LLVM: lib/Target/AArch64/AArch64TargetTransformInfo.cpp Source File (original) (raw)

1

2

3

4

5

6

7

8

23#include "llvm/IR/IntrinsicsAArch64.h"

29#include

30#include

31using namespace llvm;

33

34#define DEBUG_TYPE "aarch64tti"

35

38

40 "sve-prefer-fixed-over-scalable-if-equal", cl::Hidden);

41

44

47

50

54

58 "Penalty of calling a function that requires a change to PSTATE.SM"));

59

62 cl::desc("Penalty of inlining a call that requires a change to PSTATE.SM"));

63

66

69

70

73 cl::desc("The cost of a histcnt instruction"));

74

77 cl::desc("The number of instructions to search for a redundant dmb"));

78

79namespace {

80class TailFoldingOption {

81

82

83

84

85

86

87

88

92

93

94

95 bool NeedsDefault = true;

96

97 void setInitialBits(TailFoldingOpts Bits) { InitialBits = Bits; }

98

99 void setNeedsDefault(bool V) { NeedsDefault = V; }

100

102 EnableBits |= Bit;

103 DisableBits &= ~Bit;

104 }

105

107 EnableBits &= ~Bit;

108 DisableBits |= Bit;

109 }

110

113

114 assert((InitialBits == TailFoldingOpts::Disabled || !NeedsDefault) &&

115 "Initial bits should only include one of "

116 "(disabled|all|simple|default)");

117 Bits = NeedsDefault ? DefaultBits : InitialBits;

118 Bits |= EnableBits;

119 Bits &= ~DisableBits;

120

122 }

123

125 errs() << "invalid argument '" << Opt

126 << "' to -sve-tail-folding=; the option should be of the form\n"

127 " (disabled|all|default|simple)[+(reductions|recurrences"

128 "|reverse|noreductions|norecurrences|noreverse)]\n";

130 }

131

132public:

133

134 void operator=(const std::string &Val) {

135

136 if (Val.empty()) {

138 return;

139 }

140

141

142

143 setNeedsDefault(false);

144

147

148 unsigned StartIdx = 1;

149 if (TailFoldTypes[0] == "disabled")

150 setInitialBits(TailFoldingOpts::Disabled);

151 else if (TailFoldTypes[0] == "all")

152 setInitialBits(TailFoldingOpts::All);

153 else if (TailFoldTypes[0] == "default")

154 setNeedsDefault(true);

155 else if (TailFoldTypes[0] == "simple")

156 setInitialBits(TailFoldingOpts::Simple);

157 else {

158 StartIdx = 0;

159 setInitialBits(TailFoldingOpts::Disabled);

160 }

161

162 for (unsigned I = StartIdx; I < TailFoldTypes.size(); I++) {

163 if (TailFoldTypes[I] == "reductions")

164 setEnableBit(TailFoldingOpts::Reductions);

165 else if (TailFoldTypes[I] == "recurrences")

166 setEnableBit(TailFoldingOpts::Recurrences);

167 else if (TailFoldTypes[I] == "reverse")

168 setEnableBit(TailFoldingOpts::Reverse);

169 else if (TailFoldTypes[I] == "noreductions")

170 setDisableBit(TailFoldingOpts::Reductions);

171 else if (TailFoldTypes[I] == "norecurrences")

172 setDisableBit(TailFoldingOpts::Recurrences);

173 else if (TailFoldTypes[I] == "noreverse")

174 setDisableBit(TailFoldingOpts::Reverse);

175 else

177 }

178 }

179

182 }

183};

184}

185

187

189 "sve-tail-folding",

191 "Control the use of vectorisation using tail-folding for SVE where the"

192 " option is specified in the form (Initial)[+(Flag1|Flag2|...)]:"

193 "\ndisabled (Initial) No loop types will vectorize using "

194 "tail-folding"

195 "\ndefault (Initial) Uses the default tail-folding settings for "

196 "the target CPU"

