-
Notifications
You must be signed in to change notification settings - Fork 10.8k
/
SIInstrInfo.cpp
8304 lines (7183 loc) · 289 KB
/
SIInstrInfo.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
//===- SIInstrInfo.cpp - SI Instruction Information ----------------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
/// \file
/// SI Implementation of TargetInstrInfo.
//
//===----------------------------------------------------------------------===//
#include "SIInstrInfo.h"
#include "AMDGPU.h"
#include "AMDGPUInstrInfo.h"
#include "GCNHazardRecognizer.h"
#include "GCNSubtarget.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "SIMachineFunctionInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/CodeGen/LiveIntervals.h"
#include "llvm/CodeGen/LiveVariables.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineScheduler.h"
#include "llvm/CodeGen/RegisterScavenging.h"
#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/IntrinsicsAMDGPU.h"
#include "llvm/MC/MCContext.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Target/TargetMachine.h"
using namespace llvm;
#define DEBUG_TYPE "si-instr-info"
#define GET_INSTRINFO_CTOR_DTOR
#include "AMDGPUGenInstrInfo.inc"
namespace llvm {
class AAResults;
namespace AMDGPU {
#define GET_D16ImageDimIntrinsics_IMPL
#define GET_ImageDimIntrinsicTable_IMPL
#define GET_RsrcIntrinsics_IMPL
#include "AMDGPUGenSearchableTables.inc"
}
}
// Must be at least 4 to be able to branch over minimum unconditional branch
// code. This is only for making it possible to write reasonably small tests for
// long branches.
static cl::opt<unsigned>
BranchOffsetBits("amdgpu-s-branch-bits", cl::ReallyHidden, cl::init(16),
cl::desc("Restrict range of branch instructions (DEBUG)"));
static cl::opt<bool> Fix16BitCopies(
"amdgpu-fix-16-bit-physreg-copies",
cl::desc("Fix copies between 32 and 16 bit registers by extending to 32 bit"),
cl::init(true),
cl::ReallyHidden);
SIInstrInfo::SIInstrInfo(const GCNSubtarget &ST)
: AMDGPUGenInstrInfo(AMDGPU::ADJCALLSTACKUP, AMDGPU::ADJCALLSTACKDOWN),
RI(ST), ST(ST) {
SchedModel.init(&ST);
}
//===----------------------------------------------------------------------===//
// TargetInstrInfo callbacks
//===----------------------------------------------------------------------===//
static unsigned getNumOperandsNoGlue(SDNode *Node) {
unsigned N = Node->getNumOperands();
while (N && Node->getOperand(N - 1).getValueType() == MVT::Glue)
--N;
return N;
}
/// Returns true if both nodes have the same value for the given
/// operand \p Op, or if both nodes do not have this operand.
static bool nodesHaveSameOperandValue(SDNode *N0, SDNode* N1, unsigned OpName) {
unsigned Opc0 = N0->getMachineOpcode();
unsigned Opc1 = N1->getMachineOpcode();
int Op0Idx = AMDGPU::getNamedOperandIdx(Opc0, OpName);
int Op1Idx = AMDGPU::getNamedOperandIdx(Opc1, OpName);
if (Op0Idx == -1 && Op1Idx == -1)
return true;
if ((Op0Idx == -1 && Op1Idx != -1) ||
(Op1Idx == -1 && Op0Idx != -1))
return false;
// getNamedOperandIdx returns the index for the MachineInstr's operands,
// which includes the result as the first operand. We are indexing into the
// MachineSDNode's operands, so we need to skip the result operand to get
// the real index.
--Op0Idx;
--Op1Idx;
return N0->getOperand(Op0Idx) == N1->getOperand(Op1Idx);
}
bool SIInstrInfo::isReallyTriviallyReMaterializable(const MachineInstr &MI,
AAResults *AA) const {
if (isVOP1(MI) || isVOP2(MI) || isVOP3(MI) || isSDWA(MI) || isSALU(MI)) {
// Normally VALU use of exec would block the rematerialization, but that
// is OK in this case to have an implicit exec read as all VALU do.
// We really want all of the generic logic for this except for this.
// Another potential implicit use is mode register. The core logic of
// the RA will not attempt rematerialization if mode is set anywhere
// in the function, otherwise it is safe since mode is not changed.
