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
| //===-- AArch64ConditionalCompares.cpp --- CCMP formation for AArch64 -----===//
//
// 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 AArch64ConditionalCompares pass which reduces
// branching and code size by using the conditional compare instructions CCMP,
// CCMN, and FCMP.
//
// The CFG transformations for forming conditional compares are very similar to
// if-conversion, and this pass should run immediately before the early
// if-conversion pass.
//
//===----------------------------------------------------------------------===//
#include "AArch64.h"
#include "llvm/ADT/DepthFirstIterator.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/MachineTraceMetrics.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
#define DEBUG_TYPE "aarch64-ccmp"
// Absolute maximum number of instructions allowed per speculated block.
// This bypasses all other heuristics, so it should be set fairly high.
static cl::opt<unsigned> BlockInstrLimit(
"aarch64-ccmp-limit", cl::init(30), cl::Hidden,
cl::desc("Maximum number of instructions per speculated block."));
// Stress testing mode - disable heuristics.
static cl::opt<bool> Stress("aarch64-stress-ccmp", cl::Hidden,
cl::desc("Turn all knobs to 11"));
STATISTIC(NumConsidered, "Number of ccmps considered");
STATISTIC(NumPhiRejs, "Number of ccmps rejected (PHI)");
STATISTIC(NumPhysRejs, "Number of ccmps rejected (Physregs)");
STATISTIC(NumPhi2Rejs, "Number of ccmps rejected (PHI2)");
STATISTIC(NumHeadBranchRejs, "Number of ccmps rejected (Head branch)");
STATISTIC(NumCmpBranchRejs, "Number of ccmps rejected (CmpBB branch)");
STATISTIC(NumCmpTermRejs, "Number of ccmps rejected (CmpBB is cbz...)");
STATISTIC(NumImmRangeRejs, "Number of ccmps rejected (Imm out of range)");
STATISTIC(NumLiveDstRejs, "Number of ccmps rejected (Cmp dest live)");
STATISTIC(NumMultNZCVUses, "Number of ccmps rejected (NZCV used)");
STATISTIC(NumUnknNZCVDefs, "Number of ccmps rejected (NZCV def unknown)");
STATISTIC(NumSpeculateRejs, "Number of ccmps rejected (Can't speculate)");
STATISTIC(NumConverted, "Number of ccmp instructions created");
STATISTIC(NumCompBranches, "Number of cbz/cbnz branches converted");
//===----------------------------------------------------------------------===//
// SSACCmpConv
//===----------------------------------------------------------------------===//
//
// The SSACCmpConv class performs ccmp-conversion on SSA form machine code
// after determining if it is possible. The class contains no heuristics;
// external code should be used to determine when ccmp-conversion is a good
// idea.
//
// CCmp-formation works on a CFG representing chained conditions, typically
// from C's short-circuit || and && operators:
//
// From: Head To: Head
// / | CmpBB
// / | / |
// | CmpBB / |
// | / | Tail |
// | / | | |
// Tail | | |
// | | | |
// ... ... ... ...
//
// The Head block is terminated by a br.cond instruction, and the CmpBB block
// contains compare + br.cond. Tail must be a successor of both.
//
// The cmp-conversion turns the compare instruction in CmpBB into a conditional
// compare, and merges CmpBB into Head, speculatively executing its
// instructions. The AArch64 conditional compare instructions have an immediate
// operand that specifies the NZCV flag values when the condition is false and
// the compare isn't executed. This makes it possible to chain compares with
// different condition codes.
//
// Example:
//
// if (a == 5 || b == 17)
// foo();
//
// Head:
// cmp w0, #5
// b.eq Tail
// CmpBB:
// cmp w1, #17
// b.eq Tail
// ...
// Tail:
// bl _foo
//
// Becomes:
//
// Head:
// cmp w0, #5
// ccmp w1, #17, 4, ne ; 4 = nZcv
// b.eq Tail
// ...
// Tail:
// bl _foo
//
// The ccmp condition code is the one that would cause the Head terminator to
// branch to CmpBB.
//
// FIXME: It should also be possible to speculate a block on the critical edge
// between Head and Tail, just like if-converting a diamond.
//
// FIXME: Handle PHIs in Tail by turning them into selects (if-conversion).
namespace {
class SSACCmpConv {
MachineFunction *MF;
const TargetInstrInfo *TII;
const TargetRegisterInfo *TRI;
MachineRegisterInfo *MRI;
const MachineBranchProbabilityInfo *MBPI;
public:
/// The first block containing a conditional branch, dominating everything
/// else.