197 "\nall (Initial) All legal loop types will vectorize using "

198 "tail-folding"

199 "\nsimple (Initial) Use tail-folding for simple loops (not "

200 "reductions or recurrences)"

201 "\nreductions Use tail-folding for loops containing reductions"

202 "\nnoreductions Inverse of above"

203 "\nrecurrences Use tail-folding for loops containing fixed order "

204 "recurrences"

205 "\nnorecurrences Inverse of above"

206 "\nreverse Use tail-folding for loops requiring reversed "

207 "predicates"

208 "\nnoreverse Inverse of above"),

210

211

212

213

215 "enable-fixedwidth-autovec-in-streaming-mode", cl::init(false), cl::Hidden);

216

217

218

219

221 "enable-scalable-autovec-in-streaming-mode", cl::init(false), cl::Hidden);

222

226 .Case("__arm_sme_state", true)

227 .Case("__arm_tpidr2_save", true)

228 .Case("__arm_tpidr2_restore", true)

229 .Case("__arm_za_disable", true)

231}

232

233

234

235

239

240

241

242

243 if (isa(I) && I.isDebugOrPseudoInst() &&

244 (cast(I).isInlineAsm() || isa(I) ||

246 return true;

247 }

248 }

249 return false;

250}

251

255 StringRef FeatureStr = F.getFnAttribute(AttributeStr).getValueAsString();

257 FeatureStr.split(Features, ",");

259}

260

262 return F.hasFnAttribute("fmv-features");

263}

264

266 const Function *Callee) const {

267 SMEAttrs CallerAttrs(*Caller), CalleeAttrs(*Callee);

268

269

270

274 }

275

277 return false;

278

279 if (CallerAttrs.requiresLazySave(CalleeAttrs) ||

280 CallerAttrs.requiresSMChange(CalleeAttrs) ||

281 CallerAttrs.requiresPreservingZT0(CalleeAttrs) ||

282 CallerAttrs.requiresPreservingAllZAState(CalleeAttrs)) {

284 return false;

285 }

286

288}

289

294 return false;

295

296

297

298

299

300

301

302

303

305 auto FVTy = dyn_cast(Ty);

306 return FVTy &&

307 FVTy->getScalarSizeInBits() * FVTy->getNumElements() > 128;

308 }))

309 return false;

310

311 return true;

312}

313

314unsigned

316 unsigned DefaultCallPenalty) const {

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

339 if (F == Call.getCaller())

343 }

344

345 return DefaultCallPenalty;

346}

347

353}

354

355

356

357

359

361 return 0;

362

363 if (Val < 0)

364 Val = ~Val;

365

366

369 return Insn.size();

370}

371

372

376

378 if (BitSize == 0)

379 return ~0U;

380

381

382 APInt ImmVal = Imm;

383 if (BitSize & 0x3f)

384 ImmVal = Imm.sext((BitSize + 63) & ~0x3fU);

385

386

387

389 for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {

393 }

394

395 return std::max(1, Cost);

396}

397

403

405

406

407 if (BitSize == 0)

409

410 unsigned ImmIdx = ~0U;

411 switch (Opcode) {

412 default:

414 case Instruction::GetElementPtr:

415

416 if (Idx == 0)

419 case Instruction::Store:

420 ImmIdx = 0;

421 break;

422 case Instruction::Add:

423 case Instruction::Sub:

424 case Instruction::Mul:

425 case Instruction::UDiv:

426 case Instruction::SDiv:

427 case Instruction::URem:

428 case Instruction::SRem:

429 case Instruction::And:

430 case Instruction::Or:

431 case Instruction::Xor:

432 case Instruction::ICmp:

433 ImmIdx = 1;

434 break;

435

436 case Instruction::Shl:

437 case Instruction::LShr:

438 case Instruction::AShr:

439 if (Idx == 1)

441 break;

442 case Instruction::Trunc:

443 case Instruction::ZExt:

444 case Instruction::SExt:

