forked from llvm/llvm-project
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathShrinkWrapping.cpp
2076 lines (1902 loc) · 71.1 KB
/
ShrinkWrapping.cpp
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
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
//===- bolt/Passes/ShrinkWrapping.cpp -------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements the ShrinkWrapping class.
//
//===----------------------------------------------------------------------===//
#include "bolt/Passes/ShrinkWrapping.h"
#include "bolt/Core/MCPlus.h"
#include "bolt/Passes/DataflowInfoManager.h"
#include <numeric>
#include <stack>
#define DEBUG_TYPE "shrinkwrapping"
using namespace llvm;
namespace opts {
extern cl::opt<bool> TimeOpts;
extern cl::OptionCategory BoltOptCategory;
static cl::opt<unsigned> ShrinkWrappingThreshold(
"shrink-wrapping-threshold",
cl::desc("Percentage of prologue execution count to use as threshold when"
" evaluating whether a block is cold enough to be profitable to"
" move eligible spills there"),
cl::init(30), cl::ZeroOrMore, cl::cat(BoltOptCategory));
} // namespace opts
namespace llvm {
namespace bolt {
void CalleeSavedAnalysis::analyzeSaves() {
ReachingDefOrUse</*Def=*/true> &RD = Info.getReachingDefs();
StackReachingUses &SRU = Info.getStackReachingUses();
auto &InsnToBB = Info.getInsnToBBMap();
BitVector BlacklistedRegs(BC.MRI->getNumRegs(), false);
LLVM_DEBUG(dbgs() << "Checking spill locations\n");
for (BinaryBasicBlock &BB : BF) {
LLVM_DEBUG(dbgs() << "\tNow at BB " << BB.getName() << "\n");
const MCInst *Prev = nullptr;
for (MCInst &Inst : BB) {
if (ErrorOr<const FrameIndexEntry &> FIE = FA.getFIEFor(Inst)) {
// Blacklist weird stores we don't understand
if ((!FIE->IsSimple || FIE->StackOffset >= 0) && FIE->IsStore &&
FIE->IsStoreFromReg) {
BlacklistedRegs.set(FIE->RegOrImm);
CalleeSaved.reset(FIE->RegOrImm);
Prev = &Inst;
continue;
}
if (!FIE->IsStore || !FIE->IsStoreFromReg ||
BlacklistedRegs[FIE->RegOrImm]) {
Prev = &Inst;
continue;
}
// If this reg is defined locally, it is not a callee-saved reg
if (RD.isReachedBy(FIE->RegOrImm,
Prev ? RD.expr_begin(*Prev) : RD.expr_begin(BB))) {
BlacklistedRegs.set(FIE->RegOrImm);
CalleeSaved.reset(FIE->RegOrImm);
Prev = &Inst;
continue;
}
// If this stack position is accessed in another function, we are
// probably dealing with a parameter passed in a stack -- do not mess
// with it
if (SRU.isStoreUsed(*FIE,
Prev ? SRU.expr_begin(*Prev) : SRU.expr_begin(BB)),
/*IncludeLocalAccesses=*/false) {
BlacklistedRegs.set(FIE->RegOrImm);
CalleeSaved.reset(FIE->RegOrImm);
Prev = &Inst;
continue;
}
// If this stack position is loaded elsewhere in another reg, we can't
// update it, so blacklist it.
if (SRU.isLoadedInDifferentReg(*FIE, Prev ? SRU.expr_begin(*Prev)
: SRU.expr_begin(BB))) {
BlacklistedRegs.set(FIE->RegOrImm);
CalleeSaved.reset(FIE->RegOrImm);
Prev = &Inst;
continue;
}
// Ignore regs with multiple saves
if (CalleeSaved[FIE->RegOrImm]) {
BlacklistedRegs.set(FIE->RegOrImm);
CalleeSaved.reset(FIE->RegOrImm);
Prev = &Inst;
continue;
}
CalleeSaved.set(FIE->RegOrImm);
SaveFIEByReg[FIE->RegOrImm] = &*FIE;
SavingCost[FIE->RegOrImm] += InsnToBB[&Inst]->getKnownExecutionCount();
BC.MIB->addAnnotation(Inst, getSaveTag(), FIE->RegOrImm, AllocatorId);
OffsetsByReg[FIE->RegOrImm] = FIE->StackOffset;
LLVM_DEBUG(dbgs() << "Logging new candidate for Callee-Saved Reg: "
<< FIE->RegOrImm << "\n");
}
Prev = &Inst;
}
}
}
void CalleeSavedAnalysis::analyzeRestores() {
ReachingDefOrUse</*Def=*/false> &RU = Info.getReachingUses();
// Now compute all restores of these callee-saved regs
for (BinaryBasicBlock &BB : BF) {
const MCInst *Prev = nullptr;
for (auto I = BB.rbegin(), E = BB.rend(); I != E; ++I) {
MCInst &Inst = *I;
if (ErrorOr<const FrameIndexEntry &> FIE = FA.getFIEFor(Inst)) {
if (!FIE->IsLoad || !CalleeSaved[FIE->RegOrImm]) {
Prev = &Inst;
continue;
}
// If this reg is used locally after a restore, then we are probably
// not dealing with a callee-saved reg. Except if this use is by
// another store, but we don't cover this case yet.