MachineBasicBlock *Head;
/// The block containing cmp+br.cond with a successor shared with Head.
MachineBasicBlock *CmpBB;
/// The common successor for Head and CmpBB.
MachineBasicBlock *Tail;
/// The compare instruction in CmpBB that can be converted to a ccmp.
MachineInstr *CmpMI;
private:
/// The branch condition in Head as determined by AnalyzeBranch.
SmallVector<MachineOperand, 4> HeadCond;
/// The condition code that makes Head branch to CmpBB.
AArch64CC::CondCode HeadCmpBBCC;
/// The branch condition in CmpBB.
SmallVector<MachineOperand, 4> CmpBBCond;
/// The condition code that makes CmpBB branch to Tail.
AArch64CC::CondCode CmpBBTailCC;
/// Check if the Tail PHIs are trivially convertible.
bool trivialTailPHIs();
/// Remove CmpBB from the Tail PHIs.
void updateTailPHIs();
/// Check if an operand defining DstReg is dead.
bool isDeadDef(unsigned DstReg);
/// Find the compare instruction in MBB that controls the conditional branch.
/// Return NULL if a convertible instruction can't be found.
MachineInstr *findConvertibleCompare(MachineBasicBlock *MBB);
/// Return true if all non-terminator instructions in MBB can be safely
/// speculated.
bool canSpeculateInstrs(MachineBasicBlock *MBB, const MachineInstr *CmpMI);
public:
/// runOnMachineFunction - Initialize per-function data structures.
void runOnMachineFunction(MachineFunction &MF,
const MachineBranchProbabilityInfo *MBPI) {
this->MF = &MF;
this->MBPI = MBPI;
TII = MF.getSubtarget().getInstrInfo();
TRI = MF.getSubtarget().getRegisterInfo();
MRI = &MF.getRegInfo();
}
/// If the sub-CFG headed by MBB can be cmp-converted, initialize the
/// internal state, and return true.
bool canConvert(MachineBasicBlock *MBB);
/// Cmo-convert the last block passed to canConvertCmp(), assuming
/// it is possible. Add any erased blocks to RemovedBlocks.
void convert(SmallVectorImpl<MachineBasicBlock *> &RemovedBlocks);
/// Return the expected code size delta if the conversion into a
/// conditional compare is performed.
int expectedCodeSizeDelta() const;
};
} // end anonymous namespace
// Check that all PHIs in Tail are selecting the same value from Head and CmpBB.
// This means that no if-conversion is required when merging CmpBB into Head.
bool SSACCmpConv::trivialTailPHIs() {
for (auto &I : *Tail) {
if (!I.isPHI())
break;
unsigned HeadReg = 0, CmpBBReg = 0;
// PHI operands come in (VReg, MBB) pairs.
for (unsigned oi = 1, oe = I.getNumOperands(); oi != oe; oi += 2) {
MachineBasicBlock *MBB = I.getOperand(oi + 1).getMBB();
Register Reg = I.getOperand(oi).getReg();
if (MBB == Head) {
assert((!HeadReg || HeadReg == Reg) && "Inconsistent PHI operands");
HeadReg = Reg;
}
if (MBB == CmpBB) {
assert((!CmpBBReg || CmpBBReg == Reg) && "Inconsistent PHI operands");
CmpBBReg = Reg;
}
}
if (HeadReg != CmpBBReg)
return false;
}
return true;
}
// Assuming that trivialTailPHIs() is true, update the Tail PHIs by simply
// removing the CmpBB operands. The Head operands will be identical.
void SSACCmpConv::updateTailPHIs() {
for (auto &I : *Tail) {
if (!I.isPHI())
break;
// I is a PHI. It can have multiple entries for CmpBB.
for (unsigned oi = I.getNumOperands(); oi > 2; oi -= 2) {
// PHI operands are (Reg, MBB) at (oi-2, oi-1).
if (I.getOperand(oi - 1).getMBB() == CmpBB) {
I.RemoveOperand(oi - 1);
I.RemoveOperand(oi - 2);
}
}
}
}
// This pass runs before the AArch64DeadRegisterDefinitions pass, so compares
// are still writing virtual registers without any uses.
bool SSACCmpConv::isDeadDef(unsigned DstReg) {
// Writes to the zero register are dead.
if (DstReg == AArch64::WZR || DstReg == AArch64::XZR)
return true;
if (!Register::isVirtualRegister(DstReg))
return false;
// A virtual register def without any uses will be marked dead later, and
// eventually replaced by the zero register.
return MRI->use_nodbg_empty(DstReg);
}
// Parse a condition code returned by AnalyzeBranch, and compute the CondCode
// corresponding to TBB.