445 case Instruction::IntToPtr:

446 case Instruction::PtrToInt:

447 case Instruction::BitCast:

448 case Instruction::PHI:

449 case Instruction::Call:

450 case Instruction::Select:

451 case Instruction::Ret:

452 case Instruction::Load:

453 break;

454 }

455

456 if (Idx == ImmIdx) {

457 int NumConstants = (BitSize + 63) / 64;

462 }

464}

465

471

473

474

475 if (BitSize == 0)

477

478

479

480

481 if (IID >= Intrinsic::aarch64_addg && IID <= Intrinsic::aarch64_udiv)

483

484 switch (IID) {

485 default:

487 case Intrinsic::sadd_with_overflow:

488 case Intrinsic::uadd_with_overflow:

489 case Intrinsic::ssub_with_overflow:

490 case Intrinsic::usub_with_overflow:

491 case Intrinsic::smul_with_overflow:

492 case Intrinsic::umul_with_overflow:

493 if (Idx == 1) {

494 int NumConstants = (BitSize + 63) / 64;

499 }

500 break;

501 case Intrinsic::experimental_stackmap:

502 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))

504 break;

505 case Intrinsic::experimental_patchpoint_void:

506 case Intrinsic::experimental_patchpoint:

507 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))

509 break;

510 case Intrinsic::experimental_gc_statepoint:

511 if ((Idx < 5) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))

513 break;

514 }

516}

517

521 if (TyWidth == 32 || TyWidth == 64)

523

525}

526

530}

531

533 Type *BucketPtrsTy = ICA.getArgTypes()[0];

534 Type *EltTy = ICA.getArgTypes()[1];

535 unsigned TotalHistCnts = 1;

536

538

541

542

543

544 if (VectorType *VTy = dyn_cast(BucketPtrsTy)) {

545 unsigned EC = VTy->getElementCount().getKnownMinValue();

548

549

550 unsigned LegalEltSize = EltSize <= 32 ? 32 : 64;

551

552 if (EC == 2 || (LegalEltSize == 32 && EC == 4))

554

556 TotalHistCnts = EC / NaturalVectorWidth;

557 }

558

560}

561

565

566

567

568

570 if (auto *VTy = dyn_cast(RetTy))

573

574 switch (ICA.getID()) {

575 case Intrinsic::experimental_vector_histogram_add:

576 if (!ST->hasSVE2())

579 case Intrinsic::umin:

580 case Intrinsic::umax:

581 case Intrinsic::smin:

582 case Intrinsic::smax: {

583 static const auto ValidMinMaxTys = {MVT::v8i8, MVT::v16i8, MVT::v4i16,

584 MVT::v8i16, MVT::v2i32, MVT::v4i32,

585 MVT::nxv16i8, MVT::nxv8i16, MVT::nxv4i32,

586 MVT::nxv2i64};

588

589 if (LT.second == MVT::v2i64)

590 return LT.first * 2;

591 if (any_of(ValidMinMaxTys, [&LT](MVT M) { return M == LT.second; }))

592 return LT.first;

593 break;

594 }

595 case Intrinsic::sadd_sat:

596 case Intrinsic::ssub_sat:

597 case Intrinsic::uadd_sat:

598 case Intrinsic::usub_sat: {

599 static const auto ValidSatTys = {MVT::v8i8, MVT::v16i8, MVT::v4i16,

600 MVT::v8i16, MVT::v2i32, MVT::v4i32,

601 MVT::v2i64};

603

604

605 unsigned Instrs =

606 LT.second.getScalarSizeInBits() == RetTy->getScalarSizeInBits() ? 1 : 4;

607 if (any_of(ValidSatTys, [&LT](MVT M) { return M == LT.second; }))

608 return LT.first * Instrs;

609 break;

610 }

611 case Intrinsic::abs: {

612 static const auto ValidAbsTys = {MVT::v8i8, MVT::v16i8, MVT::v4i16,

613 MVT::v8i16, MVT::v2i32, MVT::v4i32,

614 MVT::v2i64};

616 if (any_of(ValidAbsTys, [&LT](MVT M) { return M == LT.second; }))

617 return LT.first;

618 break;

619 }

620 case Intrinsic::bswap: {

621 static const auto ValidAbsTys = {MVT::v4i16, MVT::v8i16, MVT::v2i32,

622 MVT::v4i32, MVT::v2i64};

624 if (any_of(ValidAbsTys, [&LT](MVT M) { return M == LT.second; }) &&

625 LT.second.getScalarSizeInBits() == RetTy->getScalarSizeInBits())

626 return LT.first;

627 break;

628 }

629 case Intrinsic::stepvector: {

632

633

634 if (LT.first > 1) {

638 Cost += AddCost * (LT.first - 1);

639 }

641 }

642 case Intrinsic::vector_extract:

643 case Intrinsic::vector_insert: {

644

645

646

647

648

649

652 break;

653

656 bool IsExtract = ICA.getID() == Intrinsic::vector_extract;

659

660

662 break;

663

673 break;

674 }

675 case Intrinsic::bitreverse: {

676 static const CostTblEntry BitreverseTbl[] = {

677 {Intrinsic::bitreverse, MVT::i32, 1},

678 {Intrinsic::bitreverse, MVT::i64, 1},

679 {Intrinsic::bitreverse, MVT::v8i8, 1},

680 {Intrinsic::bitreverse, MVT::v16i8, 1},

681 {Intrinsic::bitreverse, MVT::v4i16, 2},

682 {Intrinsic::bitreverse, MVT::v8i16, 2},

683 {Intrinsic::bitreverse, MVT::v2i32, 2},

684 {Intrinsic::bitreverse, MVT::v4i32, 2},

685 {Intrinsic::bitreverse, MVT::v1i64, 2},

686 {Intrinsic::bitreverse, MVT::v2i64, 2},

687 };

689 const auto *Entry =

691 if (Entry) {

692

693

696 return LegalisationCost.first * Entry->Cost + 1;

697

698 return LegalisationCost.first * Entry->Cost;

699 }

700 break;

701 }

702 case Intrinsic::ctpop: {

703 if (!ST->hasNEON()) {

704

706 }

717 };

719 MVT MTy = LT.second;

721

722

724 RetTy->getScalarSizeInBits()

725 ? 1

726 : 0;

727 return LT.first * Entry->Cost + ExtraCost;

728 }

729 break;

730 }

731 case Intrinsic::sadd_with_overflow:

732 case Intrinsic::uadd_with_overflow:

733 case Intrinsic::ssub_with_overflow:

734 case Intrinsic::usub_with_overflow:

735 case Intrinsic::smul_with_overflow:

