1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
|
//===----- SchedulePostRAList.cpp - list scheduler ------------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This implements a top-down list scheduler, using standard algorithms.
// The basic approach uses a priority queue of available nodes to schedule.
// One at a time, nodes are taken from the priority queue (thus in priority
// order), checked for legality to schedule, and emitted if legal.
//
// Nodes may not be legal to schedule either due to structural hazards (e.g.
// pipeline or resource constraints) or because an input to the instruction has
// not completed execution.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "post-RA-sched"
#include "ScheduleDAGInstrs.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/LatencyPriorityQueue.h"
#include "llvm/CodeGen/SchedulerRegistry.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/ScheduleHazardRecognizer.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/ADT/Statistic.h"
#include <map>
using namespace llvm;
STATISTIC(NumNoops, "Number of noops inserted");
STATISTIC(NumStalls, "Number of pipeline stalls");
static cl::opt<bool>
EnableAntiDepBreaking("break-anti-dependencies",
cl::desc("Break post-RA scheduling anti-dependencies"),
cl::init(true), cl::Hidden);
static cl::opt<bool>
EnablePostRAHazardAvoidance("avoid-hazards",
cl::desc("Enable simple hazard-avoidance"),
cl::init(true), cl::Hidden);
namespace {
class VISIBILITY_HIDDEN PostRAScheduler : public MachineFunctionPass {
public:
static char ID;
PostRAScheduler() : MachineFunctionPass(&ID) {}
void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<MachineDominatorTree>();
AU.addPreserved<MachineDominatorTree>();
AU.addRequired<MachineLoopInfo>();
AU.addPreserved<MachineLoopInfo>();
MachineFunctionPass::getAnalysisUsage(AU);
}
const char *getPassName() const {
return "Post RA top-down list latency scheduler";
}
bool runOnMachineFunction(MachineFunction &Fn);
};
char PostRAScheduler::ID = 0;
class VISIBILITY_HIDDEN SchedulePostRATDList : public ScheduleDAGInstrs {
/// AvailableQueue - The priority queue to use for the available SUnits.
///
LatencyPriorityQueue AvailableQueue;
/// PendingQueue - This contains all of the instructions whose operands have
/// been issued, but their results are not ready yet (due to the latency of
/// the operation). Once the operands becomes available, the instruction is
/// added to the AvailableQueue.
std::vector<SUnit*> PendingQueue;
/// Topo - A topological ordering for SUnits.
ScheduleDAGTopologicalSort Topo;
/// AllocatableSet - The set of allocatable registers.
/// We'll be ignoring anti-dependencies on non-allocatable registers,
/// because they may not be safe to break.
const BitVector AllocatableSet;
/// HazardRec - The hazard recognizer to use.
ScheduleHazardRecognizer *HazardRec;
/// Classes - For live regs that are only used in one register class in a
/// live range, the register class. If the register is not live, the
/// corresponding value is null. If the register is live but used in
/// multiple register classes, the corresponding value is -1 casted to a
/// pointer.
const TargetRegisterClass *
Classes[TargetRegisterInfo::FirstVirtualRegister];
/// RegRegs - Map registers to all their references within a live range.
std::multimap<unsigned, MachineOperand *> RegRefs;
/// The index of the most recent kill (proceding bottom-up), or ~0u if
/// the register is not live.
unsigned KillIndices[TargetRegisterInfo::FirstVirtualRegister];
/// The index of the most recent complete def (proceding bottom up), or ~0u
/// if the register is live.
unsigned DefIndices[TargetRegisterInfo::FirstVirtualRegister];
public:
SchedulePostRATDList(MachineFunction &MF,
const MachineLoopInfo &MLI,
const MachineDominatorTree &MDT,
ScheduleHazardRecognizer *HR)
: ScheduleDAGInstrs(MF, MLI, MDT), Topo(SUnits),
AllocatableSet(TRI->getAllocatableSet(MF)),
HazardRec(HR) {}
~SchedulePostRATDList() {
delete HazardRec;
}
/// StartBlock - Initialize register live-range state for scheduling in
/// this block.
///
void StartBlock(MachineBasicBlock *BB);
/// Schedule - Schedule the instruction range using list scheduling.
