1
//===---- ScheduleDAG.cpp - Implement the ScheduleDAG class ---------------===//
3
// The LLVM Compiler Infrastructure
5
// This file is distributed under the University of Illinois Open Source
6
// License. See LICENSE.TXT for details.
8
//===----------------------------------------------------------------------===//
10
// This implements 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"
26
ScheduleDAG::ScheduleDAG(MachineFunction &mf)
28
TII(TM.getInstrInfo()),
29
TRI(TM.getRegisterInfo()),
30
MF(mf), MRI(mf.getRegInfo()),
34
ScheduleDAG::~ScheduleDAG() {}
36
/// dump - dump the schedule.
37
void ScheduleDAG::dumpSchedule() const {
38
for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
39
if (SUnit *SU = Sequence[i])
42
dbgs() << "**** NOOP ****\n";
47
/// Run - perform scheduling.
49
void ScheduleDAG::Run(MachineBasicBlock *bb,
50
MachineBasicBlock::iterator insertPos) {
52
InsertPos = insertPos;
62
dbgs() << "*** Final schedule ***\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
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();
77
// Now add a corresponding succ to N.
80
SUnit *N = D.getSUnit();
81
// Update the bookkeeping.
82
if (D.getKind() == SDep::Data) {
83
assert(NumPreds < UINT_MAX && "NumPreds will overflow!");
84
assert(N->NumSuccs < UINT_MAX && "NumSuccs will overflow!");
88
if (!N->isScheduled) {
89
assert(NumPredsLeft < UINT_MAX && "NumPredsLeft will overflow!");
93
assert(N->NumSuccsLeft < UINT_MAX && "NumSuccsLeft will overflow!");
97
N->Succs.push_back(P);
98
if (P.getLatency() != 0) {
99
this->setDepthDirty();
104
/// removePred - This removes the specified edge as a pred of the current
105
/// node if it exists. It also removes the current node as a successor of
106
/// the specified node.
107
void SUnit::removePred(const SDep &D) {
108
// Find the matching predecessor.
109
for (SmallVector<SDep, 4>::iterator I = Preds.begin(), E = Preds.end();
112
bool FoundSucc = false;
113
// Find the corresponding successor in N.
116
SUnit *N = D.getSUnit();
117
for (SmallVector<SDep, 4>::iterator II = N->Succs.begin(),
118
EE = N->Succs.end(); II != EE; ++II)
124
assert(FoundSucc && "Mismatching preds / succs lists!");
126
// Update the bookkeeping.
127
if (P.getKind() == SDep::Data) {
128
assert(NumPreds > 0 && "NumPreds will underflow!");
129
assert(N->NumSuccs > 0 && "NumSuccs will underflow!");
133
if (!N->isScheduled) {
134
assert(NumPredsLeft > 0 && "NumPredsLeft will underflow!");
138
assert(N->NumSuccsLeft > 0 && "NumSuccsLeft will underflow!");
141
if (P.getLatency() != 0) {
142
this->setDepthDirty();
149
void SUnit::setDepthDirty() {
150
if (!isDepthCurrent) return;
151
SmallVector<SUnit*, 8> WorkList;
152
WorkList.push_back(this);
154
SUnit *SU = WorkList.pop_back_val();
155
SU->isDepthCurrent = false;
156
for (SUnit::const_succ_iterator I = SU->Succs.begin(),
157
E = SU->Succs.end(); I != E; ++I) {
158
SUnit *SuccSU = I->getSUnit();
159
if (SuccSU->isDepthCurrent)
160
WorkList.push_back(SuccSU);
162
} while (!WorkList.empty());
165
void SUnit::setHeightDirty() {
166
if (!isHeightCurrent) return;
167
SmallVector<SUnit*, 8> WorkList;
168
WorkList.push_back(this);
170
SUnit *SU = WorkList.pop_back_val();
171
SU->isHeightCurrent = false;
172
for (SUnit::const_pred_iterator I = SU->Preds.begin(),
173
E = SU->Preds.end(); I != E; ++I) {
174
SUnit *PredSU = I->getSUnit();
175
if (PredSU->isHeightCurrent)
176
WorkList.push_back(PredSU);
178
} while (!WorkList.empty());
181
/// setDepthToAtLeast - Update this node's successors to reflect the
182
/// fact that this node's depth just increased.