// Return
static bool parseCond(ArrayRef<MachineOperand> Cond, AArch64CC::CondCode &CC) {
// A normal br.cond simply has the condition code.
if (Cond[0].getImm() != -1) {
assert(Cond.size() == 1 && "Unknown Cond array format");
CC = (AArch64CC::CondCode)(int)Cond[0].getImm();
return true;
}
// For tbz and cbz instruction, the opcode is next.
switch (Cond[1].getImm()) {
default:
// This includes tbz / tbnz branches which can't be converted to
// ccmp + br.cond.
return false;
case AArch64::CBZW:
case AArch64::CBZX:
assert(Cond.size() == 3 && "Unknown Cond array format");
CC = AArch64CC::EQ;
return true;
case AArch64::CBNZW:
case AArch64::CBNZX:
assert(Cond.size() == 3 && "Unknown Cond array format");
CC = AArch64CC::NE;
return true;
}
}
MachineInstr *SSACCmpConv::findConvertibleCompare(MachineBasicBlock *MBB) {
MachineBasicBlock::iterator I = MBB->getFirstTerminator();
if (I == MBB->end())
return nullptr;
// The terminator must be controlled by the flags.
if (!I->readsRegister(AArch64::NZCV)) {
switch (I->getOpcode()) {
case AArch64::CBZW:
case AArch64::CBZX:
case AArch64::CBNZW:
case AArch64::CBNZX:
// These can be converted into a ccmp against #0.
return &*I;
}
++NumCmpTermRejs;
LLVM_DEBUG(dbgs() << "Flags not used by terminator: " << *I);
return nullptr;
}
// Now find the instruction controlling the terminator.
for (MachineBasicBlock::iterator B = MBB->begin(); I != B;) {
--I;
assert(!I->isTerminator() && "Spurious terminator");
switch (I->getOpcode()) {
// cmp is an alias for subs with a dead destination register.
case AArch64::SUBSWri:
case AArch64::SUBSXri:
// cmn is an alias for adds with a dead destination register.
case AArch64::ADDSWri:
case AArch64::ADDSXri:
// Check that the immediate operand is within range, ccmp wants a uimm5.
// Rd = SUBSri Rn, imm, shift
if (I->getOperand(3).getImm() || !isUInt<5>(I->getOperand(2).getImm())) {
LLVM_DEBUG(dbgs() << "Immediate out of range for ccmp: " << *I);
++NumImmRangeRejs;
return nullptr;
}
LLVM_FALLTHROUGH;
case AArch64::SUBSWrr:
case AArch64::SUBSXrr:
case AArch64::ADDSWrr:
case AArch64::ADDSXrr:
if (isDeadDef(I->getOperand(0).getReg()))
return &*I;
LLVM_DEBUG(dbgs() << "Can't convert compare with live destination: "
<< *I);
++NumLiveDstRejs;
return nullptr;
case AArch64::FCMPSrr:
case AArch64::FCMPDrr:
case AArch64::FCMPESrr:
case AArch64::FCMPEDrr:
return &*I;
}
// Check for flag reads and clobbers.
MIOperands::PhysRegInfo PRI =
MIOperands(*I).analyzePhysReg(AArch64::NZCV, TRI);
if (PRI.Read) {
// The ccmp doesn't produce exactly the same flags as the original
// compare, so reject the transform if there are uses of the flags
// besides the terminators.
LLVM_DEBUG(dbgs() << "Can't create ccmp with multiple uses: " << *I);
++NumMultNZCVUses;
return nullptr;
}
if (PRI.Defined || PRI.Clobbered) {
LLVM_DEBUG(dbgs() << "Not convertible compare: " << *I);
++NumUnknNZCVDefs;
return nullptr;
}
}
LLVM_DEBUG(dbgs() << "Flags not defined in " << printMBBReference(*MBB)
<< '\n');
return nullptr;
}
/// Determine if all the instructions in MBB can safely
/// be speculated. The terminators are not considered.
///
/// Only CmpMI is allowed to clobber the flags.