// Also not callee-saved if this load accesses caller stack or isn't
// simple.
if (!FIE->IsSimple || FIE->StackOffset >= 0 ||
RU.isReachedBy(FIE->RegOrImm,
Prev ? RU.expr_begin(*Prev) : RU.expr_begin(BB))) {
CalleeSaved.reset(FIE->RegOrImm);
Prev = &Inst;
continue;
}
// If stack offsets between saves/store don't agree with each other,
// we don't completely understand what's happening here
if (FIE->StackOffset != OffsetsByReg[FIE->RegOrImm]) {
CalleeSaved.reset(FIE->RegOrImm);
LLVM_DEBUG(dbgs() << "Dismissing Callee-Saved Reg because we found a "
"mismatching restore: "
<< FIE->RegOrImm << "\n");
Prev = &Inst;
continue;
}
LLVM_DEBUG(dbgs() << "Adding matching restore for: " << FIE->RegOrImm
<< "\n");
if (LoadFIEByReg[FIE->RegOrImm] == nullptr)
LoadFIEByReg[FIE->RegOrImm] = &*FIE;
BC.MIB->addAnnotation(Inst, getRestoreTag(), FIE->RegOrImm,
AllocatorId);
HasRestores.set(FIE->RegOrImm);
}
Prev = &Inst;
}
}
}
std::vector<MCInst *> CalleeSavedAnalysis::getSavesByReg(uint16_t Reg) {
std::vector<MCInst *> Results;
for (BinaryBasicBlock &BB : BF)
for (MCInst &Inst : BB)
if (getSavedReg(Inst) == Reg)
Results.push_back(&Inst);
return Results;
}
std::vector<MCInst *> CalleeSavedAnalysis::getRestoresByReg(uint16_t Reg) {
std::vector<MCInst *> Results;
for (BinaryBasicBlock &BB : BF)
for (MCInst &Inst : BB)
if (getRestoredReg(Inst) == Reg)
Results.push_back(&Inst);
return Results;
}
CalleeSavedAnalysis::~CalleeSavedAnalysis() {
for (BinaryBasicBlock &BB : BF) {
for (MCInst &Inst : BB) {
BC.MIB->removeAnnotation(Inst, getSaveTag());
BC.MIB->removeAnnotation(Inst, getRestoreTag());
}
}
}
void StackLayoutModifier::blacklistRegion(int64_t Offset, int64_t Size) {
if (BlacklistedRegions[Offset] < Size)
BlacklistedRegions[Offset] = Size;
}
bool StackLayoutModifier::isRegionBlacklisted(int64_t Offset, int64_t Size) {
for (std::pair<const int64_t, int64_t> Elem : BlacklistedRegions)
if (Offset + Size > Elem.first && Offset < Elem.first + Elem.second)
return true;
return false;
}
bool StackLayoutModifier::blacklistAllInConflictWith(int64_t Offset,
int64_t Size) {
bool HasConflict = false;
for (auto Iter = AvailableRegions.begin(); Iter != AvailableRegions.end();) {
std::pair<const int64_t, int64_t> &Elem = *Iter;
if (Offset + Size > Elem.first && Offset < Elem.first + Elem.second &&
(Offset != Elem.first || Size != Elem.second)) {
Iter = AvailableRegions.erase(Iter);
HasConflict = true;
continue;
}
++Iter;
}
if (HasConflict) {
blacklistRegion(Offset, Size);
return true;
}
return false;
}
void StackLayoutModifier::checkFramePointerInitialization(MCInst &Point) {
StackPointerTracking &SPT = Info.getStackPointerTracking();
if (!BC.MII->get(Point.getOpcode())
.hasDefOfPhysReg(Point, BC.MIB->getFramePointer(), *BC.MRI))
return;
int SPVal, FPVal;
std::tie(SPVal, FPVal) = *SPT.getStateBefore(Point);
std::pair<MCPhysReg, int64_t> FP;
if (FPVal != SPT.EMPTY && FPVal != SPT.SUPERPOSITION)
FP = std::make_pair(BC.MIB->getFramePointer(), FPVal);
else
FP = std::make_pair(0, 0);
std::pair<MCPhysReg, int64_t> SP;
if (SPVal != SPT.EMPTY && SPVal != SPT.SUPERPOSITION)
SP = std::make_pair(BC.MIB->getStackPointer(), SPVal);