///
void Schedule();
/// Observe - Update liveness information to account for the current
/// instruction, which will not be scheduled.
///
void Observe(MachineInstr *MI);
/// FinishBlock - Clean up register live-range state.
///
void FinishBlock();
private:
void PrescanInstruction(MachineInstr *MI);
void ScanInstruction(MachineInstr *MI, unsigned Count);
void ReleaseSucc(SUnit *SU, SDep *SuccEdge);
void ReleaseSuccessors(SUnit *SU);
void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
void ListScheduleTopDown();
bool BreakAntiDependencies();
};
/// SimpleHazardRecognizer - A *very* simple hazard recognizer. It uses
/// a coarse classification and attempts to avoid that instructions of
/// a given class aren't grouped too densely together.
class SimpleHazardRecognizer : public ScheduleHazardRecognizer {
/// Class - A simple classification for SUnits.
enum Class {
Other, Load, Store
};
/// Window - The Class values of the most recently issued
/// instructions.
Class Window[8];
/// getClass - Classify the given SUnit.
Class getClass(const SUnit *SU) {
const MachineInstr *MI = SU->getInstr();
const TargetInstrDesc &TID = MI->getDesc();
if (TID.mayLoad())
return Load;
if (TID.mayStore())
return Store;
return Other;
}
/// Step - Rotate the existing entries in Window and insert the
/// given class value in position as the most recent.
void Step(Class C) {
std::copy(Window+1, array_endof(Window), Window);
Window[array_lengthof(Window)-1] = C;
}
public:
SimpleHazardRecognizer() : Window() {}
virtual HazardType getHazardType(SUnit *SU) {
Class C = getClass(SU);
if (C == Other)
return NoHazard;
unsigned Score = 0;
for (unsigned i = 0; i != array_lengthof(Window); ++i)
if (Window[i] == C)
Score += i + 1;
if (Score > array_lengthof(Window) * 2)
return Hazard;
return NoHazard;
}
virtual void EmitInstruction(SUnit *SU) {
Step(getClass(SU));
}
virtual void AdvanceCycle() {
Step(Other);
}
};
}
/// isSchedulingBoundary - Test if the given instruction should be
/// considered a scheduling boundary. This primarily includes labels
/// and terminators.
///
static bool isSchedulingBoundary(const MachineInstr *MI,
const MachineFunction &MF) {
// Terminators and labels can't be scheduled around.
if (MI->getDesc().isTerminator() || MI->isLabel())
return true;
// Don't attempt to schedule around any instruction that modifies
// a stack-oriented pointer, as it's unlikely to be profitable. This
// saves compile time, because it doesn't require every single
// stack slot reference to depend on the instruction that does the
// modification.
const TargetLowering &TLI = *MF.getTarget().getTargetLowering();
if (MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore()))
return true;
return false;
}
bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
DOUT << "PostRAScheduler\n";
const MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
const MachineDominatorTree &MDT = getAnalysis<MachineDominatorTree>();
ScheduleHazardRecognizer *HR = EnablePostRAHazardAvoidance ?
new SimpleHazardRecognizer :
new ScheduleHazardRecognizer();
SchedulePostRATDList Scheduler(Fn, MLI, MDT, HR);
// Loop over all of the basic blocks
for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
MBB != MBBe; ++MBB) {
// Initialize register live-range state for scheduling in this block.
Scheduler.StartBlock(MBB);
// Schedule each sequence of instructions not interrupted by a label
// or anything else that effectively needs to shut down scheduling.
MachineBasicBlock::iterator Current = MBB->end();
for (MachineBasicBlock::iterator I = Current; I != MBB->begin(); ) {
MachineInstr *MI = prior(I);
if (isSchedulingBoundary(MI, Fn)) {
if (I != Current) {
Scheduler.Run(0, MBB, I, Current);
Scheduler.EmitSchedule();
}
Scheduler.Observe(MI);
Current = MI;
}
I = MI;
}
Scheduler.Run(0, MBB, MBB->begin(), Current);
Scheduler.EmitSchedule();
// Clean up register live-range state.
Scheduler.FinishBlock();
}
return true;
}
/// StartBlock - Initialize register live-range state for scheduling in
/// this block.
///
void SchedulePostRATDList::StartBlock(MachineBasicBlock *BB) {
// Call the superclass.