736 case Intrinsic::umul_with_overflow: {

737 static const CostTblEntry WithOverflowCostTbl[] = {

738 {Intrinsic::sadd_with_overflow, MVT::i8, 3},

739 {Intrinsic::uadd_with_overflow, MVT::i8, 3},

740 {Intrinsic::sadd_with_overflow, MVT::i16, 3},

741 {Intrinsic::uadd_with_overflow, MVT::i16, 3},

742 {Intrinsic::sadd_with_overflow, MVT::i32, 1},

743 {Intrinsic::uadd_with_overflow, MVT::i32, 1},

744 {Intrinsic::sadd_with_overflow, MVT::i64, 1},

745 {Intrinsic::uadd_with_overflow, MVT::i64, 1},

746 {Intrinsic::ssub_with_overflow, MVT::i8, 3},

747 {Intrinsic::usub_with_overflow, MVT::i8, 3},

748 {Intrinsic::ssub_with_overflow, MVT::i16, 3},

749 {Intrinsic::usub_with_overflow, MVT::i16, 3},

750 {Intrinsic::ssub_with_overflow, MVT::i32, 1},

751 {Intrinsic::usub_with_overflow, MVT::i32, 1},

752 {Intrinsic::ssub_with_overflow, MVT::i64, 1},

753 {Intrinsic::usub_with_overflow, MVT::i64, 1},

754 {Intrinsic::smul_with_overflow, MVT::i8, 5},

755 {Intrinsic::umul_with_overflow, MVT::i8, 4},

756 {Intrinsic::smul_with_overflow, MVT::i16, 5},

757 {Intrinsic::umul_with_overflow, MVT::i16, 4},

758 {Intrinsic::smul_with_overflow, MVT::i32, 2},

759 {Intrinsic::umul_with_overflow, MVT::i32, 2},

760 {Intrinsic::smul_with_overflow, MVT::i64, 3},

761 {Intrinsic::umul_with_overflow, MVT::i64, 3},

762 };

767 return Entry->Cost;

768 break;

769 }

770 case Intrinsic::fptosi_sat:

771 case Intrinsic::fptoui_sat: {

773 break;

774 bool IsSigned = ICA.getID() == Intrinsic::fptosi_sat;

777

778

779 if ((LT.second == MVT::f32 || LT.second == MVT::f64 ||

780 LT.second == MVT::v2f32 || LT.second == MVT::v4f32 ||

781 LT.second == MVT::v2f64)) {

783 (LT.second == MVT::f64 && MTy == MVT::i32) ||

784 (LT.second == MVT::f32 && MTy == MVT::i64)))

785 return LT.first;

786

787 if (LT.second.getScalarType() == MVT::f32 && MTy.isFixedLengthVector() &&

790 }

791

792

793 if (LT.second.getScalarType() == MVT::f16 && !ST->hasFullFP16())

800 if ((LT.second == MVT::f16 && MTy == MVT::i32) ||

801 (LT.second == MVT::f16 && MTy == MVT::i64) ||

802 ((LT.second == MVT::v4f16 || LT.second == MVT::v8f16) &&

804 return LT.first;

805

809

810

811 if (LT.second.getScalarType() == MVT::f16 && MTy.isFixedLengthVector() &&

814

815

816 if ((LT.second.getScalarType() == MVT::f32 ||

817 LT.second.getScalarType() == MVT::f64 ||

818 LT.second.getScalarType() == MVT::f16) &&

820 Type *LegalTy =

822 if (LT.second.isVector())

823 LegalTy = VectorType::get(LegalTy, LT.second.getVectorElementCount());

826 LegalTy, {LegalTy, LegalTy});

829 LegalTy, {LegalTy, LegalTy});

831 return LT.first * Cost +

832 ((LT.second.getScalarType() != MVT::f16 || ST->hasFullFP16()) ? 0

833 : 1);

834 }

835

836

839 if (LT.second.isVector()) {

840 FPTy = VectorType::get(FPTy, LT.second.getVectorElementCount());

842 }

848 getCastInstrCost(IsSigned ? Instruction::FPToSI : Instruction::FPToUI,

850 if (IsSigned) {

851 Type *CondTy = RetTy->getWithNewBitWidth(1);

856 }

857 return LT.first * Cost;

858 }

859 case Intrinsic::fshl:

860 case Intrinsic::fshr: {

862 break;

863

864

867 break;

868

871

873 {Intrinsic::fshl, MVT::v4i32, 3},

874 {Intrinsic::fshl, MVT::v2i64, 3}, {Intrinsic::fshl, MVT::v16i8, 4},

875 {Intrinsic::fshl, MVT::v8i16, 4}, {Intrinsic::fshl, MVT::v2i32, 3},

876 {Intrinsic::fshl, MVT::v8i8, 4}, {Intrinsic::fshl, MVT::v4i16, 4}};

877

878

879 const auto *Entry =

880 CostTableLookup(FshlTbl, Intrinsic::fshl, LegalisationCost.second);

881 if (Entry)

882 return LegalisationCost.first * Entry->Cost;