// There is difference to generic method which does not allow
// rematerialization if there are virtual register uses. We allow this,
// therefore this method includes SOP instructions as well.
return !MI.hasImplicitDef() &&
MI.getNumImplicitOperands() == MI.getDesc().getNumImplicitUses() &&
!MI.mayRaiseFPException();
}
return false;
}
static bool readsExecAsData(const MachineInstr &MI) {
if (MI.isCompare())
return true;
switch (MI.getOpcode()) {
default:
break;
case AMDGPU::V_READFIRSTLANE_B32:
return true;
}
return false;
}
bool SIInstrInfo::isIgnorableUse(const MachineOperand &MO) const {
// Any implicit use of exec by VALU is not a real register read.
return MO.getReg() == AMDGPU::EXEC && MO.isImplicit() &&
isVALU(*MO.getParent()) && !readsExecAsData(*MO.getParent());
}
bool SIInstrInfo::areLoadsFromSameBasePtr(SDNode *Load0, SDNode *Load1,
int64_t &Offset0,
int64_t &Offset1) const {
if (!Load0->isMachineOpcode() || !Load1->isMachineOpcode())
return false;
unsigned Opc0 = Load0->getMachineOpcode();
unsigned Opc1 = Load1->getMachineOpcode();
// Make sure both are actually loads.
if (!get(Opc0).mayLoad() || !get(Opc1).mayLoad())
return false;
if (isDS(Opc0) && isDS(Opc1)) {
// FIXME: Handle this case:
if (getNumOperandsNoGlue(Load0) != getNumOperandsNoGlue(Load1))
return false;
// Check base reg.
if (Load0->getOperand(0) != Load1->getOperand(0))
return false;
// Skip read2 / write2 variants for simplicity.
// TODO: We should report true if the used offsets are adjacent (excluded
// st64 versions).
int Offset0Idx = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
int Offset1Idx = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
if (Offset0Idx == -1 || Offset1Idx == -1)
return false;
// XXX - be careful of datalesss loads
// getNamedOperandIdx returns the index for MachineInstrs. Since they
// include the output in the operand list, but SDNodes don't, we need to
// subtract the index by one.
Offset0Idx -= get(Opc0).NumDefs;
Offset1Idx -= get(Opc1).NumDefs;
Offset0 = cast<ConstantSDNode>(Load0->getOperand(Offset0Idx))->getZExtValue();
Offset1 = cast<ConstantSDNode>(Load1->getOperand(Offset1Idx))->getZExtValue();
return true;
}
if (isSMRD(Opc0) && isSMRD(Opc1)) {
// Skip time and cache invalidation instructions.
if (AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::sbase) == -1 ||
AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::sbase) == -1)
return false;
assert(getNumOperandsNoGlue(Load0) == getNumOperandsNoGlue(Load1));
// Check base reg.
if (Load0->getOperand(0) != Load1->getOperand(0))
return false;
const ConstantSDNode *Load0Offset =
dyn_cast<ConstantSDNode>(Load0->getOperand(1));
const ConstantSDNode *Load1Offset =
dyn_cast<ConstantSDNode>(Load1->getOperand(1));
if (!Load0Offset || !Load1Offset)
return false;
Offset0 = Load0Offset->getZExtValue();
Offset1 = Load1Offset->getZExtValue();
return true;
}
// MUBUF and MTBUF can access the same addresses.
if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1))) {
// MUBUF and MTBUF have vaddr at different indices.
if (!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::soffset) ||
!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::vaddr) ||
!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::srsrc))
return false;
int OffIdx0 = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
int OffIdx1 = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
if (OffIdx0 == -1 || OffIdx1 == -1)
return false;
// getNamedOperandIdx returns the index for MachineInstrs. Since they
// include the output in the operand list, but SDNodes don't, we need to
// subtract the index by one.
OffIdx0 -= get(Opc0).NumDefs;
OffIdx1 -= get(Opc1).NumDefs;
SDValue Off0 = Load0->getOperand(OffIdx0);
SDValue Off1 = Load1->getOperand(OffIdx1);
// The offset might be a FrameIndexSDNode.
if (!isa<ConstantSDNode>(Off0) || !isa<ConstantSDNode>(Off1))
return false;
Offset0 = cast<ConstantSDNode>(Off0)->getZExtValue();
Offset1 = cast<ConstantSDNode>(Off1)->getZExtValue();
return true;
}
return false;
}
static bool isStride64(unsigned Opc) {
switch (Opc) {
case AMDGPU::DS_READ2ST64_B32:
case AMDGPU::DS_READ2ST64_B64:
case AMDGPU::DS_WRITE2ST64_B32:
case AMDGPU::DS_WRITE2ST64_B64:
return true;
default:
return false;
}
}
bool SIInstrInfo::getMemOperandsWithOffsetWidth(
const MachineInstr &LdSt, SmallVectorImpl<const MachineOperand *> &BaseOps,
int64_t &Offset, bool &OffsetIsScalable, unsigned &Width,
const TargetRegisterInfo *TRI) const {
if (!LdSt.mayLoadOrStore())
return false;
unsigned Opc = LdSt.getOpcode();
OffsetIsScalable = false;
const MachineOperand *BaseOp, *OffsetOp;
int DataOpIdx;
if (isDS(LdSt)) {
BaseOp = getNamedOperand(LdSt, AMDGPU::OpName::addr);
OffsetOp = getNamedOperand(LdSt, AMDGPU::OpName::offset);
if (OffsetOp) {
// Normal, single offset LDS instruction.
if (!BaseOp) {
// DS_CONSUME/DS_APPEND use M0 for the base address.