ScheduleDAGInstrs::StartBlock(BB);
// Clear out the register class data.
std::fill(Classes, array_endof(Classes),
static_cast<const TargetRegisterClass *>(0));
// Initialize the indices to indicate that no registers are live.
std::fill(KillIndices, array_endof(KillIndices), ~0u);
std::fill(DefIndices, array_endof(DefIndices), BB->size());
// Determine the live-out physregs for this block.
if (!BB->empty() && BB->back().getDesc().isReturn())
// In a return block, examine the function live-out regs.
for (MachineRegisterInfo::liveout_iterator I = MRI.liveout_begin(),
E = MRI.liveout_end(); I != E; ++I) {
unsigned Reg = *I;
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[Reg] = BB->size();
DefIndices[Reg] = ~0u;
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[AliasReg] = BB->size();
DefIndices[AliasReg] = ~0u;
}
}
else
// In a non-return block, examine the live-in regs of all successors.
for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
SE = BB->succ_end(); SI != SE; ++SI)
for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
E = (*SI)->livein_end(); I != E; ++I) {
unsigned Reg = *I;
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[Reg] = BB->size();
DefIndices[Reg] = ~0u;
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[AliasReg] = BB->size();
DefIndices[AliasReg] = ~0u;
}
}
// Consider callee-saved registers as live-out, since we're running after
// prologue/epilogue insertion so there's no way to add additional
// saved registers.
//
// TODO: If the callee saves and restores these, then we can potentially
// use them between the save and the restore. To do that, we could scan
// the exit blocks to see which of these registers are defined.
// Alternatively, callee-saved registers that aren't saved and restored
// could be marked live-in in every block.
for (const unsigned *I = TRI->getCalleeSavedRegs(); *I; ++I) {
unsigned Reg = *I;
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[Reg] = BB->size();
DefIndices[Reg] = ~0u;
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
KillIndices[AliasReg] = BB->size();
DefIndices[AliasReg] = ~0u;
}
}
}
/// Schedule - Schedule the instruction range using list scheduling.
///
void SchedulePostRATDList::Schedule() {
DOUT << "********** List Scheduling **********\n";
// Build the scheduling graph.
BuildSchedGraph();
if (EnableAntiDepBreaking) {
if (BreakAntiDependencies()) {
// We made changes. Update the dependency graph.
// Theoretically we could update the graph in place:
// When a live range is changed to use a different register, remove
// the def's anti-dependence *and* output-dependence edges due to
// that register, and add new anti-dependence and output-dependence
// edges based on the next live range of the register.
SUnits.clear();
EntrySU = SUnit();
ExitSU = SUnit();
BuildSchedGraph();
}
}
AvailableQueue.initNodes(SUnits);
ListScheduleTopDown();
AvailableQueue.releaseState();
}
/// Observe - Update liveness information to account for the current
/// instruction, which will not be scheduled.
///
void SchedulePostRATDList::Observe(MachineInstr *MI) {
PrescanInstruction(MI);
ScanInstruction(MI, 0);
}
/// FinishBlock - Clean up register live-range state.
///
void SchedulePostRATDList::FinishBlock() {
RegRefs.clear();
// Call the superclass.
ScheduleDAGInstrs::FinishBlock();
}
/// getInstrOperandRegClass - Return register class of the operand of an
/// instruction of the specified TargetInstrDesc.
static const TargetRegisterClass*
getInstrOperandRegClass(const TargetRegisterInfo *TRI,
const TargetInstrDesc &II, unsigned Op) {
if (Op >= II.getNumOperands())
return NULL;
if (II.OpInfo[Op].isLookupPtrRegClass())
return TRI->getPointerRegClass();
return TRI->getRegClass(II.OpInfo[Op].RegClass);
}
/// CriticalPathStep - Return the next SUnit after SU on the bottom-up
/// critical path.
static SDep *CriticalPathStep(SUnit *SU) {
SDep *Next = 0;
unsigned NextDepth = 0;
// Find the predecessor edge with the greatest depth.
for (SUnit::pred_iterator P = SU->Preds.begin(), PE = SU->Preds.end();
P != PE; ++P) {
SUnit *PredSU = P->getSUnit();
unsigned PredLatency = P->getLatency();
unsigned PredTotalLatency = PredSU->getDepth() + PredLatency;
// In the case of a latency tie, prefer an anti-dependency edge over
// other types of edges.