else
SP = std::make_pair(0, 0);
int64_t Output;
if (!BC.MIB->evaluateSimple(Point, Output, SP, FP))
return;
// Not your regular frame pointer initialization... bail
if (Output != SPVal)
blacklistRegion(0, 0);
}
void StackLayoutModifier::checkStackPointerRestore(MCInst &Point) {
StackPointerTracking &SPT = Info.getStackPointerTracking();
if (!BC.MII->get(Point.getOpcode())
.hasDefOfPhysReg(Point, BC.MIB->getStackPointer(), *BC.MRI))
return;
// Check if the definition of SP comes from FP -- in this case, this
// value may need to be updated depending on our stack layout changes
const MCInstrDesc &InstInfo = BC.MII->get(Point.getOpcode());
unsigned NumDefs = InstInfo.getNumDefs();
bool UsesFP = false;
for (unsigned I = NumDefs, E = MCPlus::getNumPrimeOperands(Point); I < E;
++I) {
MCOperand &Operand = Point.getOperand(I);
if (!Operand.isReg())
continue;
if (Operand.getReg() == BC.MIB->getFramePointer()) {
UsesFP = true;
break;
}
}
if (!UsesFP)
return;
// Setting up evaluation
int SPVal, FPVal;
std::tie(SPVal, FPVal) = *SPT.getStateBefore(Point);
std::pair<MCPhysReg, int64_t> FP;
if (FPVal != SPT.EMPTY && FPVal != SPT.SUPERPOSITION)
FP = std::make_pair(BC.MIB->getFramePointer(), FPVal);
else
FP = std::make_pair(0, 0);
std::pair<MCPhysReg, int64_t> SP;
if (SPVal != SPT.EMPTY && SPVal != SPT.SUPERPOSITION)
SP = std::make_pair(BC.MIB->getStackPointer(), SPVal);
else
SP = std::make_pair(0, 0);
int64_t Output;
if (!BC.MIB->evaluateSimple(Point, Output, SP, FP))
return;
// If the value is the same of FP, no need to adjust it
if (Output == FPVal)
return;
// If an allocation happened through FP, bail
if (Output <= SPVal) {
blacklistRegion(0, 0);
return;
}
// We are restoring SP to an old value based on FP. Mark it as a stack
// access to be fixed later.
BC.MIB->addAnnotation(Point, getSlotTag(), Output, AllocatorId);
}
void StackLayoutModifier::classifyStackAccesses() {
// Understand when stack slots are being used non-locally
StackReachingUses &SRU = Info.getStackReachingUses();
for (BinaryBasicBlock &BB : BF) {
const MCInst *Prev = nullptr;
for (auto I = BB.rbegin(), E = BB.rend(); I != E; ++I) {
MCInst &Inst = *I;
checkFramePointerInitialization(Inst);
checkStackPointerRestore(Inst);
ErrorOr<const FrameIndexEntry &> FIEX = FA.getFIEFor(Inst);
if (!FIEX) {
Prev = &Inst;
continue;
}
if (!FIEX->IsSimple || (FIEX->IsStore && !FIEX->IsStoreFromReg)) {
blacklistRegion(FIEX->StackOffset, FIEX->Size);
Prev = &Inst;
continue;
}
// If this stack position is accessed in another function, we are
// probably dealing with a parameter passed in a stack -- do not mess
// with it
if (SRU.isStoreUsed(*FIEX,
Prev ? SRU.expr_begin(*Prev) : SRU.expr_begin(BB),
/*IncludeLocalAccesses=*/false)) {
blacklistRegion(FIEX->StackOffset, FIEX->Size);
Prev = &Inst;
continue;
}
// Now we have a clear stack slot access. Check if its blacklisted or if
// it conflicts with another chunk.
if (isRegionBlacklisted(FIEX->StackOffset, FIEX->Size) ||
blacklistAllInConflictWith(FIEX->StackOffset, FIEX->Size)) {
Prev = &Inst;
continue;
}
// We are free to go. Add it as available stack slot which we know how
// to move it.