// TODO: find the implicit use operand for M0 and use that as BaseOp?
return false;
}
BaseOps.push_back(BaseOp);
Offset = OffsetOp->getImm();
// Get appropriate operand, and compute width accordingly.
DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
if (DataOpIdx == -1)
DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
Width = getOpSize(LdSt, DataOpIdx);
} else {
// The 2 offset instructions use offset0 and offset1 instead. We can treat
// these as a load with a single offset if the 2 offsets are consecutive.
// We will use this for some partially aligned loads.
const MachineOperand *Offset0Op =
getNamedOperand(LdSt, AMDGPU::OpName::offset0);
const MachineOperand *Offset1Op =
getNamedOperand(LdSt, AMDGPU::OpName::offset1);
unsigned Offset0 = Offset0Op->getImm();
unsigned Offset1 = Offset1Op->getImm();
if (Offset0 + 1 != Offset1)
return false;
// Each of these offsets is in element sized units, so we need to convert
// to bytes of the individual reads.
unsigned EltSize;
if (LdSt.mayLoad())
EltSize = TRI->getRegSizeInBits(*getOpRegClass(LdSt, 0)) / 16;
else {
assert(LdSt.mayStore());
int Data0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
EltSize = TRI->getRegSizeInBits(*getOpRegClass(LdSt, Data0Idx)) / 8;
}
if (isStride64(Opc))
EltSize *= 64;
BaseOps.push_back(BaseOp);
Offset = EltSize * Offset0;
// Get appropriate operand(s), and compute width accordingly.
DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
if (DataOpIdx == -1) {
DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
Width = getOpSize(LdSt, DataOpIdx);
DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data1);
Width += getOpSize(LdSt, DataOpIdx);
} else {
Width = getOpSize(LdSt, DataOpIdx);
}
}
return true;
}
if (isMUBUF(LdSt) || isMTBUF(LdSt)) {
const MachineOperand *RSrc = getNamedOperand(LdSt, AMDGPU::OpName::srsrc);
if (!RSrc) // e.g. BUFFER_WBINVL1_VOL
return false;
BaseOps.push_back(RSrc);
BaseOp = getNamedOperand(LdSt, AMDGPU::OpName::vaddr);
if (BaseOp && !BaseOp->isFI())
BaseOps.push_back(BaseOp);
const MachineOperand *OffsetImm =
getNamedOperand(LdSt, AMDGPU::OpName::offset);
Offset = OffsetImm->getImm();
const MachineOperand *SOffset =
getNamedOperand(LdSt, AMDGPU::OpName::soffset);
if (SOffset) {
if (SOffset->isReg())
BaseOps.push_back(SOffset);
else
Offset += SOffset->getImm();
}
// Get appropriate operand, and compute width accordingly.
DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
if (DataOpIdx == -1)
DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdata);
Width = getOpSize(LdSt, DataOpIdx);
return true;
}
if (isMIMG(LdSt)) {
int SRsrcIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::srsrc);
BaseOps.push_back(&LdSt.getOperand(SRsrcIdx));
int VAddr0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vaddr0);
if (VAddr0Idx >= 0) {
// GFX10 possible NSA encoding.
for (int I = VAddr0Idx; I < SRsrcIdx; ++I)
BaseOps.push_back(&LdSt.getOperand(I));
} else {
BaseOps.push_back(getNamedOperand(LdSt, AMDGPU::OpName::vaddr));
}
Offset = 0;
// Get appropriate operand, and compute width accordingly.
DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdata);
Width = getOpSize(LdSt, DataOpIdx);
return true;
}
if (isSMRD(LdSt)) {
BaseOp = getNamedOperand(LdSt, AMDGPU::OpName::sbase);
if (!BaseOp) // e.g. S_MEMTIME
return false;
BaseOps.push_back(BaseOp);
OffsetOp = getNamedOperand(LdSt, AMDGPU::OpName::offset);
Offset = OffsetOp ? OffsetOp->getImm() : 0;
// Get appropriate operand, and compute width accordingly.
DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::sdst);
Width = getOpSize(LdSt, DataOpIdx);
return true;
}
if (isFLAT(LdSt)) {
// Instructions have either vaddr or saddr or both or none.
BaseOp = getNamedOperand(LdSt, AMDGPU::OpName::vaddr);
if (BaseOp)
BaseOps.push_back(BaseOp);
BaseOp = getNamedOperand(LdSt, AMDGPU::OpName::saddr);
if (BaseOp)
BaseOps.push_back(BaseOp);
Offset = getNamedOperand(LdSt, AMDGPU::OpName::offset)->getImm();
// Get appropriate operand, and compute width accordingly.
DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
if (DataOpIdx == -1)
DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdata);
Width = getOpSize(LdSt, DataOpIdx);
return true;
}
return false;
}
static bool memOpsHaveSameBasePtr(const MachineInstr &MI1,
ArrayRef<const MachineOperand *> BaseOps1,
const MachineInstr &MI2,
ArrayRef<const MachineOperand *> BaseOps2) {
// Only examine the first "base" operand of each instruction, on the
// assumption that it represents the real base address of the memory access.
// Other operands are typically offsets or indices from this base address.
if (BaseOps1.front()->isIdenticalTo(*BaseOps2.front()))
return true;
if (!MI1.hasOneMemOperand() || !MI2.hasOneMemOperand())
return false;
auto MO1 = *MI1.memoperands_begin();
auto MO2 = *MI2.memoperands_begin();
if (MO1->getAddrSpace() != MO2->getAddrSpace())
return false;
auto Base1 = MO1->getValue();
auto Base2 = MO2->getValue();
if (!Base1 || !Base2)
return false;
Base1 = getUnderlyingObject(Base1);
Base2 = getUnderlyingObject(Base2);
if (isa<UndefValue>(Base1) || isa<UndefValue>(Base2))
return false;
return Base1 == Base2;
}
bool SIInstrInfo::shouldClusterMemOps(ArrayRef<const MachineOperand *> BaseOps1,
ArrayRef<const MachineOperand *> BaseOps2,
unsigned NumLoads,
unsigned NumBytes) const {
// If the mem ops (to be clustered) do not have the same base ptr, then they
// should not be clustered
if (!BaseOps1.empty() && !BaseOps2.empty()) {
const MachineInstr &FirstLdSt = *BaseOps1.front()->getParent();
const MachineInstr &SecondLdSt = *BaseOps2.front()->getParent();
if (!memOpsHaveSameBasePtr(FirstLdSt, BaseOps1, SecondLdSt, BaseOps2))
return false;
} else if (!BaseOps1.empty() || !BaseOps2.empty()) {
// If only one base op is empty, they do not have the same base ptr
return false;
}
// In order to avoid regester pressure, on an average, the number of DWORDS
// loaded together by all clustered mem ops should not exceed 8. This is an
// empirical value based on certain observations and performance related
// experiments.
// The good thing about this heuristic is - it avoids clustering of too many
// sub-word loads, and also avoids clustering of wide loads. Below is the
// brief summary of how the heuristic behaves for various `LoadSize`.
// (1) 1 <= LoadSize <= 4: cluster at max 8 mem ops
// (2) 5 <= LoadSize <= 8: cluster at max 4 mem ops
// (3) 9 <= LoadSize <= 12: cluster at max 2 mem ops
// (4) 13 <= LoadSize <= 16: cluster at max 2 mem ops
// (5) LoadSize >= 17: do not cluster
const unsigned LoadSize = NumBytes / NumLoads;
const unsigned NumDWORDs = ((LoadSize + 3) / 4) * NumLoads;
return NumDWORDs <= 8;
}
// FIXME: This behaves strangely. If, for example, you have 32 load + stores,
// the first 16 loads will be interleaved with the stores, and the next 16 will
// be clustered as expected. It should really split into 2 16 store batches.
//
// Loads are clustered until this returns false, rather than trying to schedule
// groups of stores. This also means we have to deal with saying different
// address space loads should be clustered, and ones which might cause bank
// conflicts.
//
// This might be deprecated so it might not be worth that much effort to fix.
bool SIInstrInfo::shouldScheduleLoadsNear(SDNode *Load0, SDNode *Load1,
int64_t Offset0, int64_t Offset1,
unsigned NumLoads) const {
assert(Offset1 > Offset0 &&
"Second offset should be larger than first offset!");
// If we have less than 16 loads in a row, and the offsets are within 64
// bytes, then schedule together.
// A cacheline is 64 bytes (for global memory).
return (NumLoads <= 16 && (Offset1 - Offset0) < 64);
}
static void reportIllegalCopy(const SIInstrInfo *TII, MachineBasicBlock &MBB,
MachineBasicBlock::iterator MI,
const DebugLoc &DL, MCRegister DestReg,
MCRegister SrcReg, bool KillSrc,
const char *Msg = "illegal SGPR to VGPR copy") {
MachineFunction *MF = MBB.getParent();
DiagnosticInfoUnsupported IllegalCopy(MF->getFunction(), Msg, DL, DS_Error);
LLVMContext &C = MF->getFunction().getContext();
C.diagnose(IllegalCopy);
BuildMI(MBB, MI, DL, TII->get(AMDGPU::SI_ILLEGAL_COPY), DestReg)
.addReg(SrcReg, getKillRegState(KillSrc));
}
/// Handle copying from SGPR to AGPR, or from AGPR to AGPR. It is not possible
/// to directly copy, so an intermediate VGPR needs to be used.
static void indirectCopyToAGPR(const SIInstrInfo &TII,
MachineBasicBlock &MBB,
MachineBasicBlock::iterator MI,
const DebugLoc &DL, MCRegister DestReg,
MCRegister SrcReg, bool KillSrc,
RegScavenger &RS,
Register ImpDefSuperReg = Register(),
Register ImpUseSuperReg = Register()) {
const SIRegisterInfo &RI = TII.getRegisterInfo();
assert(AMDGPU::SReg_32RegClass.contains(SrcReg) ||
AMDGPU::AGPR_32RegClass.contains(SrcReg));
// First try to find defining accvgpr_write to avoid temporary registers.
for (auto Def = MI, E = MBB.begin(); Def != E; ) {
--Def;
if (!Def->definesRegister(SrcReg, &RI))
continue;
if (Def->getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64)
break;
MachineOperand &DefOp = Def->getOperand(1);
assert(DefOp.isReg() || DefOp.isImm());
if (DefOp.isReg()) {
// Check that register source operand if not clobbered before MI.