///
bool SSACCmpConv::canSpeculateInstrs(MachineBasicBlock *MBB,
const MachineInstr *CmpMI) {
// Reject any live-in physregs. It's probably NZCV/EFLAGS, and very hard to
// get right.
if (!MBB->livein_empty()) {
LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << " has live-ins.\n");
return false;
}
unsigned InstrCount = 0;
// Check all instructions, except the terminators. It is assumed that
// terminators never have side effects or define any used register values.
for (auto &I : make_range(MBB->begin(), MBB->getFirstTerminator())) {
if (I.isDebugInstr())
continue;
if (++InstrCount > BlockInstrLimit && !Stress) {
LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << " has more than "
<< BlockInstrLimit << " instructions.\n");
return false;
}
// There shouldn't normally be any phis in a single-predecessor block.
if (I.isPHI()) {
LLVM_DEBUG(dbgs() << "Can't hoist: " << I);
return false;
}
// Don't speculate loads. Note that it may be possible and desirable to
// speculate GOT or constant pool loads that are guaranteed not to trap,
// but we don't support that for now.
if (I.mayLoad()) {
LLVM_DEBUG(dbgs() << "Won't speculate load: " << I);
return false;
}
// We never speculate stores, so an AA pointer isn't necessary.
bool DontMoveAcrossStore = true;
if (!I.isSafeToMove(nullptr, DontMoveAcrossStore)) {
LLVM_DEBUG(dbgs() << "Can't speculate: " << I);
return false;
}
// Only CmpMI is allowed to clobber the flags.
if (&I != CmpMI && I.modifiesRegister(AArch64::NZCV, TRI)) {
LLVM_DEBUG(dbgs() << "Clobbers flags: " << I);
return false;
}
}
return true;
}
/// Analyze the sub-cfg rooted in MBB, and return true if it is a potential
/// candidate for cmp-conversion. Fill out the internal state.
///
bool SSACCmpConv::canConvert(MachineBasicBlock *MBB) {
Head = MBB;
Tail = CmpBB = nullptr;
if (Head->succ_size() != 2)
return false;
MachineBasicBlock *Succ0 = Head->succ_begin()[0];
MachineBasicBlock *Succ1 = Head->succ_begin()[1];
// CmpBB can only have a single predecessor. Tail is allowed many.
if (Succ0->pred_size() != 1)
std::swap(Succ0, Succ1);
// Succ0 is our candidate for CmpBB.
if (Succ0->pred_size() != 1 || Succ0->succ_size() != 2)
return false;
CmpBB = Succ0;
Tail = Succ1;
if (!CmpBB->isSuccessor(Tail))
return false;
// The CFG topology checks out.
LLVM_DEBUG(dbgs() << "\nTriangle: " << printMBBReference(*Head) << " -> "
<< printMBBReference(*CmpBB) << " -> "
<< printMBBReference(*Tail) << '\n');
++NumConsidered;
// Tail is allowed to have many predecessors, but we can't handle PHIs yet.
//
// FIXME: Real PHIs could be if-converted as long as the CmpBB values are
// defined before The CmpBB cmp clobbers the flags. Alternatively, it should
// always be safe to sink the ccmp down to immediately before the CmpBB
// terminators.
if (!trivialTailPHIs()) {
LLVM_DEBUG(dbgs() << "Can't handle phis in Tail.\n");
++NumPhiRejs;
return false;
}
if (!Tail->livein_empty()) {
LLVM_DEBUG(dbgs() << "Can't handle live-in physregs in Tail.\n");
++NumPhysRejs;
return false;
}
// CmpBB should never have PHIs since Head is its only predecessor.
// FIXME: Clean them up if it happens.
if (!CmpBB->empty() && CmpBB->front().isPHI()) {
LLVM_DEBUG(dbgs() << "Can't handle phis in CmpBB.\n");
++NumPhi2Rejs;
return false;
}
if (!CmpBB->livein_empty()) {
LLVM_DEBUG(dbgs() << "Can't handle live-in physregs in CmpBB.\n");
++NumPhysRejs;
return false;
}
// The branch we're looking to eliminate must be analyzable.
HeadCond.clear();
MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
if (TII->analyzeBranch(*Head, TBB, FBB, HeadCond)) {
LLVM_DEBUG(dbgs() << "Head branch not analyzable.\n");
++NumHeadBranchRejs;
return false;
}
// This is weird, probably some sort of degenerate CFG, or an edge to a
// landing pad.
if (!TBB || HeadCond.empty()) {
LLVM_DEBUG(
dbgs() << "AnalyzeBranch didn't find conditional branch in Head.\n");
++NumHeadBranchRejs;
return false;
}
if (!parseCond(HeadCond, HeadCmpBBCC)) {
LLVM_DEBUG(dbgs() << "Unsupported branch type on Head\n");
++NumHeadBranchRejs;
return false;
}
// Make sure the branch direction is right.