if (NextDepth < PredTotalLatency ||
(NextDepth == PredTotalLatency && P->getKind() == SDep::Anti)) {
NextDepth = PredTotalLatency;
Next = &*P;
}
}
return Next;
}
void SchedulePostRATDList::PrescanInstruction(MachineInstr *MI) {
// Scan the register operands for this instruction and update
// Classes and RegRefs.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
const TargetRegisterClass *NewRC =
getInstrOperandRegClass(TRI, MI->getDesc(), i);
// For now, only allow the register to be changed if its register
// class is consistent across all uses.
if (!Classes[Reg] && NewRC)
Classes[Reg] = NewRC;
else if (!NewRC || Classes[Reg] != NewRC)
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
// Now check for aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
// If an alias of the reg is used during the live range, give up.
// Note that this allows us to skip checking if AntiDepReg
// overlaps with any of the aliases, among other things.
unsigned AliasReg = *Alias;
if (Classes[AliasReg]) {
Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
}
}
// If we're still willing to consider this register, note the reference.
if (Classes[Reg] != reinterpret_cast<TargetRegisterClass *>(-1))
RegRefs.insert(std::make_pair(Reg, &MO));
}
}
void SchedulePostRATDList::ScanInstruction(MachineInstr *MI,
unsigned Count) {
// Update liveness.
// Proceding upwards, registers that are defed but not used in this
// instruction are now dead.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
if (!MO.isDef()) continue;
// Ignore two-addr defs.
if (MI->isRegReDefinedByTwoAddr(i)) continue;
DefIndices[Reg] = Count;
KillIndices[Reg] = ~0u;
Classes[Reg] = 0;
RegRefs.erase(Reg);
// Repeat, for all subregs.
for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
*Subreg; ++Subreg) {
unsigned SubregReg = *Subreg;
DefIndices[SubregReg] = Count;
KillIndices[SubregReg] = ~0u;
Classes[SubregReg] = 0;
RegRefs.erase(SubregReg);
}
// Conservatively mark super-registers as unusable.
for (const unsigned *Super = TRI->getSuperRegisters(Reg);
*Super; ++Super) {
unsigned SuperReg = *Super;
Classes[SuperReg] = reinterpret_cast<TargetRegisterClass *>(-1);
}
}
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
if (!MO.isUse()) continue;
const TargetRegisterClass *NewRC =
getInstrOperandRegClass(TRI, MI->getDesc(), i);
// For now, only allow the register to be changed if its register
// class is consistent across all uses.
if (!Classes[Reg] && NewRC)
Classes[Reg] = NewRC;
else if (!NewRC || Classes[Reg] != NewRC)
Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
RegRefs.insert(std::make_pair(Reg, &MO));
// It wasn't previously live but now it is, this is a kill.
if (KillIndices[Reg] == ~0u) {
KillIndices[Reg] = Count;
DefIndices[Reg] = ~0u;
}
// Repeat, for all aliases.
for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
unsigned AliasReg = *Alias;
if (KillIndices[AliasReg] == ~0u) {
KillIndices[AliasReg] = Count;
DefIndices[AliasReg] = ~0u;
}
}
}
}
/// BreakAntiDependencies - Identifiy anti-dependencies along the critical path
/// of the ScheduleDAG and break them by renaming registers.
///
bool SchedulePostRATDList::BreakAntiDependencies() {
// The code below assumes that there is at least one instruction,
// so just duck out immediately if the block is empty.
if (SUnits.empty()) return false;
// Find the node at the bottom of the critical path.
SUnit *Max = 0;
for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
SUnit *SU = &SUnits[i];
if (!Max || SU->getDepth() + SU->Latency > Max->getDepth() + Max->Latency)
Max = SU;
}
DOUT << "Critical path has total latency "
<< (Max->getDepth() + Max->Latency) << "\n";
// Track progress along the critical path through the SUnit graph as we walk
// the instructions.
SUnit *CriticalPathSU = Max;
MachineInstr *CriticalPathMI = CriticalPathSU->getInstr();
// Consider this pattern:
// A = ...
// ... = A
// A = ...
