
454 void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) {
455 int UpperBound, LowerBound;
456 LowerBound = Node2Index[Y->NodeNum];
457 UpperBound = Node2Index[X->NodeNum];
458 bool HasLoop = false;
459 // Is Ord(X) < Ord(Y) ?
460 if (LowerBound < UpperBound) {
461 // Update the topological order.
462 Visited.reset();
463 DFS(Y, UpperBound, HasLoop);
464 assert(!HasLoop && "Inserted edge creates a loop!");
465 // Recompute topological indexes.
466 Shift(Visited, LowerBound, UpperBound);
470 /// RemovePred - Updates the topological ordering to accomodate an
471 /// an edge to be removed from the specified node N from the predecessors
472 /// of the current node M.
473 void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) {
474 // InitDAGTopologicalSorting();
477 /// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark
478 /// all nodes affected by the edge insertion. These nodes will later get new
479 /// topological indexes by means of the Shift method.
480 void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound,
481 bool& HasLoop) {
482 std::vector<const SUnit*> WorkList;
483 WorkList.reserve(SUnits.size());
485 WorkList.push_back(SU);
486 do {
487 SU = WorkList.back();
488 WorkList.pop_back();
489 Visited.set(SU->NodeNum);
490 for (int I = SU->Succs.size()-1; I >= 0; --I) {
491 int s = SU->Succs[I].getSUnit()->NodeNum;
492 if (Node2Index[s] == UpperBound) {
493 HasLoop = true;
494 return;
496 // Visit successors if not already and in affected region.
497 if (!Visited.test(s) && Node2Index[s] < UpperBound) {
498 WorkList.push_back(SU->Succs[I].getSUnit());
501 } while (!WorkList.empty());
504 /// Shift - Renumber the nodes so that the topological ordering is
505 /// preserved.
506 void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
507 int UpperBound) {
508 std::vector<int> L;
509 int shift = 0;
510 int i;
512 for (i = LowerBound; i <= UpperBound; ++i) {
513 // w is node at topological index i.
514 int w = Index2Node[i];
515 if (Visited.test(w)) {
516 // Unmark.
517 Visited.reset(w);
518 L.push_back(w);
519 shift = shift + 1;
520 } else {
521 Allocate(w, i - shift);
525 for (unsigned j = 0; j < L.size(); ++j) {
526 Allocate(L[j], i - shift);
527 i = i + 1;
532 /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
533 /// create a cycle.
534 bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
535 if (IsReachable(TargetSU, SU))
536 return true;
537 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
538 I != E; ++I)
539 if (I->isAssignedRegDep() &&
540 IsReachable(TargetSU, I->getSUnit()))
541 return true;
542 return false;
545 /// IsReachable - Checks if SU is reachable from TargetSU.
546 bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU,
547 const SUnit *TargetSU) {
548 // If insertion of the edge SU->TargetSU would create a cycle
549 // then there is a path from TargetSU to SU.
550 int UpperBound, LowerBound;
551 LowerBound = Node2Index[TargetSU->NodeNum];
552 UpperBound = Node2Index[SU->NodeNum];
553 bool HasLoop = false;
554 // Is Ord(TargetSU) < Ord(SU) ?
555 if (LowerBound < UpperBound) {
556 Visited.reset();
557 // There may be a path from TargetSU to SU. Check for it.
558 DFS(TargetSU, UpperBound, HasLoop);
560 return HasLoop;
563 /// Allocate - assign the topological index to the node n.
564 void ScheduleDAGTopologicalSort::Allocate(int n, int index) {
565 Node2Index[n] = index;
566 Index2Node[index] = n;
569 ScheduleDAGTopologicalSort::ScheduleDAGTopologicalSort(
570 std::vector<SUnit> &sunits)
571 : SUnits(sunits) {}
573 ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {}