llvm.org GIT mirror llvm / master lib / Target / AMDGPU / AMDGPUSubtarget.cpp
master

Tree @master (Download .tar.gz)

AMDGPUSubtarget.cpp @masterraw · history · blame

  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
//===-- AMDGPUSubtarget.cpp - AMDGPU Subtarget 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
/// Implements the AMDGPU specific subclass of TargetSubtarget.
//
//===----------------------------------------------------------------------===//

#include "AMDGPUSubtarget.h"
#include "AMDGPU.h"
#include "AMDGPUTargetMachine.h"
#include "AMDGPUCallLowering.h"
#include "AMDGPUInstructionSelector.h"
#include "AMDGPULegalizerInfo.h"
#include "AMDGPURegisterBankInfo.h"
#include "SIMachineFunctionInfo.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "llvm/ADT/SmallString.h"
#include "llvm/CodeGen/MachineScheduler.h"
#include "llvm/MC/MCSubtargetInfo.h"
#include "llvm/IR/MDBuilder.h"
#include "llvm/CodeGen/TargetFrameLowering.h"
#include <algorithm>

using namespace llvm;

#define DEBUG_TYPE "amdgpu-subtarget"

#define GET_SUBTARGETINFO_TARGET_DESC
#define GET_SUBTARGETINFO_CTOR
#define AMDGPUSubtarget GCNSubtarget
#include "AMDGPUGenSubtargetInfo.inc"
#define GET_SUBTARGETINFO_TARGET_DESC
#define GET_SUBTARGETINFO_CTOR
#undef AMDGPUSubtarget
#include "R600GenSubtargetInfo.inc"

static cl::opt<bool> DisablePowerSched(
  "amdgpu-disable-power-sched",
  cl::desc("Disable scheduling to minimize mAI power bursts"),
  cl::init(false));

GCNSubtarget::~GCNSubtarget() = default;

R600Subtarget &
R600Subtarget::initializeSubtargetDependencies(const Triple &TT,
                                               StringRef GPU, StringRef FS) {
  SmallString<256> FullFS("+promote-alloca,");
  FullFS += FS;
  ParseSubtargetFeatures(GPU, FullFS);

  // FIXME: I don't think think Evergreen has any useful support for
  // denormals, but should be checked. Should we issue a warning somewhere
  // if someone tries to enable these?
  if (getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
    FP32Denormals = false;
  }

  HasMulU24 = getGeneration() >= EVERGREEN;
  HasMulI24 = hasCaymanISA();

  return *this;
}

GCNSubtarget &
GCNSubtarget::initializeSubtargetDependencies(const Triple &TT,
                                              StringRef GPU, StringRef FS) {
  // Determine default and user-specified characteristics
  // On SI+, we want FP64 denormals to be on by default. FP32 denormals can be
  // enabled, but some instructions do not respect them and they run at the
  // double precision rate, so don't enable by default.
  //
  // We want to be able to turn these off, but making this a subtarget feature
  // for SI has the unhelpful behavior that it unsets everything else if you
  // disable it.
  //
  // Similarly we want enable-prt-strict-null to be on by default and not to
  // unset everything else if it is disabled

  // Assuming ECC is enabled is the conservative default.
  SmallString<256> FullFS("+promote-alloca,+load-store-opt,+sram-ecc,+xnack,");

  if (isAmdHsaOS()) // Turn on FlatForGlobal for HSA.
    FullFS += "+flat-for-global,+unaligned-buffer-access,+trap-handler,";

  // FIXME: I don't think think Evergreen has any useful support for
  // denormals, but should be checked. Should we issue a warning somewhere
  // if someone tries to enable these?
  if (getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
    FullFS += "+fp64-fp16-denormals,";
  } else {
    FullFS += "-fp32-denormals,";
  }

  FullFS += "+enable-prt-strict-null,"; // This is overridden by a disable in FS

  // Disable mutually exclusive bits.
  if (FS.find_lower("+wavefrontsize") != StringRef::npos) {
    if (FS.find_lower("wavefrontsize16") == StringRef::npos)
      FullFS += "-wavefrontsize16,";
    if (FS.find_lower("wavefrontsize32") == StringRef::npos)
      FullFS += "-wavefrontsize32,";
    if (FS.find_lower("wavefrontsize64") == StringRef::npos)
      FullFS += "-wavefrontsize64,";
  }

  FullFS += FS;

  ParseSubtargetFeatures(GPU, FullFS);

  // We don't support FP64 for EG/NI atm.
  assert(!hasFP64() || (getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS));