// ... = A
// A = ...
// ... = A
// A = ...
// ... = A
// There are three anti-dependencies here, and without special care,
// we'd break all of them using the same register:
// A = ...
// ... = A
// B = ...
// ... = B
// B = ...
// ... = B
// B = ...
// ... = B
// because at each anti-dependence, B is the first register that
// isn't A which is free. This re-introduces anti-dependencies
// at all but one of the original anti-dependencies that we were
// trying to break. To avoid this, keep track of the most recent
// register that each register was replaced with, avoid avoid
// using it to repair an anti-dependence on the same register.
// This lets us produce this:
// A = ...
// ... = A
// B = ...
// ... = B
// C = ...
// ... = C
// B = ...
// ... = B
// This still has an anti-dependence on B, but at least it isn't on the
// original critical path.
//
// TODO: If we tracked more than one register here, we could potentially
// fix that remaining critical edge too. This is a little more involved,
// because unlike the most recent register, less recent registers should
// still be considered, though only if no other registers are available.
unsigned LastNewReg[TargetRegisterInfo::FirstVirtualRegister] = {};
// Attempt to break anti-dependence edges on the critical path. Walk the
// instructions from the bottom up, tracking information about liveness
// as we go to help determine which registers are available.
bool Changed = false;
unsigned Count = SUnits.size() - 1;
for (MachineBasicBlock::iterator I = End, E = Begin;
I != E; --Count) {
MachineInstr *MI = --I;
// After regalloc, IMPLICIT_DEF instructions aren't safe to treat as
// dependence-breaking. In the case of an INSERT_SUBREG, the IMPLICIT_DEF
// is left behind appearing to clobber the super-register, while the
// subregister needs to remain live. So we just ignore them.
if (MI->getOpcode() == TargetInstrInfo::IMPLICIT_DEF)
continue;
// Check if this instruction has a dependence on the critical path that
// is an anti-dependence that we may be able to break. If it is, set
// AntiDepReg to the non-zero register associated with the anti-dependence.
//
// We limit our attention to the critical path as a heuristic to avoid
// breaking anti-dependence edges that aren't going to significantly
// impact the overall schedule. There are a limited number of registers
// and we want to save them for the important edges.
//
// TODO: Instructions with multiple defs could have multiple
// anti-dependencies. The current code here only knows how to break one
// edge per instruction. Note that we'd have to be able to break all of
// the anti-dependencies in an instruction in order to be effective.
unsigned AntiDepReg = 0;
if (MI == CriticalPathMI) {
if (SDep *Edge = CriticalPathStep(CriticalPathSU)) {
SUnit *NextSU = Edge->getSUnit();
// Only consider anti-dependence edges.
if (Edge->getKind() == SDep::Anti) {
AntiDepReg = Edge->getReg();
assert(AntiDepReg != 0 && "Anti-dependence on reg0?");
// Don't break anti-dependencies on non-allocatable registers.
if (!AllocatableSet.test(AntiDepReg))
AntiDepReg = 0;
else {
// If the SUnit has other dependencies on the SUnit that it
// anti-depends on, don't bother breaking the anti-dependency
// since those edges would prevent such units from being
// scheduled past each other regardless.
//
// Also, if there are dependencies on other SUnits with the
// same register as the anti-dependency, don't attempt to
// break it.
for (SUnit::pred_iterator P = CriticalPathSU->Preds.begin(),
PE = CriticalPathSU->Preds.end(); P != PE; ++P)
if (P->getSUnit() == NextSU ?
(P->getKind() != SDep::Anti || P->getReg() != AntiDepReg) :
(P->getKind() == SDep::Data && P->getReg() == AntiDepReg)) {
AntiDepReg = 0;
break;
}
}
}
CriticalPathSU = NextSU;
CriticalPathMI = CriticalPathSU->getInstr();
} else {
// We've reached the end of the critical path.