// Immediate operands are always safe to propagate.
bool SafeToPropagate = true;
for (auto I = Def; I != MI && SafeToPropagate; ++I)
if (I->modifiesRegister(DefOp.getReg(), &RI))
SafeToPropagate = false;
if (!SafeToPropagate)
break;
DefOp.setIsKill(false);
}
MachineInstrBuilder Builder =
BuildMI(MBB, MI, DL, TII.get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), DestReg)
.add(DefOp);
if (ImpDefSuperReg)
Builder.addReg(ImpDefSuperReg, RegState::Define | RegState::Implicit);
if (ImpUseSuperReg) {
Builder.addReg(ImpUseSuperReg,
getKillRegState(KillSrc) | RegState::Implicit);
}
return;
}
RS.enterBasicBlock(MBB);
RS.forward(MI);
// Ideally we want to have three registers for a long reg_sequence copy
// to hide 2 waitstates between v_mov_b32 and accvgpr_write.
unsigned MaxVGPRs = RI.getRegPressureLimit(&AMDGPU::VGPR_32RegClass,
*MBB.getParent());
// Registers in the sequence are allocated contiguously so we can just
// use register number to pick one of three round-robin temps.
unsigned RegNo = DestReg % 3;
Register Tmp = RS.scavengeRegister(&AMDGPU::VGPR_32RegClass, 0);
if (!Tmp)
report_fatal_error("Cannot scavenge VGPR to copy to AGPR");
RS.setRegUsed(Tmp);
if (!TII.getSubtarget().hasGFX90AInsts()) {
// Only loop through if there are any free registers left, otherwise
// scavenger may report a fatal error without emergency spill slot
// or spill with the slot.
while (RegNo-- && RS.FindUnusedReg(&AMDGPU::VGPR_32RegClass)) {
Register Tmp2 = RS.scavengeRegister(&AMDGPU::VGPR_32RegClass, 0);
if (!Tmp2 || RI.getHWRegIndex(Tmp2) >= MaxVGPRs)
break;
Tmp = Tmp2;
RS.setRegUsed(Tmp);
}
}
// Insert copy to temporary VGPR.
unsigned TmpCopyOp = AMDGPU::V_MOV_B32_e32;
if (AMDGPU::AGPR_32RegClass.contains(SrcReg)) {
TmpCopyOp = AMDGPU::V_ACCVGPR_READ_B32_e64;
} else {
assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
}
MachineInstrBuilder UseBuilder = BuildMI(MBB, MI, DL, TII.get(TmpCopyOp), Tmp)
.addReg(SrcReg, getKillRegState(KillSrc));
if (ImpUseSuperReg) {
UseBuilder.addReg(ImpUseSuperReg,
getKillRegState(KillSrc) | RegState::Implicit);
}
MachineInstrBuilder DefBuilder
= BuildMI(MBB, MI, DL, TII.get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), DestReg)
.addReg(Tmp, RegState::Kill);
if (ImpDefSuperReg)
DefBuilder.addReg(ImpDefSuperReg, RegState::Define | RegState::Implicit);
}
static void expandSGPRCopy(const SIInstrInfo &TII, MachineBasicBlock &MBB,
MachineBasicBlock::iterator MI, const DebugLoc &DL,
MCRegister DestReg, MCRegister SrcReg, bool KillSrc,
const TargetRegisterClass *RC, bool Forward) {
const SIRegisterInfo &RI = TII.getRegisterInfo();
ArrayRef<int16_t> BaseIndices = RI.getRegSplitParts(RC, 4);
MachineBasicBlock::iterator I = MI;
MachineInstr *FirstMI = nullptr, *LastMI = nullptr;
for (unsigned Idx = 0; Idx < BaseIndices.size(); ++Idx) {
int16_t SubIdx = BaseIndices[Idx];
Register Reg = RI.getSubReg(DestReg, SubIdx);
unsigned Opcode = AMDGPU::S_MOV_B32;
// Is SGPR aligned? If so try to combine with next.