AvailableRegions[FIEX->StackOffset] = FIEX->Size;
BC.MIB->addAnnotation(Inst, getSlotTag(), FIEX->StackOffset, AllocatorId);
RegionToRegMap[FIEX->StackOffset].insert(FIEX->RegOrImm);
RegToRegionMap[FIEX->RegOrImm].insert(FIEX->StackOffset);
LLVM_DEBUG(dbgs() << "Adding region " << FIEX->StackOffset << " size "
<< (int)FIEX->Size << "\n");
}
}
}
void StackLayoutModifier::classifyCFIs() {
std::stack<std::pair<int64_t, uint16_t>> CFIStack;
int64_t CfaOffset = -8;
uint16_t CfaReg = 7;
auto recordAccess = [&](MCInst *Inst, int64_t Offset) {
const uint16_t Reg = *BC.MRI->getLLVMRegNum(CfaReg, /*isEH=*/false);
if (Reg == BC.MIB->getStackPointer() || Reg == BC.MIB->getFramePointer()) {
BC.MIB->addAnnotation(*Inst, getSlotTag(), Offset, AllocatorId);
LLVM_DEBUG(dbgs() << "Recording CFI " << Offset << "\n");
} else {
IsSimple = false;
return;
}
};
for (BinaryBasicBlock *&BB : BF.layout()) {
for (MCInst &Inst : *BB) {
if (!BC.MIB->isCFI(Inst))
continue;
const MCCFIInstruction *CFI = BF.getCFIFor(Inst);
switch (CFI->getOperation()) {
case MCCFIInstruction::OpDefCfa:
CfaOffset = -CFI->getOffset();
recordAccess(&Inst, CfaOffset);
LLVM_FALLTHROUGH;
case MCCFIInstruction::OpDefCfaRegister:
CfaReg = CFI->getRegister();
break;
case MCCFIInstruction::OpDefCfaOffset:
CfaOffset = -CFI->getOffset();
recordAccess(&Inst, CfaOffset);
break;
case MCCFIInstruction::OpOffset:
recordAccess(&Inst, CFI->getOffset());
BC.MIB->addAnnotation(Inst, getOffsetCFIRegTag(),
BC.MRI->getLLVMRegNum(CFI->getRegister(),
/*isEH=*/false),
AllocatorId);
break;
case MCCFIInstruction::OpSameValue:
BC.MIB->addAnnotation(Inst, getOffsetCFIRegTag(),
BC.MRI->getLLVMRegNum(CFI->getRegister(),
/*isEH=*/false),
AllocatorId);
break;
case MCCFIInstruction::OpRememberState:
CFIStack.push(std::make_pair(CfaOffset, CfaReg));
break;
case MCCFIInstruction::OpRestoreState: {
assert(!CFIStack.empty() && "Corrupt CFI stack");
std::pair<int64_t, uint16_t> &Elem = CFIStack.top();
CFIStack.pop();
CfaOffset = Elem.first;
CfaReg = Elem.second;
break;
}
case MCCFIInstruction::OpRelOffset:
case MCCFIInstruction::OpAdjustCfaOffset:
llvm_unreachable("Unhandled AdjustCfaOffset");
break;
default:
break;
}
}
}
}
void StackLayoutModifier::scheduleChange(
MCInst &Inst, StackLayoutModifier::WorklistItem Item) {
auto &WList = BC.MIB->getOrCreateAnnotationAs<std::vector<WorklistItem>>(
Inst, getTodoTag(), AllocatorId);
WList.push_back(Item);
}
bool StackLayoutModifier::canCollapseRegion(MCInst *DeletedPush) {
if (!IsSimple || !BC.MIB->isPush(*DeletedPush))
return false;
ErrorOr<const FrameIndexEntry &> FIE = FA.getFIEFor(*DeletedPush);
if (!FIE)
return false;
return canCollapseRegion(FIE->StackOffset);
}
bool StackLayoutModifier::canCollapseRegion(int64_t RegionAddr) {
if (!IsInitialized)
initialize();
if (!IsSimple)
return false;
if (CollapsedRegions.count(RegionAddr))
return true;
// Check if it is possible to readjust all accesses below RegionAddr
if (!BlacklistedRegions.empty())
return false;
return true;
}
bool StackLayoutModifier::collapseRegion(MCInst *DeletedPush) {
ErrorOr<const FrameIndexEntry &> FIE = FA.getFIEFor(*DeletedPush);
if (!FIE)
return false;
int64_t RegionAddr = FIE->StackOffset;
int64_t RegionSz = FIE->Size;
return collapseRegion(DeletedPush, RegionAddr, RegionSz);
}
bool StackLayoutModifier::collapseRegion(MCInst *Alloc, int64_t RegionAddr,
int64_t RegionSz) {
if (!canCollapseRegion(RegionAddr))
return false;
assert(IsInitialized);
StackAllocationAnalysis &SAA = Info.getStackAllocationAnalysis();
for (BinaryBasicBlock &BB : BF) {
for (MCInst &Inst : BB) {
if (!BC.MIB->hasAnnotation(Inst, getSlotTag()))
continue;
auto Slot =
BC.MIB->getAnnotationAs<decltype(FrameIndexEntry::StackOffset)>(
Inst, getSlotTag());
if (!AvailableRegions.count(Slot))
continue;
// We need to ensure this access is affected by the deleted push
if (!(*SAA.getStateBefore(Inst))[SAA.ExprToIdx[Alloc]])
continue;
if (BC.MIB->isCFI(Inst)) {
if (Slot > RegionAddr)
continue;
scheduleChange(Inst, WorklistItem(WorklistItem::AdjustCFI, RegionSz));
continue;
}
ErrorOr<const FrameIndexEntry &> FIE = FA.getFIEFor(Inst);