883 }

884

886 if (RetTy->isIntegerTy())

887 break;

888

889

890

891 bool HigherCost = (RetTy->getScalarSizeInBits() != 32 &&

892 RetTy->getScalarSizeInBits() < 64) ||

893 (RetTy->getScalarSizeInBits() % 64 != 0);

894 unsigned ExtraCost = HigherCost ? 1 : 0;

895 if (RetTy->getScalarSizeInBits() == 32 ||

896 RetTy->getScalarSizeInBits() == 64)

897 ExtraCost = 0;

898

899 else if (HigherCost)

900 ExtraCost = 1;

901 else

902 break;

903 return TyL.first + ExtraCost;

904 }

905 case Intrinsic::get_active_lane_mask: {

910 if (!getTLI()->shouldExpandGetActiveLaneMask(RetVT, OpVT) &&

912

913

914

915

916

917

918

919

920

921 return RetTy->getNumElements() * 2;

922 }

923 }

924 break;

925 }

926 case Intrinsic::experimental_vector_match: {

927 auto *NeedleTy = cast(ICA.getArgTypes()[1]);

929 unsigned SearchSize = NeedleTy->getNumElements();

930 if (!getTLI()->shouldExpandVectorMatch(SearchVT, SearchSize)) {

931

932

933

934

935

937 if (isa(RetTy))

940 }

941 break;

942 }

943 default:

944 break;

945 }

947}

948

949

950

954 auto RequiredType = II.getType();

955

956 auto *PN = dyn_cast(II.getArgOperand(0));

957 assert(PN && "Expected Phi Node!");

958

959

960 if (!PN->hasOneUse())

961 return std::nullopt;

962

963 for (Value *IncValPhi : PN->incoming_values()) {

964 auto *Reinterpret = dyn_cast(IncValPhi);

965 if (!Reinterpret ||

966 Reinterpret->getIntrinsicID() !=

967 Intrinsic::aarch64_sve_convert_to_svbool ||

968 RequiredType != Reinterpret->getArgOperand(0)->getType())

969 return std::nullopt;

970 }

971

972

976

977 for (unsigned I = 0; I < PN->getNumIncomingValues(); I++) {

978 auto *Reinterpret = cast(PN->getIncomingValue(I));

979 NPN->addIncoming(Reinterpret->getOperand(0), PN->getIncomingBlock(I));

981 }

982

983

985}

986

987

988

989

990

991

992

993

994

995

996static std::optional<Instruction *>

998 auto BinOp = dyn_cast(II.getOperand(0));

999 if (!BinOp)

1000 return std::nullopt;

1001

1002 auto IntrinsicID = BinOp->getIntrinsicID();

1003 switch (IntrinsicID) {

1004 case Intrinsic::aarch64_sve_and_z:

1005 case Intrinsic::aarch64_sve_bic_z:

1006 case Intrinsic::aarch64_sve_eor_z:

1007 case Intrinsic::aarch64_sve_nand_z:

1008 case Intrinsic::aarch64_sve_nor_z:

1009 case Intrinsic::aarch64_sve_orn_z:

1010 case Intrinsic::aarch64_sve_orr_z:

1011 break;

1012 default:

1013 return std::nullopt;

1014 }

1015

1016 auto BinOpPred = BinOp->getOperand(0);

1017 auto BinOpOp1 = BinOp->getOperand(1);

1018 auto BinOpOp2 = BinOp->getOperand(2);

1019

1020 auto PredIntr = dyn_cast(BinOpPred);

1021 if (!PredIntr ||

1022 PredIntr->getIntrinsicID() != Intrinsic::aarch64_sve_convert_to_svbool)

1023 return std::nullopt;

1024

1025 auto PredOp = PredIntr->getOperand(0);

1026 auto PredOpTy = cast(PredOp->getType());

1027 if (PredOpTy != II.getType())

1028 return std::nullopt;