Register Src = RI.getSubReg(SrcReg, SubIdx);
bool AlignedDest = ((Reg - AMDGPU::SGPR0) % 2) == 0;
bool AlignedSrc = ((Src - AMDGPU::SGPR0) % 2) == 0;
if (AlignedDest && AlignedSrc && (Idx + 1 < BaseIndices.size())) {
// Can use SGPR64 copy
unsigned Channel = RI.getChannelFromSubReg(SubIdx);
SubIdx = RI.getSubRegFromChannel(Channel, 2);
Opcode = AMDGPU::S_MOV_B64;
Idx++;
}
LastMI = BuildMI(MBB, I, DL, TII.get(Opcode), RI.getSubReg(DestReg, SubIdx))
.addReg(RI.getSubReg(SrcReg, SubIdx))
.addReg(SrcReg, RegState::Implicit);
if (!FirstMI)
FirstMI = LastMI;
if (!Forward)
I--;
}
assert(FirstMI && LastMI);
if (!Forward)
std::swap(FirstMI, LastMI);
FirstMI->addOperand(
MachineOperand::CreateReg(DestReg, true /*IsDef*/, true /*IsImp*/));
if (KillSrc)
LastMI->addRegisterKilled(SrcReg, &RI);
}
void SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
MachineBasicBlock::iterator MI,
const DebugLoc &DL, MCRegister DestReg,
MCRegister SrcReg, bool KillSrc) const {
const TargetRegisterClass *RC = RI.getPhysRegClass(DestReg);
// FIXME: This is hack to resolve copies between 16 bit and 32 bit
// registers until all patterns are fixed.
if (Fix16BitCopies &&
((RI.getRegSizeInBits(*RC) == 16) ^
(RI.getRegSizeInBits(*RI.getPhysRegClass(SrcReg)) == 16))) {
MCRegister &RegToFix = (RI.getRegSizeInBits(*RC) == 16) ? DestReg : SrcReg;
MCRegister Super = RI.get32BitRegister(RegToFix);
assert(RI.getSubReg(Super, AMDGPU::lo16) == RegToFix);
RegToFix = Super;
if (DestReg == SrcReg) {
// Insert empty bundle since ExpandPostRA expects an instruction here.
BuildMI(MBB, MI, DL, get(AMDGPU::BUNDLE));
return;
}
RC = RI.getPhysRegClass(DestReg);
}
if (RC == &AMDGPU::VGPR_32RegClass) {
assert(AMDGPU::VGPR_32RegClass.contains(SrcReg) ||
AMDGPU::SReg_32RegClass.contains(SrcReg) ||
AMDGPU::AGPR_32RegClass.contains(SrcReg));
unsigned Opc = AMDGPU::AGPR_32RegClass.contains(SrcReg) ?
AMDGPU::V_ACCVGPR_READ_B32_e64 : AMDGPU::V_MOV_B32_e32;
BuildMI(MBB, MI, DL, get(Opc), DestReg)
.addReg(SrcReg, getKillRegState(KillSrc));
return;
}
if (RC == &AMDGPU::SReg_32_XM0RegClass ||
RC == &AMDGPU::SReg_32RegClass) {
if (SrcReg == AMDGPU::SCC) {
BuildMI(MBB, MI, DL, get(AMDGPU::S_CSELECT_B32), DestReg)
.addImm(1)
.addImm(0);
return;
}
if (DestReg == AMDGPU::VCC_LO) {
if (AMDGPU::SReg_32RegClass.contains(SrcReg)) {
BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), AMDGPU::VCC_LO)
.addReg(SrcReg, getKillRegState(KillSrc));
} else {
// FIXME: Hack until VReg_1 removed.
assert(AMDGPU::VGPR_32RegClass.contains(SrcReg));
BuildMI(MBB, MI, DL, get(AMDGPU::V_CMP_NE_U32_e32))
.addImm(0)
.addReg(SrcReg, getKillRegState(KillSrc));
}
return;
}
if (!AMDGPU::SReg_32RegClass.contains(SrcReg)) {
reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc);
return;
}
BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
.addReg(SrcReg, getKillRegState(KillSrc));
return;
}
if (RC == &AMDGPU::SReg_64RegClass) {
if (SrcReg == AMDGPU::SCC) {
BuildMI(MBB, MI, DL, get(AMDGPU::S_CSELECT_B64), DestReg)
.addImm(1)
.addImm(0);
return;
}
if (DestReg == AMDGPU::VCC) {
if (AMDGPU::SReg_64RegClass.contains(SrcReg)) {
BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), AMDGPU::VCC)
.addReg(SrcReg, getKillRegState(KillSrc));
} else {
// FIXME: Hack until VReg_1 removed.
assert(AMDGPU::VGPR_32RegClass.contains(SrcReg));
BuildMI(MBB, MI, DL, get(AMDGPU::V_CMP_NE_U32_e32))
.addImm(0)
.addReg(SrcReg, getKillRegState(KillSrc));
}
return;
}
if (!AMDGPU::SReg_64RegClass.contains(SrcReg)) {
reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc);
return;
}
BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
.addReg(SrcReg, getKillRegState(KillSrc));
return;
}
if (DestReg == AMDGPU::SCC) {
// Copying 64-bit or 32-bit sources to SCC barely makes sense,
// but SelectionDAG emits such copies for i1 sources.
if (AMDGPU::SReg_64RegClass.contains(SrcReg)) {
// This copy can only be produced by patterns
// with explicit SCC, which are known to be enabled
// only for subtargets with S_CMP_LG_U64 present.