if (!FIE) {
if (Slot > RegionAddr)
continue;
// SP update based on frame pointer
scheduleChange(
Inst, WorklistItem(WorklistItem::AdjustLoadStoreOffset, RegionSz));
continue;
}
if (Slot == RegionAddr) {
BC.MIB->addAnnotation(Inst, "AccessesDeletedPos", 0U, AllocatorId);
continue;
}
if (BC.MIB->isPush(Inst) || BC.MIB->isPop(Inst))
continue;
if (FIE->StackPtrReg == BC.MIB->getStackPointer() && Slot < RegionAddr)
continue;
if (FIE->StackPtrReg == BC.MIB->getFramePointer() && Slot > RegionAddr)
continue;
scheduleChange(
Inst, WorklistItem(WorklistItem::AdjustLoadStoreOffset, RegionSz));
}
}
CollapsedRegions.insert(RegionAddr);
return true;
}
void StackLayoutModifier::setOffsetForCollapsedAccesses(int64_t NewOffset) {
for (BinaryBasicBlock &BB : BF) {
for (MCInst &Inst : BB) {
if (!BC.MIB->hasAnnotation(Inst, "AccessesDeletedPos"))
continue;
BC.MIB->removeAnnotation(Inst, "AccessesDeletedPos");
scheduleChange(
Inst, WorklistItem(WorklistItem::AdjustLoadStoreOffset, NewOffset));
}
}
}
bool StackLayoutModifier::canInsertRegion(ProgramPoint P) {
if (!IsInitialized)
initialize();
if (!IsSimple)
return false;
StackPointerTracking &SPT = Info.getStackPointerTracking();
int64_t RegionAddr = SPT.getStateBefore(P)->first;
if (RegionAddr == SPT.SUPERPOSITION || RegionAddr == SPT.EMPTY)
return false;
if (InsertedRegions.count(RegionAddr))
return true;
// Check if we are going to screw up stack accesses at call sites that
// pass parameters via stack
if (!BlacklistedRegions.empty())
return false;
return true;
}
bool StackLayoutModifier::insertRegion(ProgramPoint P, int64_t RegionSz) {
if (!canInsertRegion(P))
return false;
assert(IsInitialized);
StackPointerTracking &SPT = Info.getStackPointerTracking();
// This RegionAddr is slightly different from the one seen in collapseRegion
// This is the value of SP before the allocation the user wants to make.
int64_t RegionAddr = SPT.getStateBefore(P)->first;
if (RegionAddr == SPT.SUPERPOSITION || RegionAddr == SPT.EMPTY)
return false;
DominatorAnalysis<false> &DA = Info.getDominatorAnalysis();
for (BinaryBasicBlock &BB : BF) {
for (MCInst &Inst : BB) {
if (!BC.MIB->hasAnnotation(Inst, getSlotTag()))
continue;
auto Slot =
BC.MIB->getAnnotationAs<decltype(FrameIndexEntry::StackOffset)>(
Inst, getSlotTag());
if (!AvailableRegions.count(Slot))
continue;
if (!(DA.doesADominateB(P, Inst)))
continue;
if (BC.MIB->isCFI(Inst)) {
if (Slot >= RegionAddr)
continue;
scheduleChange(Inst, WorklistItem(WorklistItem::AdjustCFI, -RegionSz));
continue;
}
ErrorOr<const FrameIndexEntry &> FIE = FA.getFIEFor(Inst);
if (!FIE) {
if (Slot >= RegionAddr)
continue;
scheduleChange(
Inst, WorklistItem(WorklistItem::AdjustLoadStoreOffset, -RegionSz));
continue;
}
if (FIE->StackPtrReg == BC.MIB->getStackPointer() && Slot < RegionAddr)
continue;
if (FIE->StackPtrReg == BC.MIB->getFramePointer() && Slot >= RegionAddr)
continue;
if (BC.MIB->isPush(Inst) || BC.MIB->isPop(Inst))
continue;
scheduleChange(
Inst, WorklistItem(WorklistItem::AdjustLoadStoreOffset, -RegionSz));
}
}
InsertedRegions.insert(RegionAddr);
return true;
}
void StackLayoutModifier::performChanges() {
std::set<uint32_t> ModifiedCFIIndices;
for (BinaryBasicBlock &BB : BF) {
for (auto I = BB.rbegin(), E = BB.rend(); I != E; ++I) {
MCInst &Inst = *I;
if (BC.MIB->hasAnnotation(Inst, "AccessesDeletedPos")) {
assert(BC.MIB->isPop(Inst) || BC.MIB->isPush(Inst));
BC.MIB->removeAnnotation(Inst, "AccessesDeletedPos");
}
if (!BC.MIB->hasAnnotation(Inst, getTodoTag()))
continue;
auto &WList = BC.MIB->getAnnotationAs<std::vector<WorklistItem>>(
Inst, getTodoTag());
int64_t Adjustment = 0;
WorklistItem::ActionType AdjustmentType = WorklistItem::None;
for (WorklistItem &WI : WList) {
if (WI.Action == WorklistItem::None)
continue;
assert(WI.Action == WorklistItem::AdjustLoadStoreOffset ||
WI.Action == WorklistItem::AdjustCFI);
assert((AdjustmentType == WorklistItem::None ||
AdjustmentType == WI.Action) &&
"Conflicting actions requested at the same program point");
AdjustmentType = WI.Action;
Adjustment += WI.OffsetUpdate;
}
if (!Adjustment)