184
void SUnit::setDepthToAtLeast(unsigned NewDepth) {
185
if (NewDepth <= getDepth())
189
isDepthCurrent = true;
192
/// setHeightToAtLeast - Update this node's predecessors to reflect the
193
/// fact that this node's height just increased.
195
void SUnit::setHeightToAtLeast(unsigned NewHeight) {
196
if (NewHeight <= getHeight())
200
isHeightCurrent = true;
203
/// ComputeDepth - Calculate the maximal path from the node to the exit.
205
void SUnit::ComputeDepth() {
206
SmallVector<SUnit*, 8> WorkList;
207
WorkList.push_back(this);
209
SUnit *Cur = WorkList.back();
212
unsigned MaxPredDepth = 0;
213
for (SUnit::const_pred_iterator I = Cur->Preds.begin(),
214
E = Cur->Preds.end(); I != E; ++I) {
215
SUnit *PredSU = I->getSUnit();
216
if (PredSU->isDepthCurrent)
217
MaxPredDepth = std::max(MaxPredDepth,
218
PredSU->Depth + I->getLatency());
221
WorkList.push_back(PredSU);
227
if (MaxPredDepth != Cur->Depth) {
228
Cur->setDepthDirty();
229
Cur->Depth = MaxPredDepth;
231
Cur->isDepthCurrent = true;
233
} while (!WorkList.empty());
236
/// ComputeHeight - Calculate the maximal path from the node to the entry.
238
void SUnit::ComputeHeight() {
239
SmallVector<SUnit*, 8> WorkList;
240
WorkList.push_back(this);
242
SUnit *Cur = WorkList.back();
245
unsigned MaxSuccHeight = 0;
246
for (SUnit::const_succ_iterator I = Cur->Succs.begin(),
247
E = Cur->Succs.end(); I != E; ++I) {
248
SUnit *SuccSU = I->getSUnit();
249
if (SuccSU->isHeightCurrent)
250
MaxSuccHeight = std::max(MaxSuccHeight,
251
SuccSU->Height + I->getLatency());
254
WorkList.push_back(SuccSU);
260
if (MaxSuccHeight != Cur->Height) {
261
Cur->setHeightDirty();
262
Cur->Height = MaxSuccHeight;
264
Cur->isHeightCurrent = true;
266
} while (!WorkList.empty());
269
/// SUnit - Scheduling unit. It's an wrapper around either a single SDNode or
270
/// a group of nodes flagged together.
271
void SUnit::dump(const ScheduleDAG *G) const {
272
dbgs() << "SU(" << NodeNum << "): ";
276
void SUnit::dumpAll(const ScheduleDAG *G) const {
279
dbgs() << " # preds left : " << NumPredsLeft << "\n";
280
dbgs() << " # succs left : " << NumSuccsLeft << "\n";
281
dbgs() << " Latency : " << Latency << "\n";
282
dbgs() << " Depth : " << Depth << "\n";
283
dbgs() << " Height : " << Height << "\n";
285
if (Preds.size() != 0) {
286
dbgs() << " Predecessors:\n";
287
for (SUnit::const_succ_iterator I = Preds.begin(), E = Preds.end();
290
switch (I->getKind()) {
291
case SDep::Data: dbgs() << "val "; break;
292
case SDep::Anti: dbgs() << "anti"; break;
293
case SDep::Output: dbgs() << "out "; break;
294
case SDep::Order: dbgs() << "ch "; break;
297
dbgs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
298
if (I->isArtificial())
300
dbgs() << ": Latency=" << I->getLatency();
304
if (Succs.size() != 0) {
305
dbgs() << " Successors:\n";
306
for (SUnit::const_succ_iterator I = Succs.begin(), E = Succs.end();
309
switch (I->getKind()) {
310
case SDep::Data: dbgs() << "val "; break;
311
case SDep::Anti: dbgs() << "anti"; break;
312
case SDep::Output: dbgs() << "out "; break;
313
case SDep::Order: dbgs() << "ch "; break;
316
dbgs() << I->getSUnit() << " - SU(" << I->getSUnit()->NodeNum << ")";
317
if (I->isArtificial())
319
dbgs() << ": Latency=" << I->getLatency();
327
/// VerifySchedule - Verify that all SUnits were scheduled and that
328
/// their state is consistent.