assert(ST.hasScalarCompareEq64());
BuildMI(MBB, MI, DL, get(AMDGPU::S_CMP_LG_U64))
.addReg(SrcReg, getKillRegState(KillSrc))
.addImm(0);
} else {
assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
BuildMI(MBB, MI, DL, get(AMDGPU::S_CMP_LG_U32))
.addReg(SrcReg, getKillRegState(KillSrc))
.addImm(0);
}
return;
}
if (RC == &AMDGPU::AGPR_32RegClass) {
if (AMDGPU::VGPR_32RegClass.contains(SrcReg)) {
BuildMI(MBB, MI, DL, get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), DestReg)
.addReg(SrcReg, getKillRegState(KillSrc));
return;
}
if (AMDGPU::AGPR_32RegClass.contains(SrcReg) && ST.hasGFX90AInsts()) {
BuildMI(MBB, MI, DL, get(AMDGPU::V_ACCVGPR_MOV_B32), DestReg)
.addReg(SrcReg, getKillRegState(KillSrc));
return;
}
// FIXME: Pass should maintain scavenger to avoid scan through the block on
// every AGPR spill.
RegScavenger RS;
indirectCopyToAGPR(*this, MBB, MI, DL, DestReg, SrcReg, KillSrc, RS);
return;
}
const unsigned Size = RI.getRegSizeInBits(*RC);
if (Size == 16) {
assert(AMDGPU::VGPR_LO16RegClass.contains(SrcReg) ||
AMDGPU::VGPR_HI16RegClass.contains(SrcReg) ||
AMDGPU::SReg_LO16RegClass.contains(SrcReg) ||
AMDGPU::AGPR_LO16RegClass.contains(SrcReg));
bool IsSGPRDst = AMDGPU::SReg_LO16RegClass.contains(DestReg);
bool IsSGPRSrc = AMDGPU::SReg_LO16RegClass.contains(SrcReg);
bool IsAGPRDst = AMDGPU::AGPR_LO16RegClass.contains(DestReg);
bool IsAGPRSrc = AMDGPU::AGPR_LO16RegClass.contains(SrcReg);
bool DstLow = AMDGPU::VGPR_LO16RegClass.contains(DestReg) ||
AMDGPU::SReg_LO16RegClass.contains(DestReg) ||
AMDGPU::AGPR_LO16RegClass.contains(DestReg);
bool SrcLow = AMDGPU::VGPR_LO16RegClass.contains(SrcReg) ||
AMDGPU::SReg_LO16RegClass.contains(SrcReg) ||
AMDGPU::AGPR_LO16RegClass.contains(SrcReg);
MCRegister NewDestReg = RI.get32BitRegister(DestReg);
MCRegister NewSrcReg = RI.get32BitRegister(SrcReg);
if (IsSGPRDst) {
if (!IsSGPRSrc) {
reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc);
return;
}
BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), NewDestReg)
.addReg(NewSrcReg, getKillRegState(KillSrc));
return;
}
if (IsAGPRDst || IsAGPRSrc) {
if (!DstLow || !SrcLow) {
reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc,
"Cannot use hi16 subreg with an AGPR!");
}
copyPhysReg(MBB, MI, DL, NewDestReg, NewSrcReg, KillSrc);
return;
}
if (IsSGPRSrc && !ST.hasSDWAScalar()) {
if (!DstLow || !SrcLow) {
reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc,
"Cannot use hi16 subreg on VI!");
}
BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), NewDestReg)
.addReg(NewSrcReg, getKillRegState(KillSrc));
return;
}
auto MIB = BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_sdwa), NewDestReg)
.addImm(0) // src0_modifiers
.addReg(NewSrcReg)
.addImm(0) // clamp
.addImm(DstLow ? AMDGPU::SDWA::SdwaSel::WORD_0
: AMDGPU::SDWA::SdwaSel::WORD_1)
.addImm(AMDGPU::SDWA::DstUnused::UNUSED_PRESERVE)
.addImm(SrcLow ? AMDGPU::SDWA::SdwaSel::WORD_0
: AMDGPU::SDWA::SdwaSel::WORD_1)
.addReg(NewDestReg, RegState::Implicit | RegState::Undef);
// First implicit operand is $exec.