  // Unless +-flat-for-global is specified, turn on FlatForGlobal for all OS-es
  // on VI and newer hardware to avoid assertion failures due to missing ADDR64
  // variants of MUBUF instructions.
  if (!hasAddr64() && !FS.contains("flat-for-global")) {
    FlatForGlobal = true;
  }

  // Set defaults if needed.
  if (MaxPrivateElementSize == 0)
    MaxPrivateElementSize = 4;

  if (LDSBankCount == 0)
    LDSBankCount = 32;

  if (TT.getArch() == Triple::amdgcn) {
    if (LocalMemorySize == 0)
      LocalMemorySize = 32768;

    // Do something sensible for unspecified target.
    if (!HasMovrel && !HasVGPRIndexMode)
      HasMovrel = true;
  }

  // Don't crash on invalid devices.
  if (WavefrontSize == 0)
    WavefrontSize = 64;

  HasFminFmaxLegacy = getGeneration() < AMDGPUSubtarget::VOLCANIC_ISLANDS;

  if (DoesNotSupportXNACK && EnableXNACK) {
    ToggleFeature(AMDGPU::FeatureXNACK);
    EnableXNACK = false;
  }

  // ECC is on by default, but turn it off if the hardware doesn't support it
  // anyway. This matters for the gfx9 targets with d16 loads, but don't support
  // ECC.
  if (DoesNotSupportSRAMECC && EnableSRAMECC) {
    ToggleFeature(AMDGPU::FeatureSRAMECC);
    EnableSRAMECC = false;
  }

  return *this;
}

AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT) :
  TargetTriple(TT),
  Has16BitInsts(false),
  HasMadMixInsts(false),
  FP32Denormals(false),
  FPExceptions(false),
  HasSDWA(false),
  HasVOP3PInsts(false),
  HasMulI24(true),
  HasMulU24(true),
  HasInv2PiInlineImm(false),
  HasFminFmaxLegacy(true),
  EnablePromoteAlloca(false),
  HasTrigReducedRange(false),
  MaxWavesPerEU(10),
  LocalMemorySize(0),
  WavefrontSize(0)
  { }

GCNSubtarget::GCNSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
                           const GCNTargetMachine &TM) :
    AMDGPUGenSubtargetInfo(TT, GPU, FS),
    AMDGPUSubtarget(TT),
    TargetTriple(TT),
    Gen(TT.getOS() == Triple::AMDHSA ? SEA_ISLANDS : SOUTHERN_ISLANDS),
    InstrItins(getInstrItineraryForCPU(GPU)),
    LDSBankCount(0),
    MaxPrivateElementSize(0),

    FastFMAF32(false),
    HalfRate64Ops(false),

    FP64FP16Denormals(false),
    FlatForGlobal(false),
    AutoWaitcntBeforeBarrier(false),
    CodeObjectV3(false),
    UnalignedScratchAccess(false),
    UnalignedBufferAccess(false),

    HasApertureRegs(false),
    EnableXNACK(false),
    DoesNotSupportXNACK(false),
    EnableCuMode(false),
    TrapHandler(false),

    EnableLoadStoreOpt(false),
    EnableUnsafeDSOffsetFolding(false),
    EnableSIScheduler(false),
    EnableDS128(false),
    EnablePRTStrictNull(false),
    DumpCode(false),

    FP64(false),
    GCN3Encoding(false),
    CIInsts(false),
    GFX8Insts(false),
    GFX9Insts(false),
    GFX10Insts(false),
    GFX7GFX8GFX9Insts(false),
    SGPRInitBug(false),
    HasSMemRealTime(false),
    HasIntClamp(false),
    HasFmaMixInsts(false),
    HasMovrel(false),
    HasVGPRIndexMode(false),
    HasScalarStores(false),
    HasScalarAtomics(false),
    HasSDWAOmod(false),
    HasSDWAScalar(false),
    HasSDWASdst(false),
    HasSDWAMac(false),
    HasSDWAOutModsVOPC(false),
    HasDPP(false),
    HasDPP8(false),
    HasR128A16(false),
    HasNSAEncoding(false),
    HasDLInsts(false),
    HasDot1Insts(false),
    HasDot2Insts(false),
    HasDot3Insts(false),
    HasDot4Insts(false),
    HasDot5Insts(false),
    HasDot6Insts(false),
    HasMAIInsts(false),
    HasPkFmacF16Inst(false),
    HasAtomicFaddInsts(false),
    EnableSRAMECC(false),
    DoesNotSupportSRAMECC(false),
    HasNoSdstCMPX(false),
    HasVscnt(false),
    HasRegisterBanking(false),
    HasVOP3Literal(false),
    HasNoDataDepHazard(false),
    FlatAddressSpace(false),
    FlatInstOffsets(false),
    FlatGlobalInsts(false),
    FlatScratchInsts(false),
    ScalarFlatScratchInsts(false),
    AddNoCarryInsts(false),
    HasUnpackedD16VMem(false),
    LDSMisalignedBug(false),
    HasMFMAInlineLiteralBug(false),

    ScalarizeGlobal(false),

    HasVcmpxPermlaneHazard(false),
    HasVMEMtoScalarWriteHazard(false),
    HasSMEMtoVectorWriteHazard(false),
    HasInstFwdPrefetchBug(false),
    HasVcmpxExecWARHazard(false),
    HasLdsBranchVmemWARHazard(false),
    HasNSAtoVMEMBug(false),
    HasOffset3fBug(false),
    HasFlatSegmentOffsetBug(false),

    FeatureDisable(false),
    InstrInfo(initializeSubtargetDependencies(TT, GPU, FS)),
    TLInfo(TM, *this),
    FrameLowering(TargetFrameLowering::StackGrowsUp, getStackAlignment(), 0) {
  MaxWavesPerEU = AMDGPU::IsaInfo::getMaxWavesPerEU(this);
  CallLoweringInfo.reset(new AMDGPUCallLowering(*getTargetLowering()));
  Legalizer.reset(new AMDGPULegalizerInfo(*this, TM));
  RegBankInfo.reset(new AMDGPURegisterBankInfo(*this));
  InstSelector.reset(new AMDGPUInstructionSelector(
  *this, *static_cast<AMDGPURegisterBankInfo *>(RegBankInfo.get()), TM));
}

unsigned GCNSubtarget::getConstantBusLimit(unsigned Opcode) const {
  if (getGeneration() < GFX10)
    return 1;

  switch (Opcode) {
  case AMDGPU::V_LSHLREV_B64:
  case AMDGPU::V_LSHLREV_B64_gfx10:
  case AMDGPU::V_LSHL_B64:
  case AMDGPU::V_LSHRREV_B64:
  case AMDGPU::V_LSHRREV_B64_gfx10:
  case AMDGPU::V_LSHR_B64:
  case AMDGPU::V_ASHRREV_I64:
  case AMDGPU::V_ASHRREV_I64_gfx10:
  case AMDGPU::V_ASHR_I64:
    return 1;
  }

  return 2;
}

unsigned AMDGPUSubtarget::getMaxLocalMemSizeWithWaveCount(unsigned NWaves,
  const Function &F) const {
  if (NWaves == 1)
    return getLocalMemorySize();
  unsigned WorkGroupSize = getFlatWorkGroupSizes(F).second;
  unsigned WorkGroupsPerCu = getMaxWorkGroupsPerCU(WorkGroupSize);
  if (!WorkGroupsPerCu)
    return 0;
  unsigned MaxWaves = getMaxWavesPerEU();
  return getLocalMemorySize() * MaxWaves / WorkGroupsPerCu / NWaves;
}

unsigned AMDGPUSubtarget::getOccupancyWithLocalMemSize(uint32_t Bytes,
  const Function &F) const {
  unsigned WorkGroupSize = getFlatWorkGroupSizes(F).second;
  unsigned WorkGroupsPerCu = getMaxWorkGroupsPerCU(WorkGroupSize);
  if (!WorkGroupsPerCu)
    return 0;
  unsigned MaxWaves = getMaxWavesPerEU();
  unsigned Limit = getLocalMemorySize() * MaxWaves / WorkGroupsPerCu;
  unsigned NumWaves = Limit / (Bytes ? Bytes : 1u);
  NumWaves = std::min(NumWaves, MaxWaves);
  NumWaves = std::max(NumWaves, 1u);
  return NumWaves;
}

unsigned
AMDGPUSubtarget::getOccupancyWithLocalMemSize(const MachineFunction &MF) const {
  const auto *MFI = MF.getInfo<SIMachineFunctionInfo>();
  return getOccupancyWithLocalMemSize(MFI->getLDSSize(), MF.getFunction());
}

std::pair<unsigned, unsigned>
AMDGPUSubtarget::getDefaultFlatWorkGroupSize(CallingConv::ID CC) const {
  switch (CC) {
  case CallingConv::AMDGPU_CS:
  case CallingConv::AMDGPU_KERNEL:
  case CallingConv::SPIR_KERNEL:
    return std::make_pair(getWavefrontSize() * 2,
                          std::max(getWavefrontSize() * 4, 256u));
  case CallingConv::AMDGPU_VS:
  case CallingConv::AMDGPU_LS:
  case CallingConv::AMDGPU_HS:
  case CallingConv::AMDGPU_ES:
  case CallingConv::AMDGPU_GS:
  case CallingConv::AMDGPU_PS:
    return std::make_pair(1, getWavefrontSize());
  default:
    return std::make_pair(1, 16 * getWavefrontSize());
  }
}

std::pair<unsigned, unsigned> AMDGPUSubtarget::getFlatWorkGroupSizes(
  const Function &F) const {
  // FIXME: 1024 if function.
  // Default minimum/maximum flat work group sizes.
  std::pair<unsigned, unsigned> Default =
    getDefaultFlatWorkGroupSize(F.getCallingConv());

  // Requested minimum/maximum flat work group sizes.
  std::pair<unsigned, unsigned> Requested = AMDGPU::getIntegerPairAttribute(
    F, "amdgpu-flat-work-group-size", Default);

  // Make sure requested minimum is less than requested maximum.
  if (Requested.first > Requested.second)
    return Default;

  // Make sure requested values do not violate subtarget's specifications.
  if (Requested.first < getMinFlatWorkGroupSize())
    return Default;
  if (Requested.second > getMaxFlatWorkGroupSize())
    return Default;

  return Requested;
}

std::pair<unsigned, unsigned> AMDGPUSubtarget::getWavesPerEU(
  const Function &F) const {
  // Default minimum/maximum number of waves per execution unit.
  std::pair<unsigned, unsigned> Default(1, getMaxWavesPerEU());

  // Default/requested minimum/maximum flat work group sizes.
  std::pair<unsigned, unsigned> FlatWorkGroupSizes = getFlatWorkGroupSizes(F);

  // If minimum/maximum flat work group sizes were explicitly requested using
  // "amdgpu-flat-work-group-size" attribute, then set default minimum/maximum
  // number of waves per execution unit to values implied by requested
  // minimum/maximum flat work group sizes.
  unsigned MinImpliedByFlatWorkGroupSize =
    getMaxWavesPerEU(FlatWorkGroupSizes.second);
  bool RequestedFlatWorkGroupSize = false;

  if (F.hasFnAttribute("amdgpu-flat-work-group-size")) {
    Default.first = MinImpliedByFlatWorkGroupSize;
    RequestedFlatWorkGroupSize = true;
  }

  // Requested minimum/maximum number of waves per execution unit.
  std::pair<unsigned, unsigned> Requested = AMDGPU::getIntegerPairAttribute(
    F, "amdgpu-waves-per-eu", Default, true);

  // Make sure requested minimum is less than requested maximum.
  if (Requested.second && Requested.first > Requested.second)
    return Default;

  // Make sure requested values do not violate subtarget's specifications.
  if (Requested.first < getMinWavesPerEU() ||
      Requested.first > getMaxWavesPerEU())
    return Default;
  if (Requested.second > getMaxWavesPerEU())
    return Default;

  // Make sure requested values are compatible with values implied by requested
  // minimum/maximum flat work group sizes.
  if (RequestedFlatWorkGroupSize &&
      Requested.first < MinImpliedByFlatWorkGroupSize)
    return Default;

  return Requested;
}

bool AMDGPUSubtarget::makeLIDRangeMetadata(Instruction *I) const {
  Function *Kernel = I->getParent()->getParent();
  unsigned MinSize = 0;
  unsigned MaxSize = getFlatWorkGroupSizes(*Kernel).second;
  bool IdQuery = false;

  // If reqd_work_group_size is present it narrows value down.
  if (auto *CI = dyn_cast<CallInst>(I)) {
    const Function *F = CI->getCalledFunction();
    if (F) {
      unsigned Dim = UINT_MAX;
      switch (F->getIntrinsicID()) {
      case Intrinsic::amdgcn_workitem_id_x:
      case Intrinsic::r600_read_tidig_x:
        IdQuery = true;
        LLVM_FALLTHROUGH;
      case Intrinsic::r600_read_local_size_x:
        Dim = 0;
        break;
      case Intrinsic::amdgcn_workitem_id_y:
      case Intrinsic::r600_read_tidig_y:
        IdQuery = true;
        LLVM_FALLTHROUGH;
      case Intrinsic::r600_read_local_size_y:
        Dim = 1;
        break;
      case Intrinsic::amdgcn_workitem_id_z:
      case Intrinsic::r600_read_tidig_z:
        IdQuery = true;
        LLVM_FALLTHROUGH;
      case Intrinsic::r600_read_local_size_z:
        Dim = 2;
        break;
      default:
        break;
      }
      if (Dim <= 3) {
        if (auto Node = Kernel->getMetadata("reqd_work_group_size"))
          if (Node->getNumOperands() == 3)
            MinSize = MaxSize = mdconst::extract<ConstantInt>(
                                  Node->getOperand(Dim))->getZExtValue();
      }
    }
  }

  if (!MaxSize)
    return false;

  // Range metadata is [Lo, Hi). For ID query we need to pass max size
  // as Hi. For size query we need to pass Hi + 1.
  if (IdQuery)
    MinSize = 0;
  else
    ++MaxSize;

  MDBuilder MDB(I->getContext());
  MDNode *MaxWorkGroupSizeRange = MDB.createRange(APInt(32, MinSize),
                                                  APInt(32, MaxSize));
  I->setMetadata(LLVMContext::MD_range, MaxWorkGroupSizeRange);
  return true;
}

uint64_t AMDGPUSubtarget::getExplicitKernArgSize(const Function &F,
                                                 unsigned &MaxAlign) const {
  assert(F.getCallingConv() == CallingConv::AMDGPU_KERNEL ||
         F.getCallingConv() == CallingConv::SPIR_KERNEL);

  const DataLayout &DL = F.getParent()->getDataLayout();
  uint64_t ExplicitArgBytes = 0;
  MaxAlign = 1;

  for (const Argument &Arg : F.args()) {
    Type *ArgTy = Arg.getType();

    unsigned Align = DL.getABITypeAlignment(ArgTy);
    uint64_t AllocSize = DL.getTypeAllocSize(ArgTy);
    ExplicitArgBytes = alignTo(ExplicitArgBytes, Align) + AllocSize;
    MaxAlign = std::max(MaxAlign, Align);
  }

  return ExplicitArgBytes;
}

unsigned AMDGPUSubtarget::getKernArgSegmentSize(const Function &F,
                                                unsigned &MaxAlign) const {
  uint64_t ExplicitArgBytes = getExplicitKernArgSize(F, MaxAlign);

  unsigned ExplicitOffset = getExplicitKernelArgOffset(F);

  uint64_t TotalSize = ExplicitOffset + ExplicitArgBytes;
  unsigned ImplicitBytes = getImplicitArgNumBytes(F);
  if (ImplicitBytes != 0) {
    unsigned Alignment = getAlignmentForImplicitArgPtr();
    TotalSize = alignTo(ExplicitArgBytes, Alignment) + ImplicitBytes;
  }

  // Being able to dereference past the end is useful for emitting scalar loads.
  return alignTo(TotalSize, 4);
}

R600Subtarget::R600Subtarget(const Triple &TT, StringRef GPU, StringRef FS,
                             const TargetMachine &TM) :
  R600GenSubtargetInfo(TT, GPU, FS),
  AMDGPUSubtarget(TT),
  InstrInfo(*this),
  FrameLowering(TargetFrameLowering::StackGrowsUp, getStackAlignment(), 0),
  FMA(false),
  CaymanISA(false),
  CFALUBug(false),
  HasVertexCache(false),
  R600ALUInst(false),
  FP64(false),
  TexVTXClauseSize(0),
  Gen(R600),
  TLInfo(TM, initializeSubtargetDependencies(TT, GPU, FS)),
  InstrItins(getInstrItineraryForCPU(GPU)) { }

void GCNSubtarget::overrideSchedPolicy(MachineSchedPolicy &Policy,
                                      unsigned NumRegionInstrs) const {
  // Track register pressure so the scheduler can try to decrease
  // pressure once register usage is above the threshold defined by
  // SIRegisterInfo::getRegPressureSetLimit()
  Policy.ShouldTrackPressure = true;

  // Enabling both top down and bottom up scheduling seems to give us less
  // register spills than just using one of these approaches on its own.
  Policy.OnlyTopDown = false;
  Policy.OnlyBottomUp = false;

  // Enabling ShouldTrackLaneMasks crashes the SI Machine Scheduler.
  if (!enableSIScheduler())
    Policy.ShouldTrackLaneMasks = true;
}

bool GCNSubtarget::hasMadF16() const {
  return InstrInfo.pseudoToMCOpcode(AMDGPU::V_MAD_F16) != -1;
}

unsigned GCNSubtarget::getOccupancyWithNumSGPRs(unsigned SGPRs) const {
  if (getGeneration() >= AMDGPUSubtarget::GFX10)
    return getMaxWavesPerEU();

  if (getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
    if (SGPRs <= 80)
      return 10;
    if (SGPRs <= 88)
      return 9;
    if (SGPRs <= 100)
      return 8;
    return 7;
  }
  if (SGPRs <= 48)
    return 10;
  if (SGPRs <= 56)
    return 9;
  if (SGPRs <= 64)
    return 8;
  if (SGPRs <= 72)
    return 7;
  if (SGPRs <= 80)
    return 6;
  return 5;
}

unsigned GCNSubtarget::getOccupancyWithNumVGPRs(unsigned VGPRs) const {
  unsigned MaxWaves = getMaxWavesPerEU();
  unsigned Granule = getVGPRAllocGranule();
  if (VGPRs < Granule)
    return MaxWaves;
  unsigned RoundedRegs = ((VGPRs + Granule - 1) / Granule) * Granule;
  return std::min(std::max(getTotalNumVGPRs() / RoundedRegs, 1u), MaxWaves);
}

unsigned GCNSubtarget::getReservedNumSGPRs(const MachineFunction &MF) const {
  const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();
  if (getGeneration() >= AMDGPUSubtarget::GFX10)
    return 2; // VCC. FLAT_SCRATCH and XNACK are no longer in SGPRs.

  if (MFI.hasFlatScratchInit()) {
    if (getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
      return 6; // FLAT_SCRATCH, XNACK, VCC (in that order).
    if (getGeneration() == AMDGPUSubtarget::SEA_ISLANDS)
      return 4; // FLAT_SCRATCH, VCC (in that order).
  }

  if (isXNACKEnabled())
    return 4; // XNACK, VCC (in that order).
  return 2; // VCC.
}

unsigned GCNSubtarget::computeOccupancy(const MachineFunction &MF,
                                        unsigned LDSSize,
                                        unsigned NumSGPRs,
                                        unsigned NumVGPRs) const {
  unsigned Occupancy =
    std::min(getMaxWavesPerEU(),
             getOccupancyWithLocalMemSize(LDSSize, MF.getFunction()));
  if (NumSGPRs)
    Occupancy = std::min(Occupancy, getOccupancyWithNumSGPRs(NumSGPRs));
  if (NumVGPRs)
    Occupancy = std::min(Occupancy, getOccupancyWithNumVGPRs(NumVGPRs));
  return Occupancy;
}

unsigned GCNSubtarget::getMaxNumSGPRs(const MachineFunction &MF) const {
  const Function &F = MF.getFunction();
  const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();

  // Compute maximum number of SGPRs function can use using default/requested
  // minimum number of waves per execution unit.
  std::pair<unsigned, unsigned> WavesPerEU = MFI.getWavesPerEU();
  unsigned MaxNumSGPRs = getMaxNumSGPRs(WavesPerEU.first, false);
  unsigned MaxAddressableNumSGPRs = getMaxNumSGPRs(WavesPerEU.first, true);

  // Check if maximum number of SGPRs was explicitly requested using
  // "amdgpu-num-sgpr" attribute.
  if (F.hasFnAttribute("amdgpu-num-sgpr")) {
    unsigned Requested = AMDGPU::getIntegerAttribute(
      F, "amdgpu-num-sgpr", MaxNumSGPRs);

    // Make sure requested value does not violate subtarget's specifications.
    if (Requested && (Requested <= getReservedNumSGPRs(MF)))
      Requested = 0;

    // If more SGPRs are required to support the input user/system SGPRs,
    // increase to accommodate them.
    //
    // FIXME: This really ends up using the requested number of SGPRs + number
    // of reserved special registers in total. Theoretically you could re-use
    // the last input registers for these special registers, but this would
    // require a lot of complexity to deal with the weird aliasing.
    unsigned InputNumSGPRs = MFI.getNumPreloadedSGPRs();
    if (Requested && Requested < InputNumSGPRs)
      Requested = InputNumSGPRs;

    // Make sure requested value is compatible with values implied by
    // default/requested minimum/maximum number of waves per execution unit.
    if (Requested && Requested > getMaxNumSGPRs(WavesPerEU.first, false))
      Requested = 0;
    if (WavesPerEU.second &&
        Requested && Requested < getMinNumSGPRs(WavesPerEU.second))
      Requested = 0;

    if (Requested)
      MaxNumSGPRs = Requested;
  }

  if (hasSGPRInitBug())
    MaxNumSGPRs = AMDGPU::IsaInfo::FIXED_NUM_SGPRS_FOR_INIT_BUG;

  return std::min(MaxNumSGPRs - getReservedNumSGPRs(MF),
                  MaxAddressableNumSGPRs);
}

unsigned GCNSubtarget::getMaxNumVGPRs(const MachineFunction &MF) const {
  const Function &F = MF.getFunction();
  const SIMachineFunctionInfo &MFI = *MF.getInfo<SIMachineFunctionInfo>();