330
void ScheduleDAG::VerifySchedule(bool isBottomUp) {
331
bool AnyNotSched = false;
332
unsigned DeadNodes = 0;
334
for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
335
if (!SUnits[i].isScheduled) {
336
if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
341
dbgs() << "*** Scheduling failed! ***\n";
342
SUnits[i].dump(this);
343
dbgs() << "has not been scheduled!\n";
346
if (SUnits[i].isScheduled &&
347
(isBottomUp ? SUnits[i].getHeight() : SUnits[i].getDepth()) >
350
dbgs() << "*** Scheduling failed! ***\n";
351
SUnits[i].dump(this);
352
dbgs() << "has an unexpected "
353
<< (isBottomUp ? "Height" : "Depth") << " value!\n";
357
if (SUnits[i].NumSuccsLeft != 0) {
359
dbgs() << "*** Scheduling failed! ***\n";
360
SUnits[i].dump(this);
361
dbgs() << "has successors left!\n";
365
if (SUnits[i].NumPredsLeft != 0) {
367
dbgs() << "*** Scheduling failed! ***\n";
368
SUnits[i].dump(this);
369
dbgs() << "has predecessors left!\n";
374
for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
377
assert(!AnyNotSched);
378
assert(Sequence.size() + DeadNodes - Noops == SUnits.size() &&
379
"The number of nodes scheduled doesn't match the expected number!");
383
/// InitDAGTopologicalSorting - create the initial topological
384
/// ordering from the DAG to be scheduled.
386
/// The idea of the algorithm is taken from
387
/// "Online algorithms for managing the topological order of
388
/// a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
389
/// This is the MNR algorithm, which was first introduced by
390
/// A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
391
/// "Maintaining a topological order under edge insertions".
393
/// Short description of the algorithm:
395
/// Topological ordering, ord, of a DAG maps each node to a topological
396
/// index so that for all edges X->Y it is the case that ord(X) < ord(Y).
398
/// This means that if there is a path from the node X to the node Z,
399
/// then ord(X) < ord(Z).
401
/// This property can be used to check for reachability of nodes:
402
/// if Z is reachable from X, then an insertion of the edge Z->X would
405
/// The algorithm first computes a topological ordering for the DAG by
406
/// initializing the Index2Node and Node2Index arrays and then tries to keep
407
/// the ordering up-to-date after edge insertions by reordering the DAG.
409
/// On insertion of the edge X->Y, the algorithm first marks by calling DFS
410
/// the nodes reachable from Y, and then shifts them using Shift to lie
411
/// immediately after X in Index2Node.
412
void ScheduleDAGTopologicalSort::InitDAGTopologicalSorting() {
413
unsigned DAGSize = SUnits.size();
414
std::vector<SUnit*> WorkList;
415
WorkList.reserve(DAGSize);
417
Index2Node.resize(DAGSize);
418
Node2Index.resize(DAGSize);
420
// Initialize the data structures.
421
for (unsigned i = 0, e = DAGSize; i != e; ++i) {
422
SUnit *SU = &SUnits[i];
423
int NodeNum = SU->NodeNum;
424
unsigned Degree = SU->Succs.size();
425
// Temporarily use the Node2Index array as scratch space for degree counts.
426
Node2Index[NodeNum] = Degree;
428
// Is it a node without dependencies?
430
assert(SU->Succs.empty() && "SUnit should have no successors");
431
// Collect leaf nodes.
432
WorkList.push_back(SU);
437
while (!WorkList.empty()) {
438
SUnit *SU = WorkList.back();
440
Allocate(SU->NodeNum, --Id);
441
for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
443
SUnit *SU = I->getSUnit();
444
if (!--Node2Index[SU->NodeNum])
445
// If all dependencies of the node are processed already,
446
// then the node can be computed now.