continue;
if (AdjustmentType != WorklistItem::AdjustLoadStoreOffset) {
assert(BC.MIB->isCFI(Inst));
uint32_t CFINum = Inst.getOperand(0).getImm();
if (ModifiedCFIIndices.count(CFINum))
continue;
ModifiedCFIIndices.insert(CFINum);
const MCCFIInstruction *CFI = BF.getCFIFor(Inst);
const MCCFIInstruction::OpType Operation = CFI->getOperation();
if (Operation == MCCFIInstruction::OpDefCfa ||
Operation == MCCFIInstruction::OpDefCfaOffset)
Adjustment = 0 - Adjustment;
LLVM_DEBUG(dbgs() << "Changing CFI offset from " << CFI->getOffset()
<< " to " << (CFI->getOffset() + Adjustment) << "\n");
BF.mutateCFIOffsetFor(Inst, CFI->getOffset() + Adjustment);
continue;
}
int32_t SrcImm = 0;
MCPhysReg Reg = 0;
MCPhysReg StackPtrReg = 0;
int64_t StackOffset = 0;
bool IsIndexed = false;
bool IsLoad = false;
bool IsStore = false;
bool IsSimple = false;
bool IsStoreFromReg = false;
uint8_t Size = 0;
bool Success = false;
Success = BC.MIB->isStackAccess(Inst, IsLoad, IsStore, IsStoreFromReg,
Reg, SrcImm, StackPtrReg, StackOffset,
Size, IsSimple, IsIndexed);
if (!Success) {
// SP update based on FP value
Success = BC.MIB->addToImm(Inst, Adjustment, &*BC.Ctx);
assert(Success);
continue;
}
assert(Success && IsSimple && !IsIndexed && (!IsStore || IsStoreFromReg));
if (StackPtrReg != BC.MIB->getFramePointer())
Adjustment = -Adjustment;
if (IsLoad)
Success = BC.MIB->createRestoreFromStack(
Inst, StackPtrReg, StackOffset + Adjustment, Reg, Size);
else if (IsStore)
Success = BC.MIB->createSaveToStack(
Inst, StackPtrReg, StackOffset + Adjustment, Reg, Size);
LLVM_DEBUG({
dbgs() << "Adjusted instruction: ";
Inst.dump();
});
assert(Success);
}
}
}
void StackLayoutModifier::initialize() {
classifyStackAccesses();
classifyCFIs();
IsInitialized = true;
}
uint64_t ShrinkWrapping::SpillsMovedRegularMode = 0;
uint64_t ShrinkWrapping::SpillsMovedPushPopMode = 0;
using BBIterTy = BinaryBasicBlock::iterator;
void ShrinkWrapping::classifyCSRUses() {
DominatorAnalysis<false> &DA = Info.getDominatorAnalysis();
StackPointerTracking &SPT = Info.getStackPointerTracking();
UsesByReg = std::vector<BitVector>(BC.MRI->getNumRegs(),
BitVector(DA.NumInstrs, false));
const BitVector &FPAliases = BC.MIB->getAliases(BC.MIB->getFramePointer());
for (BinaryBasicBlock &BB : BF) {
for (MCInst &Inst : BB) {
if (BC.MIB->isCFI(Inst))
continue;
BitVector BV = BitVector(BC.MRI->getNumRegs(), false);
BC.MIB->getTouchedRegs(Inst, BV);
BV &= CSA.CalleeSaved;
for (int I = BV.find_first(); I != -1; I = BV.find_next(I)) {
if (I == 0)
continue;
if (CSA.getSavedReg(Inst) != I && CSA.getRestoredReg(Inst) != I)
UsesByReg[I].set(DA.ExprToIdx[&Inst]);
}
if (!SPT.HasFramePointer || !BC.MIB->isCall(Inst))
continue;
BV = CSA.CalleeSaved;
BV &= FPAliases;
for (int I = BV.find_first(); I > 0; I = BV.find_next(I))
UsesByReg[I].set(DA.ExprToIdx[&Inst]);
}
}
}
void ShrinkWrapping::pruneUnwantedCSRs() {
BitVector ParamRegs = BC.MIB->getRegsUsedAsParams();
for (unsigned I = 0, E = BC.MRI->getNumRegs(); I != E; ++I) {
if (!CSA.CalleeSaved[I])
continue;
if (ParamRegs[I]) {
CSA.CalleeSaved.reset(I);
continue;
}
if (UsesByReg[I].empty()) {
LLVM_DEBUG(
dbgs()
<< "Dismissing Callee-Saved Reg because we found no uses of it:" << I
<< "\n");
CSA.CalleeSaved.reset(I);
continue;
}
if (!CSA.HasRestores[I]) {
LLVM_DEBUG(
dbgs() << "Dismissing Callee-Saved Reg because it does not have "
"restores:"
<< I << "\n");
CSA.CalleeSaved.reset(I);
}
}
}
void ShrinkWrapping::computeSaveLocations() {
SavePos = std::vector<SmallSetVector<MCInst *, 4>>(BC.MRI->getNumRegs());
ReachingInsns<true> &RI = Info.getReachingInsnsBackwards();
DominatorAnalysis<false> &DA = Info.getDominatorAnalysis();
StackPointerTracking &SPT = Info.getStackPointerTracking();
LLVM_DEBUG(dbgs() << "Checking save/restore possibilities\n");
for (BinaryBasicBlock &BB : BF) {
LLVM_DEBUG(dbgs() << "\tNow at BB " << BB.getName() << "\n");
MCInst *First = BB.begin() != BB.end() ? &*BB.begin() : nullptr;
if (!First)
continue;
// Use reaching instructions to detect if we are inside a loop - if we
// are, do not consider this BB as valid placement for saves.