  // Compute maximum number of VGPRs function can use using default/requested
  // minimum number of waves per execution unit.
  std::pair<unsigned, unsigned> WavesPerEU = MFI.getWavesPerEU();
  unsigned MaxNumVGPRs = getMaxNumVGPRs(WavesPerEU.first);

  // Check if maximum number of VGPRs was explicitly requested using
  // "amdgpu-num-vgpr" attribute.
  if (F.hasFnAttribute("amdgpu-num-vgpr")) {
    unsigned Requested = AMDGPU::getIntegerAttribute(
      F, "amdgpu-num-vgpr", MaxNumVGPRs);

    // Make sure requested value is compatible with values implied by
    // default/requested minimum/maximum number of waves per execution unit.
    if (Requested && Requested > getMaxNumVGPRs(WavesPerEU.first))
      Requested = 0;
    if (WavesPerEU.second &&
        Requested && Requested < getMinNumVGPRs(WavesPerEU.second))
      Requested = 0;

    if (Requested)
      MaxNumVGPRs = Requested;
  }

  return MaxNumVGPRs;
}

namespace {
struct MemOpClusterMutation : ScheduleDAGMutation {
  const SIInstrInfo *TII;

  MemOpClusterMutation(const SIInstrInfo *tii) : TII(tii) {}

  void apply(ScheduleDAGInstrs *DAG) override {
    SUnit *SUa = nullptr;
    // Search for two consequent memory operations and link them
    // to prevent scheduler from moving them apart.
    // In DAG pre-process SUnits are in the original order of
    // the instructions before scheduling.
    for (SUnit &SU : DAG->SUnits) {
      MachineInstr &MI2 = *SU.getInstr();
      if (!MI2.mayLoad() && !MI2.mayStore()) {
        SUa = nullptr;
        continue;
      }
      if (!SUa) {
        SUa = &SU;
        continue;
      }

      MachineInstr &MI1 = *SUa->getInstr();
      if ((TII->isVMEM(MI1) && TII->isVMEM(MI2)) ||
          (TII->isFLAT(MI1) && TII->isFLAT(MI2)) ||
          (TII->isSMRD(MI1) && TII->isSMRD(MI2)) ||
          (TII->isDS(MI1)   && TII->isDS(MI2))) {
        SU.addPredBarrier(SUa);

        for (const SDep &SI : SU.Preds) {
          if (SI.getSUnit() != SUa)
            SUa->addPred(SDep(SI.getSUnit(), SDep::Artificial));
        }

        if (&SU != &DAG->ExitSU) {
          for (const SDep &SI : SUa->Succs) {
            if (SI.getSUnit() != &SU)
              SI.getSUnit()->addPred(SDep(&SU, SDep::Artificial));
          }
        }
      }

      SUa = &SU;
    }
  }
};

struct FillMFMAShadowMutation : ScheduleDAGMutation {
  const SIInstrInfo *TII;

  ScheduleDAGMI *DAG;

  FillMFMAShadowMutation(const SIInstrInfo *tii) : TII(tii) {}

  bool isSALU(const SUnit *SU) const {
    const MachineInstr *MI = SU->getInstr();
    return MI && TII->isSALU(*MI) && !MI->isTerminator();
  }

  bool canAddEdge(const SUnit *Succ, const SUnit *Pred) const {
    if (Pred->NodeNum < Succ->NodeNum)
      return true;

    SmallVector<const SUnit*, 64> Succs({Succ}), Preds({Pred});

    for (unsigned I = 0; I < Succs.size(); ++I) {
      for (const SDep &SI : Succs[I]->Succs) {
        const SUnit *SU = SI.getSUnit();
        if (SU != Succs[I] && llvm::find(Succs, SU) == Succs.end())
          Succs.push_back(SU);
      }
    }

    SmallPtrSet<const SUnit*, 32> Visited;
    while (!Preds.empty()) {
      const SUnit *SU = Preds.pop_back_val();
      if (llvm::find(Succs, SU) != Succs.end())
        return false;
      Visited.insert(SU);
      for (const SDep &SI : SU->Preds)
        if (SI.getSUnit() != SU && !Visited.count(SI.getSUnit()))
          Preds.push_back(SI.getSUnit());
    }

    return true;
  }

  // Link as much SALU intructions in chain as possible. Return the size
  // of the chain. Links up to MaxChain instructions.
  unsigned linkSALUChain(SUnit *From, SUnit *To, unsigned MaxChain,
                         SmallPtrSetImpl<SUnit *> &Visited) const {
    SmallVector<SUnit *, 8> Worklist({To});
    unsigned Linked = 0;

    while (!Worklist.empty() && MaxChain-- > 0) {
      SUnit *SU = Worklist.pop_back_val();
      if (!Visited.insert(SU).second)
        continue;