1029

1032 Intrinsic::aarch64_sve_convert_from_svbool, {PredOpTy}, {BinOpOp1});

1033 NarrowedBinOpArgs.push_back(NarrowBinOpOp1);

1034 if (BinOpOp1 == BinOpOp2)

1035 NarrowedBinOpArgs.push_back(NarrowBinOpOp1);

1036 else

1038 Intrinsic::aarch64_sve_convert_from_svbool, {PredOpTy}, {BinOpOp2}));

1039

1040 auto NarrowedBinOp =

1043}

1044

1045static std::optional<Instruction *>

1047

1048 if (isa(II.getArgOperand(0)))

1050

1052 return BinOpCombine;

1053

1054

1055 if (isa(II.getArgOperand(0)->getType()) ||

1056 isa(II.getType()))

1057 return std::nullopt;

1058

1060 Value *Cursor = II.getOperand(0), *EarliestReplacement = nullptr;

1061

1062 const auto *IVTy = cast(II.getType());

1063

1064

1065 while (Cursor) {

1066

1067

1068 const auto *CursorVTy = cast(Cursor->getType());

1069 if (CursorVTy->getElementCount().getKnownMinValue() <

1070 IVTy->getElementCount().getKnownMinValue())

1071 break;

1072

1073

1074 if (Cursor->getType() == IVTy)

1075 EarliestReplacement = Cursor;

1076

1077 auto *IntrinsicCursor = dyn_cast(Cursor);

1078

1079

1080 if (!IntrinsicCursor || !(IntrinsicCursor->getIntrinsicID() ==

1081 Intrinsic::aarch64_sve_convert_to_svbool ||

1082 IntrinsicCursor->getIntrinsicID() ==

1083 Intrinsic::aarch64_sve_convert_from_svbool))

1084 break;

1085

1086 CandidatesForRemoval.insert(CandidatesForRemoval.begin(), IntrinsicCursor);

1087 Cursor = IntrinsicCursor->getOperand(0);

1088 }

1089

1090

1091

1092 if (!EarliestReplacement)

1093 return std::nullopt;

1094

1096}

1097

1099

1100 Value *UncastedPred;

1101 if (match(Pred, m_IntrinsicIntrinsic::aarch64\_sve\_convert\_from\_svbool(

1102 m_IntrinsicIntrinsic::aarch64\_sve\_convert\_to\_svbool(

1103 m_Value(UncastedPred)))))

1104

1105

1106 if (cast(Pred->getType())->getMinNumElements() <=

1107 cast(UncastedPred->getType())->getMinNumElements())

1108 Pred = UncastedPred;

1109

1110 return match(Pred, m_IntrinsicIntrinsic::aarch64\_sve\_ptrue(

1111 m_ConstantIntAArch64SVEPredPattern::all()));

1112}

1113

1114

1115

1116static std::optional<Instruction *>

1118 bool hasInactiveVector) {

1119 int PredOperand = hasInactiveVector ? 1 : 0;

1120 int ReplaceOperand = hasInactiveVector ? 0 : 1;

1124 }

1125 return std::nullopt;

1126}

1127

1128

1129

1130static std::optional<Instruction *>

1133 !isallvm::UndefValue(II.getOperand(0)) &&

1134 !isallvm::PoisonValue(II.getOperand(0))) {

1137 }

1139}

1140

1141

1142static std::optional<Instruction *>

1144 int PredPos) {

1147 }

1148 return std::nullopt;

1149}

1150

1151

1152

1153static std::optional<Instruction *>

1158 if (RetTy->isStructTy()) {

1159 auto StructT = cast(RetTy);

1160 auto VecT = StructT->getElementType(0);

1162 for (unsigned i = 0; i < StructT->getNumElements(); i++) {

1163 ZerVec.push_back(VecT->isFPOrFPVectorTy() ? ConstantFP::get(VecT, 0.0)

1164 : ConstantInt::get(VecT, 0));