MIB->tieOperands(0, MIB->getNumOperands() - 1);
return;
}
const TargetRegisterClass *SrcRC = RI.getPhysRegClass(SrcReg);
if (RC == RI.getVGPR64Class() && (SrcRC == RC || RI.isSGPRClass(SrcRC))) {
if (ST.hasPackedFP32Ops()) {
BuildMI(MBB, MI, DL, get(AMDGPU::V_PK_MOV_B32), DestReg)
.addImm(SISrcMods::OP_SEL_1)
.addReg(SrcReg)
.addImm(SISrcMods::OP_SEL_0 | SISrcMods::OP_SEL_1)
.addReg(SrcReg)
.addImm(0) // op_sel_lo
.addImm(0) // op_sel_hi
.addImm(0) // neg_lo
.addImm(0) // neg_hi
.addImm(0) // clamp
.addReg(SrcReg, getKillRegState(KillSrc) | RegState::Implicit);
return;
}
}
const bool Forward = RI.getHWRegIndex(DestReg) <= RI.getHWRegIndex(SrcReg);
if (RI.isSGPRClass(RC)) {
if (!RI.isSGPRClass(SrcRC)) {
reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc);
return;
}
expandSGPRCopy(*this, MBB, MI, DL, DestReg, SrcReg, KillSrc, RC, Forward);
return;
}
unsigned EltSize = 4;
unsigned Opcode = AMDGPU::V_MOV_B32_e32;
if (RI.isAGPRClass(RC)) {
if (ST.hasGFX90AInsts() && RI.isAGPRClass(SrcRC))
Opcode = AMDGPU::V_ACCVGPR_MOV_B32;
else if (RI.hasVGPRs(SrcRC))
Opcode = AMDGPU::V_ACCVGPR_WRITE_B32_e64;
else
Opcode = AMDGPU::INSTRUCTION_LIST_END;
} else if (RI.hasVGPRs(RC) && RI.isAGPRClass(SrcRC)) {
Opcode = AMDGPU::V_ACCVGPR_READ_B32_e64;
} else if ((Size % 64 == 0) && RI.hasVGPRs(RC) &&
(RI.isProperlyAlignedRC(*RC) &&
(SrcRC == RC || RI.isSGPRClass(SrcRC)))) {
// TODO: In 96-bit case, could do a 64-bit mov and then a 32-bit mov.
if (ST.hasPackedFP32Ops()) {
Opcode = AMDGPU::V_PK_MOV_B32;
EltSize = 8;
}
}
// For the cases where we need an intermediate instruction/temporary register
// (destination is an AGPR), we need a scavenger.
//
// FIXME: The pass should maintain this for us so we don't have to re-scan the
// whole block for every handled copy.
std::unique_ptr<RegScavenger> RS;
if (Opcode == AMDGPU::INSTRUCTION_LIST_END)
RS.reset(new RegScavenger());
ArrayRef<int16_t> SubIndices = RI.getRegSplitParts(RC, EltSize);
// If there is an overlap, we can't kill the super-register on the last
// instruction, since it will also kill the components made live by this def.
const bool CanKillSuperReg = KillSrc && !RI.regsOverlap(SrcReg, DestReg);
for (unsigned Idx = 0; Idx < SubIndices.size(); ++Idx) {
unsigned SubIdx;
if (Forward)
SubIdx = SubIndices[Idx];
else
SubIdx = SubIndices[SubIndices.size() - Idx - 1];
bool UseKill = CanKillSuperReg && Idx == SubIndices.size() - 1;
if (Opcode == AMDGPU::INSTRUCTION_LIST_END) {
Register ImpDefSuper = Idx == 0 ? Register(DestReg) : Register();
Register ImpUseSuper = SrcReg;
indirectCopyToAGPR(*this, MBB, MI, DL, RI.getSubReg(DestReg, SubIdx),
RI.getSubReg(SrcReg, SubIdx), UseKill, *RS,
ImpDefSuper, ImpUseSuper);
} else if (Opcode == AMDGPU::V_PK_MOV_B32) {
Register DstSubReg = RI.getSubReg(DestReg, SubIdx);
Register SrcSubReg = RI.getSubReg(SrcReg, SubIdx);
MachineInstrBuilder MIB =
BuildMI(MBB, MI, DL, get(AMDGPU::V_PK_MOV_B32), DstSubReg)
.addImm(SISrcMods::OP_SEL_1)
.addReg(SrcSubReg)
.addImm(SISrcMods::OP_SEL_0 | SISrcMods::OP_SEL_1)
.addReg(SrcSubReg)
.addImm(0) // op_sel_lo
.addImm(0) // op_sel_hi
.addImm(0) // neg_lo
.addImm(0) // neg_hi
.addImm(0) // clamp
.addReg(SrcReg, getKillRegState(UseKill) | RegState::Implicit);
if (Idx == 0)
MIB.addReg(DestReg, RegState::Define | RegState::Implicit);
} else {
MachineInstrBuilder Builder =
BuildMI(MBB, MI, DL, get(Opcode), RI.getSubReg(DestReg, SubIdx))
.addReg(RI.getSubReg(SrcReg, SubIdx));
if (Idx == 0)
Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
Builder.addReg(SrcReg, getKillRegState(UseKill) | RegState::Implicit);
}
}
}
int SIInstrInfo::commuteOpcode(unsigned Opcode) const {
int NewOpc;
// Try to map original to commuted opcode
NewOpc = AMDGPU::getCommuteRev(Opcode);
if (NewOpc != -1)
// Check if the commuted (REV) opcode exists on the target.
return pseudoToMCOpcode(NewOpc) != -1 ? NewOpc : -1;