CriticalPathSU = 0;
CriticalPathMI = 0;
}
}
PrescanInstruction(MI);
// If this instruction has a use of AntiDepReg, breaking it
// is invalid.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
if (!MO.isReg()) continue;
unsigned Reg = MO.getReg();
if (Reg == 0) continue;
if (MO.isUse() && AntiDepReg == Reg) {
AntiDepReg = 0;
break;
}
}
// Determine AntiDepReg's register class, if it is live and is
// consistently used within a single class.
const TargetRegisterClass *RC = AntiDepReg != 0 ? Classes[AntiDepReg] : 0;
assert((AntiDepReg == 0 || RC != NULL) &&
"Register should be live if it's causing an anti-dependence!");
if (RC == reinterpret_cast<TargetRegisterClass *>(-1))
AntiDepReg = 0;
// Look for a suitable register to use to break the anti-depenence.
//
// TODO: Instead of picking the first free register, consider which might
// be the best.
if (AntiDepReg != 0) {
for (TargetRegisterClass::iterator R = RC->allocation_order_begin(MF),
RE = RC->allocation_order_end(MF); R != RE; ++R) {
unsigned NewReg = *R;
// Don't replace a register with itself.
if (NewReg == AntiDepReg) continue;
// Don't replace a register with one that was recently used to repair
// an anti-dependence with this AntiDepReg, because that would
// re-introduce that anti-dependence.
if (NewReg == LastNewReg[AntiDepReg]) continue;
// If NewReg is dead and NewReg's most recent def is not before
// AntiDepReg's kill, it's safe to replace AntiDepReg with NewReg.
assert(((KillIndices[AntiDepReg] == ~0u) != (DefIndices[AntiDepReg] == ~0u)) &&
"Kill and Def maps aren't consistent for AntiDepReg!");
assert(((KillIndices[NewReg] == ~0u) != (DefIndices[NewReg] == ~0u)) &&
"Kill and Def maps aren't consistent for NewReg!");
if (KillIndices[NewReg] == ~0u &&
Classes[NewReg] != reinterpret_cast<TargetRegisterClass *>(-1) &&
KillIndices[AntiDepReg] <= DefIndices[NewReg]) {
DOUT << "Breaking anti-dependence edge on "
<< TRI->getName(AntiDepReg)
<< " with " << RegRefs.count(AntiDepReg) << " references"
<< " using " << TRI->getName(NewReg) << "!\n";
// Update the references to the old register to refer to the new
// register.
std::pair<std::multimap<unsigned, MachineOperand *>::iterator,
std::multimap<unsigned, MachineOperand *>::iterator>
Range = RegRefs.equal_range(AntiDepReg);
for (std::multimap<unsigned, MachineOperand *>::iterator
Q = Range.first, QE = Range.second; Q != QE; ++Q)
Q->second->setReg(NewReg);
// We just went back in time and modified history; the
// liveness information for the anti-depenence reg is now
// inconsistent. Set the state as if it were dead.
Classes[NewReg] = Classes[AntiDepReg];
DefIndices[NewReg] = DefIndices[AntiDepReg];
KillIndices[NewReg] = KillIndices[AntiDepReg];
Classes[AntiDepReg] = 0;
DefIndices[AntiDepReg] = KillIndices[AntiDepReg];
KillIndices[AntiDepReg] = ~0u;
RegRefs.erase(AntiDepReg);
Changed = true;
LastNewReg[AntiDepReg] = NewReg;
break;
}
}
}
ScanInstruction(MI, Count);
}
assert(Count == ~0u && "Count mismatch!");
return Changed;
}
//===----------------------------------------------------------------------===//
// Top-Down Scheduling
//===----------------------------------------------------------------------===//
/// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
/// the PendingQueue if the count reaches zero. Also update its cycle bound.
void SchedulePostRATDList::ReleaseSucc(SUnit *SU, SDep *SuccEdge) {
SUnit *SuccSU = SuccEdge->getSUnit();
--SuccSU->NumPredsLeft;
#ifndef NDEBUG
if (SuccSU->NumPredsLeft < 0) {
cerr << "*** Scheduling failed! ***\n";
SuccSU->dump(this);
cerr << " has been released too many times!\n";
assert(0);
}
#endif
// Compute how many cycles it will be before this actually becomes
// available. This is the max of the start time of all predecessors plus
// their latencies.