if (TBB != CmpBB) {
assert(TBB == Tail && "Unexpected TBB");
HeadCmpBBCC = AArch64CC::getInvertedCondCode(HeadCmpBBCC);
}
CmpBBCond.clear();
TBB = FBB = nullptr;
if (TII->analyzeBranch(*CmpBB, TBB, FBB, CmpBBCond)) {
LLVM_DEBUG(dbgs() << "CmpBB branch not analyzable.\n");
++NumCmpBranchRejs;
return false;
}
if (!TBB || CmpBBCond.empty()) {
LLVM_DEBUG(
dbgs() << "AnalyzeBranch didn't find conditional branch in CmpBB.\n");
++NumCmpBranchRejs;
return false;
}
if (!parseCond(CmpBBCond, CmpBBTailCC)) {
LLVM_DEBUG(dbgs() << "Unsupported branch type on CmpBB\n");
++NumCmpBranchRejs;
return false;
}
if (TBB != Tail)
CmpBBTailCC = AArch64CC::getInvertedCondCode(CmpBBTailCC);
LLVM_DEBUG(dbgs() << "Head->CmpBB on "
<< AArch64CC::getCondCodeName(HeadCmpBBCC)
<< ", CmpBB->Tail on "
<< AArch64CC::getCondCodeName(CmpBBTailCC) << '\n');
CmpMI = findConvertibleCompare(CmpBB);
if (!CmpMI)
return false;
if (!canSpeculateInstrs(CmpBB, CmpMI)) {
++NumSpeculateRejs;
return false;
}
return true;
}
void SSACCmpConv::convert(SmallVectorImpl<MachineBasicBlock *> &RemovedBlocks) {
LLVM_DEBUG(dbgs() << "Merging " << printMBBReference(*CmpBB) << " into "
<< printMBBReference(*Head) << ":\n"
<< *CmpBB);
// All CmpBB instructions are moved into Head, and CmpBB is deleted.
// Update the CFG first.
updateTailPHIs();
// Save successor probabilties before removing CmpBB and Tail from their
// parents.
BranchProbability Head2CmpBB = MBPI->getEdgeProbability(Head, CmpBB);
BranchProbability CmpBB2Tail = MBPI->getEdgeProbability(CmpBB, Tail);
Head->removeSuccessor(CmpBB);
CmpBB->removeSuccessor(Tail);
// If Head and CmpBB had successor probabilties, udpate the probabilities to
// reflect the ccmp-conversion.
if (Head->hasSuccessorProbabilities() && CmpBB->hasSuccessorProbabilities()) {
// Head is allowed two successors. We've removed CmpBB, so the remaining
// successor is Tail. We need to increase the successor probability for
// Tail to account for the CmpBB path we removed.
//
// Pr(Tail|Head) += Pr(CmpBB|Head) * Pr(Tail|CmpBB).
assert(*Head->succ_begin() == Tail && "Head successor is not Tail");
BranchProbability Head2Tail = MBPI->getEdgeProbability(Head, Tail);
Head->setSuccProbability(Head->succ_begin(),
Head2Tail + Head2CmpBB * CmpBB2Tail);
// We will transfer successors of CmpBB to Head in a moment without
// normalizing the successor probabilities. Set the successor probabilites
// before doing so.
//
// Pr(I|Head) = Pr(CmpBB|Head) * Pr(I|CmpBB).
for (auto I = CmpBB->succ_begin(), E = CmpBB->succ_end(); I != E; ++I) {
BranchProbability CmpBB2I = MBPI->getEdgeProbability(CmpBB, *I);
CmpBB->setSuccProbability(I, Head2CmpBB * CmpBB2I);
}
}
Head->transferSuccessorsAndUpdatePHIs(CmpBB);
DebugLoc TermDL = Head->getFirstTerminator()->getDebugLoc();
TII->removeBranch(*Head);
// If the Head terminator was one of the cbz / tbz branches with built-in
// compare, we need to insert an explicit compare instruction in its place.
if (HeadCond[0].getImm() == -1) {
++NumCompBranches;
unsigned Opc = 0;
switch (HeadCond[1].getImm()) {
case AArch64::CBZW:
case AArch64::CBNZW:
Opc = AArch64::SUBSWri;
break;
case AArch64::CBZX:
case AArch64::CBNZX:
Opc = AArch64::SUBSXri;
break;
default:
llvm_unreachable("Cannot convert Head branch");
}
const MCInstrDesc &MCID = TII->get(Opc);
// Create a dummy virtual register for the SUBS def.