447
WorkList.push_back(SU);
451
Visited.resize(DAGSize);
454
// Check correctness of the ordering
455
for (unsigned i = 0, e = DAGSize; i != e; ++i) {
456
SUnit *SU = &SUnits[i];
457
for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
459
assert(Node2Index[SU->NodeNum] > Node2Index[I->getSUnit()->NodeNum] &&
460
"Wrong topological sorting");
466
/// AddPred - Updates the topological ordering to accomodate an edge
467
/// to be added from SUnit X to SUnit Y.
468
void ScheduleDAGTopologicalSort::AddPred(SUnit *Y, SUnit *X) {
469
int UpperBound, LowerBound;
470
LowerBound = Node2Index[Y->NodeNum];
471
UpperBound = Node2Index[X->NodeNum];
472
bool HasLoop = false;
473
// Is Ord(X) < Ord(Y) ?
474
if (LowerBound < UpperBound) {
475
// Update the topological order.
477
DFS(Y, UpperBound, HasLoop);
478
assert(!HasLoop && "Inserted edge creates a loop!");
479
// Recompute topological indexes.
480
Shift(Visited, LowerBound, UpperBound);
484
/// RemovePred - Updates the topological ordering to accomodate an
485
/// an edge to be removed from the specified node N from the predecessors
486
/// of the current node M.
487
void ScheduleDAGTopologicalSort::RemovePred(SUnit *M, SUnit *N) {
488
// InitDAGTopologicalSorting();
491
/// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark
492
/// all nodes affected by the edge insertion. These nodes will later get new
493
/// topological indexes by means of the Shift method.
494
void ScheduleDAGTopologicalSort::DFS(const SUnit *SU, int UpperBound,
496
std::vector<const SUnit*> WorkList;
497
WorkList.reserve(SUnits.size());
499
WorkList.push_back(SU);
501
SU = WorkList.back();
503
Visited.set(SU->NodeNum);
504
for (int I = SU->Succs.size()-1; I >= 0; --I) {
505
int s = SU->Succs[I].getSUnit()->NodeNum;
506
if (Node2Index[s] == UpperBound) {
510
// Visit successors if not already and in affected region.
511
if (!Visited.test(s) && Node2Index[s] < UpperBound) {
512
WorkList.push_back(SU->Succs[I].getSUnit());
515
} while (!WorkList.empty());
518
/// Shift - Renumber the nodes so that the topological ordering is
520
void ScheduleDAGTopologicalSort::Shift(BitVector& Visited, int LowerBound,
526
for (i = LowerBound; i <= UpperBound; ++i) {
527
// w is node at topological index i.
528
int w = Index2Node[i];
529
if (Visited.test(w)) {
535
Allocate(w, i - shift);
539
for (unsigned j = 0; j < L.size(); ++j) {
540
Allocate(L[j], i - shift);
546
/// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
548
bool ScheduleDAGTopologicalSort::WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
549
if (IsReachable(TargetSU, SU))
551
for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
553
if (I->isAssignedRegDep() &&
554
IsReachable(TargetSU, I->getSUnit()))
559
/// IsReachable - Checks if SU is reachable from TargetSU.
560
bool ScheduleDAGTopologicalSort::IsReachable(const SUnit *SU,
561
const SUnit *TargetSU) {
562
// If insertion of the edge SU->TargetSU would create a cycle
563
// then there is a path from TargetSU to SU.
564
int UpperBound, LowerBound;
565
LowerBound = Node2Index[TargetSU->NodeNum];
566
UpperBound = Node2Index[SU->NodeNum];
567
bool HasLoop = false;
568
// Is Ord(TargetSU) < Ord(SU) ?
569
if (LowerBound < UpperBound) {
571
// There may be a path from TargetSU to SU. Check for it.
572
DFS(TargetSU, UpperBound, HasLoop);
577
/// Allocate - assign the topological index to the node n.
578
void ScheduleDAGTopologicalSort::Allocate(int n, int index) {
579
Node2Index[n] = index;
580
Index2Node[index] = n;
583
ScheduleDAGTopologicalSort::
584
ScheduleDAGTopologicalSort(std::vector<SUnit> &sunits) : SUnits(sunits) {}
586
ScheduleHazardRecognizer::~ScheduleHazardRecognizer() {}