SuccSU->setDepthToAtLeast(SU->getDepth() + SuccEdge->getLatency());
// If all the node's predecessors are scheduled, this node is ready
// to be scheduled. Ignore the special ExitSU node.
if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
PendingQueue.push_back(SuccSU);
}
/// ReleaseSuccessors - Call ReleaseSucc on each of SU's successors.
void SchedulePostRATDList::ReleaseSuccessors(SUnit *SU) {
for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
I != E; ++I)
ReleaseSucc(SU, &*I);
}
/// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
/// count of its successors. If a successor pending count is zero, add it to
/// the Available queue.
void SchedulePostRATDList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
DOUT << "*** Scheduling [" << CurCycle << "]: ";
DEBUG(SU->dump(this));
Sequence.push_back(SU);
assert(CurCycle >= SU->getDepth() && "Node scheduled above its depth!");
SU->setDepthToAtLeast(CurCycle);
ReleaseSuccessors(SU);
SU->isScheduled = true;
AvailableQueue.ScheduledNode(SU);
}
/// ListScheduleTopDown - The main loop of list scheduling for top-down
/// schedulers.
void SchedulePostRATDList::ListScheduleTopDown() {
unsigned CurCycle = 0;
// Release any successors of the special Entry node.
ReleaseSuccessors(&EntrySU);
// All leaves to Available queue.
for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
// It is available if it has no predecessors.
if (SUnits[i].Preds.empty()) {
AvailableQueue.push(&SUnits[i]);
SUnits[i].isAvailable = true;
}
}
// While Available queue is not empty, grab the node with the highest
// priority. If it is not ready put it back. Schedule the node.
std::vector<SUnit*> NotReady;
Sequence.reserve(SUnits.size());
while (!AvailableQueue.empty() || !PendingQueue.empty()) {
// Check to see if any of the pending instructions are ready to issue. If
// so, add them to the available queue.
unsigned MinDepth = ~0u;
for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
if (PendingQueue[i]->getDepth() <= CurCycle) {
AvailableQueue.push(PendingQueue[i]);
PendingQueue[i]->isAvailable = true;
PendingQueue[i] = PendingQueue.back();
PendingQueue.pop_back();
--i; --e;
} else if (PendingQueue[i]->getDepth() < MinDepth)
MinDepth = PendingQueue[i]->getDepth();
}
// If there are no instructions available, don't try to issue anything, and
// don't advance the hazard recognizer.
if (AvailableQueue.empty()) {
CurCycle = MinDepth != ~0u ? MinDepth : CurCycle + 1;
continue;
}
SUnit *FoundSUnit = 0;
bool HasNoopHazards = false;
while (!AvailableQueue.empty()) {
SUnit *CurSUnit = AvailableQueue.pop();
ScheduleHazardRecognizer::HazardType HT =
HazardRec->getHazardType(CurSUnit);
if (HT == ScheduleHazardRecognizer::NoHazard) {
FoundSUnit = CurSUnit;
break;
}
// Remember if this is a noop hazard.
HasNoopHazards |= HT == ScheduleHazardRecognizer::NoopHazard;
NotReady.push_back(CurSUnit);
}
// Add the nodes that aren't ready back onto the available list.
if (!NotReady.empty()) {
AvailableQueue.push_all(NotReady);
NotReady.clear();
}
// If we found a node to schedule, do it now.
if (FoundSUnit) {
ScheduleNodeTopDown(FoundSUnit, CurCycle);
HazardRec->EmitInstruction(FoundSUnit);
// If this is a pseudo-op node, we don't want to increment the current
// cycle.
if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
++CurCycle;
} else if (!HasNoopHazards) {
// Otherwise, we have a pipeline stall, but no other problem, just advance
// the current cycle and try again.
DOUT << "*** Advancing cycle, no work to do\n";
HazardRec->AdvanceCycle();
++NumStalls;
++CurCycle;
} else {
// Otherwise, we have no instructions to issue and we have instructions
// that will fault if we don't do this right. This is the case for
// processors without pipeline interlocks and other cases.
DOUT << "*** Emitting noop\n";
HazardRec->EmitNoop();
Sequence.push_back(0); // NULL here means noop
++NumNoops;
++CurCycle;
}
}
#ifndef NDEBUG
VerifySchedule(/*isBottomUp=*/false);
#endif
}
//===----------------------------------------------------------------------===//
// Public Constructor Functions
//===----------------------------------------------------------------------===//
FunctionPass *llvm::createPostRAScheduler() {
return new PostRAScheduler();
}
|