Register DestReg =
MRI->createVirtualRegister(TII->getRegClass(MCID, 0, TRI, *MF));
// Insert a SUBS Rn, #0 instruction instead of the cbz / cbnz.
BuildMI(*Head, Head->end(), TermDL, MCID)
.addReg(DestReg, RegState::Define | RegState::Dead)
.add(HeadCond[2])
.addImm(0)
.addImm(0);
// SUBS uses the GPR*sp register classes.
MRI->constrainRegClass(HeadCond[2].getReg(),
TII->getRegClass(MCID, 1, TRI, *MF));
}
Head->splice(Head->end(), CmpBB, CmpBB->begin(), CmpBB->end());
// Now replace CmpMI with a ccmp instruction that also considers the incoming
// flags.
unsigned Opc = 0;
unsigned FirstOp = 1; // First CmpMI operand to copy.
bool isZBranch = false; // CmpMI is a cbz/cbnz instruction.
switch (CmpMI->getOpcode()) {
default:
llvm_unreachable("Unknown compare opcode");
case AArch64::SUBSWri: Opc = AArch64::CCMPWi; break;
case AArch64::SUBSWrr: Opc = AArch64::CCMPWr; break;
case AArch64::SUBSXri: Opc = AArch64::CCMPXi; break;
case AArch64::SUBSXrr: Opc = AArch64::CCMPXr; break;
case AArch64::ADDSWri: Opc = AArch64::CCMNWi; break;
case AArch64::ADDSWrr: Opc = AArch64::CCMNWr; break;
case AArch64::ADDSXri: Opc = AArch64::CCMNXi; break;
case AArch64::ADDSXrr: Opc = AArch64::CCMNXr; break;
case AArch64::FCMPSrr: Opc = AArch64::FCCMPSrr; FirstOp = 0; break;
case AArch64::FCMPDrr: Opc = AArch64::FCCMPDrr; FirstOp = 0; break;
case AArch64::FCMPESrr: Opc = AArch64::FCCMPESrr; FirstOp = 0; break;
case AArch64::FCMPEDrr: Opc = AArch64::FCCMPEDrr; FirstOp = 0; break;
case AArch64::CBZW:
case AArch64::CBNZW:
Opc = AArch64::CCMPWi;
FirstOp = 0;
isZBranch = true;
break;
case AArch64::CBZX:
case AArch64::CBNZX:
Opc = AArch64::CCMPXi;
FirstOp = 0;
isZBranch = true;
break;
}
// The ccmp instruction should set the flags according to the comparison when
// Head would have branched to CmpBB.
// The NZCV immediate operand should provide flags for the case where Head
// would have branched to Tail. These flags should cause the new Head
// terminator to branch to tail.
unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(CmpBBTailCC);
const MCInstrDesc &MCID = TII->get(Opc);
MRI->constrainRegClass(CmpMI->getOperand(FirstOp).getReg(),
TII->getRegClass(MCID, 0, TRI, *MF));
if (CmpMI->getOperand(FirstOp + 1).isReg())
MRI->constrainRegClass(CmpMI->getOperand(FirstOp + 1).getReg(),
TII->getRegClass(MCID, 1, TRI, *MF));
MachineInstrBuilder MIB = BuildMI(*Head, CmpMI, CmpMI->getDebugLoc(), MCID)
.add(CmpMI->getOperand(FirstOp)); // Register Rn
if (isZBranch)
MIB.addImm(0); // cbz/cbnz Rn -> ccmp Rn, #0
else
MIB.add(CmpMI->getOperand(FirstOp + 1)); // Register Rm / Immediate
MIB.addImm(NZCV).addImm(HeadCmpBBCC);
// If CmpMI was a terminator, we need a new conditional branch to replace it.
// This now becomes a Head terminator.
if (isZBranch) {
bool isNZ = CmpMI->getOpcode() == AArch64::CBNZW ||
CmpMI->getOpcode() == AArch64::CBNZX;
BuildMI(*Head, CmpMI, CmpMI->getDebugLoc(), TII->get(AArch64::Bcc))
.addImm(isNZ ? AArch64CC::NE : AArch64CC::EQ)
.add(CmpMI->getOperand(1)); // Branch target.