if (RI.isInLoop(BB))
continue;
const std::pair<int, int> SPFP = *SPT.getStateBefore(*First);
// If we don't know stack state at this point, bail
if ((SPFP.first == SPT.SUPERPOSITION || SPFP.first == SPT.EMPTY) &&
(SPFP.second == SPT.SUPERPOSITION || SPFP.second == SPT.EMPTY))
continue;
for (unsigned I = 0, E = BC.MRI->getNumRegs(); I != E; ++I) {
if (!CSA.CalleeSaved[I])
continue;
BitVector BBDominatedUses = BitVector(DA.NumInstrs, false);
for (int J = UsesByReg[I].find_first(); J > 0;
J = UsesByReg[I].find_next(J))
if (DA.doesADominateB(*First, J))
BBDominatedUses.set(J);
LLVM_DEBUG(dbgs() << "\t\tBB " << BB.getName() << " dominates "
<< BBDominatedUses.count() << " uses for reg " << I
<< ". Total uses for reg is " << UsesByReg[I].count()
<< "\n");
BBDominatedUses &= UsesByReg[I];
if (BBDominatedUses == UsesByReg[I]) {
LLVM_DEBUG(dbgs() << "\t\t\tAdded " << BB.getName()
<< " as a save pos for " << I << "\n");
SavePos[I].insert(First);
LLVM_DEBUG({
dbgs() << "Dominated uses are:\n";
for (int J = UsesByReg[I].find_first(); J > 0;
J = UsesByReg[I].find_next(J)) {
dbgs() << "Idx " << J << ": ";
DA.Expressions[J]->dump();
}
});
}
}
}
BestSaveCount = std::vector<uint64_t>(BC.MRI->getNumRegs(),
std::numeric_limits<uint64_t>::max());
BestSavePos = std::vector<MCInst *>(BC.MRI->getNumRegs(), nullptr);
auto &InsnToBB = Info.getInsnToBBMap();
for (unsigned I = 0, E = BC.MRI->getNumRegs(); I != E; ++I) {
if (!CSA.CalleeSaved[I])
continue;
for (MCInst *Pos : SavePos[I]) {
BinaryBasicBlock *BB = InsnToBB[Pos];
uint64_t Count = BB->getExecutionCount();
if (Count != BinaryBasicBlock::COUNT_NO_PROFILE &&
Count < BestSaveCount[I]) {
BestSavePos[I] = Pos;
BestSaveCount[I] = Count;
}
}
}
}
void ShrinkWrapping::computeDomOrder() {
std::vector<MCPhysReg> Order;
for (MCPhysReg I = 0, E = BC.MRI->getNumRegs(); I != E; ++I) {
Order.push_back(I);
}
DominatorAnalysis<false> &DA = Info.getDominatorAnalysis();
auto &InsnToBB = Info.getInsnToBBMap();
std::sort(Order.begin(), Order.end(),
[&](const MCPhysReg &A, const MCPhysReg &B) {
BinaryBasicBlock *BBA =
BestSavePos[A] ? InsnToBB[BestSavePos[A]] : nullptr;
BinaryBasicBlock *BBB =
BestSavePos[B] ? InsnToBB[BestSavePos[B]] : nullptr;
if (BBA == BBB)
return A < B;
if (!BBA && BBB)
return false;
if (BBA && !BBB)
return true;
if (DA.doesADominateB(*BestSavePos[A], *BestSavePos[B]))
return true;
if (DA.doesADominateB(*BestSavePos[B], *BestSavePos[A]))
return false;
return A < B;
});
for (MCPhysReg I = 0, E = BC.MRI->getNumRegs(); I != E; ++I)
DomOrder[Order[I]] = I;
}
bool ShrinkWrapping::isBestSavePosCold(unsigned CSR, MCInst *&BestPosSave,
uint64_t &TotalEstimatedWin) {
const uint64_t CurSavingCost = CSA.SavingCost[CSR];
if (!CSA.CalleeSaved[CSR])
return false;
uint64_t BestCount = BestSaveCount[CSR];
BestPosSave = BestSavePos[CSR];
bool ShouldMove = false;
if (BestCount != std::numeric_limits<uint64_t>::max() &&
BestCount < (opts::ShrinkWrappingThreshold / 100.0) * CurSavingCost) {
LLVM_DEBUG({
auto &InsnToBB = Info.getInsnToBBMap();
dbgs() << "Better position for saves found in func " << BF.getPrintName()
<< " count << " << BF.getKnownExecutionCount() << "\n";
dbgs() << "Reg: " << CSR
<< "; New BB: " << InsnToBB[BestPosSave]->getName()
<< " Freq reduction: " << (CurSavingCost - BestCount) << "\n";
});
TotalEstimatedWin += CurSavingCost - BestCount;
ShouldMove = true;
}
if (!ShouldMove)
return false;
if (!BestPosSave) {
LLVM_DEBUG({