      LLVM_DEBUG(dbgs() << "Inserting edge from\n" ; DAG->dumpNode(*From);
                 dbgs() << "to\n"; DAG->dumpNode(*SU); dbgs() << '\n');

      if (SU->addPred(SDep(From, SDep::Artificial), false))
        ++Linked;

      for (SDep &SI : From->Succs) {
        SUnit *SUv = SI.getSUnit();
        if (SUv != From && TII->isVALU(*SUv->getInstr()) && canAddEdge(SUv, SU))
          SUv->addPred(SDep(SU, SDep::Artificial), false);
      }

      for (SDep &SI : SU->Succs) {
        SUnit *Succ = SI.getSUnit();
        if (Succ != SU && isSALU(Succ) && canAddEdge(From, Succ))
          Worklist.push_back(Succ);
      }
    }

    return Linked;
  }

  void apply(ScheduleDAGInstrs *DAGInstrs) override {
    const GCNSubtarget &ST = DAGInstrs->MF.getSubtarget<GCNSubtarget>();
    if (!ST.hasMAIInsts() || DisablePowerSched)
      return;
    DAG = static_cast<ScheduleDAGMI*>(DAGInstrs);
    const TargetSchedModel *TSchedModel = DAGInstrs->getSchedModel();
    if (!TSchedModel || DAG->SUnits.empty())
      return;

    // Scan for MFMA long latency instructions and try to add a dependency
    // of available SALU instructions to give them a chance to fill MFMA
    // shadow. That is desirable to fill MFMA shadow with SALU instructions
    // rather than VALU to prevent power consumption bursts and throttle.
    auto LastSALU = DAG->SUnits.begin();
    auto E = DAG->SUnits.end();
    SmallPtrSet<SUnit*, 32> Visited;
    for (SUnit &SU : DAG->SUnits) {
      MachineInstr &MAI = *SU.getInstr();
      if (!TII->isMAI(MAI) ||
           MAI.getOpcode() == AMDGPU::V_ACCVGPR_WRITE_B32 ||
           MAI.getOpcode() == AMDGPU::V_ACCVGPR_READ_B32)
        continue;

      unsigned Lat = TSchedModel->computeInstrLatency(&MAI) - 1;

      LLVM_DEBUG(dbgs() << "Found MFMA: "; DAG->dumpNode(SU);
                 dbgs() << "Need " << Lat
                        << " instructions to cover latency.\n");

      // Find up to Lat independent scalar instructions as early as
      // possible such that they can be scheduled after this MFMA.
      for ( ; Lat && LastSALU != E; ++LastSALU) {
        if (Visited.count(&*LastSALU))
          continue;

        if (!isSALU(&*LastSALU) || !canAddEdge(&*LastSALU, &SU))
          continue;

        Lat -= linkSALUChain(&SU, &*LastSALU, Lat, Visited);
      }
    }
  }
};
} // namespace

void GCNSubtarget::getPostRAMutations(
    std::vector<std::unique_ptr<ScheduleDAGMutation>> &Mutations) const {
  Mutations.push_back(std::make_unique<MemOpClusterMutation>(&InstrInfo));
  Mutations.push_back(std::make_unique<FillMFMAShadowMutation>(&InstrInfo));
}

const AMDGPUSubtarget &AMDGPUSubtarget::get(const MachineFunction &MF) {
  if (MF.getTarget().getTargetTriple().getArch() == Triple::amdgcn)
    return static_cast<const AMDGPUSubtarget&>(MF.getSubtarget<GCNSubtarget>());
  else
    return static_cast<const AMDGPUSubtarget&>(MF.getSubtarget<R600Subtarget>());
}

const AMDGPUSubtarget &AMDGPUSubtarget::get(const TargetMachine &TM, const Function &F) {
  if (TM.getTargetTriple().getArch() == Triple::amdgcn)
    return static_cast<const AMDGPUSubtarget&>(TM.getSubtarget<GCNSubtarget>(F));
  else
    return static_cast<const AMDGPUSubtarget&>(TM.getSubtarget<R600Subtarget>(F));
}