}
CmpMI->eraseFromParent();
Head->updateTerminator();
RemovedBlocks.push_back(CmpBB);
CmpBB->eraseFromParent();
LLVM_DEBUG(dbgs() << "Result:\n" << *Head);
++NumConverted;
}
int SSACCmpConv::expectedCodeSizeDelta() const {
int delta = 0;
// If the Head terminator was one of the cbz / tbz branches with built-in
// compare, we need to insert an explicit compare instruction in its place
// plus a branch instruction.
if (HeadCond[0].getImm() == -1) {
switch (HeadCond[1].getImm()) {
case AArch64::CBZW:
case AArch64::CBNZW:
case AArch64::CBZX:
case AArch64::CBNZX:
// Therefore delta += 1
delta = 1;
break;
default:
llvm_unreachable("Cannot convert Head branch");
}
}
// If the Cmp terminator was one of the cbz / tbz branches with
// built-in compare, it will be turned into a compare instruction
// into Head, but we do not save any instruction.
// Otherwise, we save the branch instruction.
switch (CmpMI->getOpcode()) {
default:
--delta;
break;
case AArch64::CBZW:
case AArch64::CBNZW:
case AArch64::CBZX:
case AArch64::CBNZX:
break;
}
return delta;
}
//===----------------------------------------------------------------------===//
// AArch64ConditionalCompares Pass
//===----------------------------------------------------------------------===//
namespace {
class AArch64ConditionalCompares : public MachineFunctionPass {
const MachineBranchProbabilityInfo *MBPI;
const TargetInstrInfo *TII;
const TargetRegisterInfo *TRI;
MCSchedModel SchedModel;
// Does the proceeded function has Oz attribute.
bool MinSize;
MachineRegisterInfo *MRI;
MachineDominatorTree *DomTree;
MachineLoopInfo *Loops;
MachineTraceMetrics *Traces;
MachineTraceMetrics::Ensemble *MinInstr;
SSACCmpConv CmpConv;
public:
static char ID;
AArch64ConditionalCompares() : MachineFunctionPass(ID) {
initializeAArch64ConditionalComparesPass(*PassRegistry::getPassRegistry());
}
void getAnalysisUsage(AnalysisUsage &AU) const override;
bool runOnMachineFunction(MachineFunction &MF) override;
StringRef getPassName() const override {
return "AArch64 Conditional Compares";
}
private:
bool tryConvert(MachineBasicBlock *);
void updateDomTree(ArrayRef<MachineBasicBlock *> Removed);
void updateLoops(ArrayRef<MachineBasicBlock *> Removed);
void invalidateTraces();
bool shouldConvert();
};
} // end anonymous namespace
char AArch64ConditionalCompares::ID = 0;
INITIALIZE_PASS_BEGIN(AArch64ConditionalCompares, "aarch64-ccmp",
"AArch64 CCMP Pass", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
INITIALIZE_PASS_DEPENDENCY(MachineTraceMetrics)
INITIALIZE_PASS_END(AArch64ConditionalCompares, "aarch64-ccmp",
"AArch64 CCMP Pass", false, false)
FunctionPass *llvm::createAArch64ConditionalCompares() {
return new AArch64ConditionalCompares();
}
void AArch64ConditionalCompares::getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<MachineBranchProbabilityInfo>();
AU.addRequired<MachineDominatorTree>();
AU.addPreserved<MachineDominatorTree>();
AU.addRequired<MachineLoopInfo>();
AU.addPreserved<MachineLoopInfo>();
AU.addRequired<MachineTraceMetrics>();
AU.addPreserved<MachineTraceMetrics>();
MachineFunctionPass::getAnalysisUsage(AU);
}
/// Update the dominator tree after if-conversion erased some blocks.
void AArch64ConditionalCompares::updateDomTree(
ArrayRef<MachineBasicBlock *> Removed) {
// convert() removes CmpBB which was previously dominated by Head.
// CmpBB children should be transferred to Head.
MachineDomTreeNode *HeadNode = DomTree->getNode(CmpConv.Head);
for (MachineBasicBlock *RemovedMBB : Removed) {
MachineDomTreeNode *Node = DomTree->getNode(RemovedMBB);
assert(Node != HeadNode && "Cannot erase the head node");
assert(Node->getIDom() == HeadNode && "CmpBB should be dominated by Head");
while (Node->getNumChildren())
DomTree->changeImmediateDominator(Node->getChildren().back(), HeadNode);
DomTree->eraseNode(RemovedMBB);
}
}
/// Update LoopInfo after if-conversion.
void
AArch64ConditionalCompares::updateLoops(ArrayRef<MachineBasicBlock *> Removed) {
if (!Loops)
return;
for (MachineBasicBlock *RemovedMBB : Removed)
Loops->removeBlock(RemovedMBB);
}
/// Invalidate MachineTraceMetrics before if-conversion.
void AArch64ConditionalCompares::invalidateTraces() {
Traces->invalidate(CmpConv.Head);
Traces->invalidate(CmpConv.CmpBB);
}
/// Apply cost model and heuristics to the if-conversion in IfConv.