dbgs() << "Dropping opportunity because we don't know where to put "
"stores -- total est. freq reduc: "
<< TotalEstimatedWin << "\n";
});
return false;
}
return true;
}
/// Auxiliar function used to create basic blocks for critical edges and update
/// the dominance frontier with these new locations
void ShrinkWrapping::splitFrontierCritEdges(
BinaryFunction *Func, SmallVector<ProgramPoint, 4> &Frontier,
const SmallVector<bool, 4> &IsCritEdge,
const SmallVector<BinaryBasicBlock *, 4> &From,
const SmallVector<SmallVector<BinaryBasicBlock *, 4>, 4> &To) {
LLVM_DEBUG(dbgs() << "splitFrontierCritEdges: Now handling func "
<< BF.getPrintName() << "\n");
// For every FromBB, there might be one or more critical edges, with
// To[I] containing destination BBs. It's important to memorize
// the original size of the Frontier as we may append to it while splitting
// critical edges originating with blocks with multiple destinations.
for (size_t I = 0, IE = Frontier.size(); I < IE; ++I) {
if (!IsCritEdge[I])
continue;
if (To[I].empty())
continue;
BinaryBasicBlock *FromBB = From[I];
LLVM_DEBUG(dbgs() << " - Now handling FrontierBB " << FromBB->getName()
<< "\n");
// Split edge for every DestinationBBs
for (size_t DI = 0, DIE = To[I].size(); DI < DIE; ++DI) {
BinaryBasicBlock *DestinationBB = To[I][DI];
LLVM_DEBUG(dbgs() << " - Dest : " << DestinationBB->getName() << "\n");
BinaryBasicBlock *NewBB = Func->splitEdge(FromBB, DestinationBB);
// Insert dummy instruction so this BB is never empty (we need this for
// PredictiveStackPointerTracking to work, since it annotates instructions
// and not BBs).
if (NewBB->empty()) {
MCInst NewInst;
BC.MIB->createNoop(NewInst);
NewBB->addInstruction(std::move(NewInst));
scheduleChange(&*NewBB->begin(), WorklistItem(WorklistItem::Erase, 0));
}
// Update frontier
ProgramPoint NewFrontierPP = ProgramPoint::getLastPointAt(*NewBB);
if (DI == 0) {
// Update frontier inplace
Frontier[I] = NewFrontierPP;
LLVM_DEBUG(dbgs() << " - Update frontier with " << NewBB->getName()
<< '\n');
} else {
// Append new frontier to the end of the list
Frontier.push_back(NewFrontierPP);
LLVM_DEBUG(dbgs() << " - Append frontier " << NewBB->getName()
<< '\n');
}
}
}
}
SmallVector<ProgramPoint, 4>
ShrinkWrapping::doRestorePlacement(MCInst *BestPosSave, unsigned CSR,
uint64_t TotalEstimatedWin) {
SmallVector<ProgramPoint, 4> Frontier;
SmallVector<bool, 4> IsCritEdge;
bool CannotPlace = false;
DominatorAnalysis<false> &DA = Info.getDominatorAnalysis();
SmallVector<BinaryBasicBlock *, 4> CritEdgesFrom;
SmallVector<SmallVector<BinaryBasicBlock *, 4>, 4> CritEdgesTo;
// In case of a critical edge, we need to create extra BBs to host restores
// into edges transitioning to the dominance frontier, otherwise we pull these
// restores to inside the dominated area.
Frontier = DA.getDominanceFrontierFor(*BestPosSave).takeVector();
LLVM_DEBUG({
dbgs() << "Dumping dominance frontier for ";
BC.printInstruction(dbgs(), *BestPosSave);
for (ProgramPoint &PP : Frontier)
if (PP.isInst())
BC.printInstruction(dbgs(), *PP.getInst());
else
dbgs() << PP.getBB()->getName() << "\n";
});
for (ProgramPoint &PP : Frontier) {
bool HasCritEdges = false;
if (PP.isInst() && BC.MIB->isTerminator(*PP.getInst()) &&
doesInstUsesCSR(*PP.getInst(), CSR))
CannotPlace = true;
BinaryBasicBlock *FrontierBB = Info.getParentBB(PP);
CritEdgesFrom.emplace_back(FrontierBB);
CritEdgesTo.emplace_back(0);