/// Return true if the conversion is a good idea.
///
bool AArch64ConditionalCompares::shouldConvert() {
// Stress testing mode disables all cost considerations.
if (Stress)
return true;
if (!MinInstr)
MinInstr = Traces->getEnsemble(MachineTraceMetrics::TS_MinInstrCount);
// Head dominates CmpBB, so it is always included in its trace.
MachineTraceMetrics::Trace Trace = MinInstr->getTrace(CmpConv.CmpBB);
// If code size is the main concern
if (MinSize) {
int CodeSizeDelta = CmpConv.expectedCodeSizeDelta();
LLVM_DEBUG(dbgs() << "Code size delta: " << CodeSizeDelta << '\n');
// If we are minimizing the code size, do the conversion whatever
// the cost is.
if (CodeSizeDelta < 0)
return true;
if (CodeSizeDelta > 0) {
LLVM_DEBUG(dbgs() << "Code size is increasing, give up on this one.\n");
return false;
}
// CodeSizeDelta == 0, continue with the regular heuristics
}
// Heuristic: The compare conversion delays the execution of the branch
// instruction because we must wait for the inputs to the second compare as
// well. The branch has no dependent instructions, but delaying it increases
// the cost of a misprediction.
//
// Set a limit on the delay we will accept.
unsigned DelayLimit = SchedModel.MispredictPenalty * 3 / 4;
// Instruction depths can be computed for all trace instructions above CmpBB.
unsigned HeadDepth =
Trace.getInstrCycles(*CmpConv.Head->getFirstTerminator()).Depth;
unsigned CmpBBDepth =
Trace.getInstrCycles(*CmpConv.CmpBB->getFirstTerminator()).Depth;
LLVM_DEBUG(dbgs() << "Head depth: " << HeadDepth
<< "\nCmpBB depth: " << CmpBBDepth << '\n');
if (CmpBBDepth > HeadDepth + DelayLimit) {
LLVM_DEBUG(dbgs() << "Branch delay would be larger than " << DelayLimit
<< " cycles.\n");
return false;
}
// Check the resource depth at the bottom of CmpBB - these instructions will
// be speculated.
unsigned ResDepth = Trace.getResourceDepth(true);
LLVM_DEBUG(dbgs() << "Resources: " << ResDepth << '\n');
// Heuristic: The speculatively executed instructions must all be able to
// merge into the Head block. The Head critical path should dominate the
// resource cost of the speculated instructions.
if (ResDepth > HeadDepth) {
LLVM_DEBUG(dbgs() << "Too many instructions to speculate.\n");
return false;
}
return true;
}
bool AArch64ConditionalCompares::tryConvert(MachineBasicBlock *MBB) {
bool Changed = false;
while (CmpConv.canConvert(MBB) && shouldConvert()) {
invalidateTraces();
SmallVector<MachineBasicBlock *, 4> RemovedBlocks;
CmpConv.convert(RemovedBlocks);
Changed = true;
updateDomTree(RemovedBlocks);
updateLoops(RemovedBlocks);
}
return Changed;
}
bool AArch64ConditionalCompares::runOnMachineFunction(MachineFunction &MF) {
LLVM_DEBUG(dbgs() << "********** AArch64 Conditional Compares **********\n"
<< "********** Function: " << MF.getName() << '\n');
if (skipFunction(MF.getFunction()))
return false;
TII = MF.getSubtarget().getInstrInfo();
TRI = MF.getSubtarget().getRegisterInfo();
SchedModel = MF.getSubtarget().getSchedModel();
MRI = &MF.getRegInfo();
DomTree = &getAnalysis<MachineDominatorTree>();
Loops = getAnalysisIfAvailable<MachineLoopInfo>();
MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
Traces = &getAnalysis<MachineTraceMetrics>();
MinInstr = nullptr;
MinSize = MF.getFunction().hasMinSize();
bool Changed = false;
CmpConv.runOnMachineFunction(MF, MBPI);
// Visit blocks in dominator tree pre-order. The pre-order enables multiple
// cmp-conversions from the same head block.
// Note that updateDomTree() modifies the children of the DomTree node
// currently being visited. The df_iterator supports that; it doesn't look at
// child_begin() / child_end() until after a node has been visited.
for (auto *I : depth_first(DomTree))
if (tryConvert(I->getBlock()))
Changed = true;
return Changed;
}
|