1165 }

1167 } else

1168 Node = RetTy->isFPOrFPVectorTy() ? ConstantFP::get(RetTy, 0.0)

1169 : ConstantInt::get(II.getType(), 0);

1170

1173 }

1174 return std::nullopt;

1175}

1176

1179

1180 auto *OpPredicate = II.getOperand(0);

1183

1187}

1188

1191 IntrinsicInst *Pg = dyn_cast(II.getArgOperand(1));

1192 if (!Pg)

1193 return std::nullopt;

1194

1195 if (Pg->getIntrinsicID() != Intrinsic::aarch64_sve_ptrue)

1196 return std::nullopt;

1197

1198 const auto PTruePattern =

1199 cast(Pg->getOperand(0))->getZExtValue();

1200 if (PTruePattern != AArch64SVEPredPattern::vl1)

1201 return std::nullopt;

1202

1203

1206 II.getArgOperand(0), II.getArgOperand(2), ConstantInt::get(IdxTy, 0));

1207 Insert->insertBefore(&II);

1208 Insert->takeName(&II);

1209

1211}

1212

1215

1216 auto *RetTy = cast(II.getType());

1218 II.getArgOperand(0));

1221}

1222

1226

1227

1229 return II_NA;

1230

1231

1232 auto *Pg = dyn_cast(II.getArgOperand(0));

1233 if (!Pg || Pg->getIntrinsicID() != Intrinsic::aarch64_sve_ptrue)

1234 return std::nullopt;

1235

1236 const auto PTruePattern =

1237 cast(Pg->getOperand(0))->getZExtValue();

1238 if (PTruePattern != AArch64SVEPredPattern::all)

1239 return std::nullopt;

1240

1241

1242 auto *SplatValue =

1243 dyn_cast_or_null(getSplatValue(II.getArgOperand(2)));

1244 if (!SplatValue || !SplatValue->isZero())

1245 return std::nullopt;

1246

1247

1248 auto *DupQLane = dyn_cast(II.getArgOperand(1));

1249 if (!DupQLane ||

1250 DupQLane->getIntrinsicID() != Intrinsic::aarch64_sve_dupq_lane)

1251 return std::nullopt;

1252

1253

1254 auto *DupQLaneIdx = dyn_cast(DupQLane->getArgOperand(1));

1255 if (!DupQLaneIdx || !DupQLaneIdx->isZero())

1256 return std::nullopt;

1257

1258 auto *VecIns = dyn_cast(DupQLane->getArgOperand(0));

1259 if (!VecIns || VecIns->getIntrinsicID() != Intrinsic::vector_insert)

1260 return std::nullopt;

1261

1262

1263

1264 if (!isa(VecIns->getArgOperand(0)))

1265 return std::nullopt;

1266

1267 if (!cast(VecIns->getArgOperand(2))->isZero())

1268 return std::nullopt;

1269

1270 auto *ConstVec = dyn_cast(VecIns->getArgOperand(1));

1271 if (!ConstVec)

1272 return std::nullopt;

1273

1274 auto *VecTy = dyn_cast(ConstVec->getType());

1275 auto *OutTy = dyn_cast(II.getType());

1276 if (!VecTy || !OutTy || VecTy->getNumElements() != OutTy->getMinNumElements())

1277 return std::nullopt;

1278

1279 unsigned NumElts = VecTy->getNumElements();

1280 unsigned PredicateBits = 0;

1281

1282

1283 for (unsigned I = 0; I < NumElts; ++I) {

1284 auto *Arg = dyn_cast(ConstVec->getAggregateElement(I));

1285 if (!Arg)

1286 return std::nullopt;

1287 if (!Arg->isZero())

1288 PredicateBits |= 1 << (I * (16 / NumElts));

1289 }

1290

1291

1292 if (PredicateBits == 0) {

1294 PFalse->takeName(&II);

1296 }

1297

1298

1299 unsigned Mask = 8;

1300 for (unsigned I = 0; I < 16; ++I)

1301 if ((PredicateBits & (1 << I)) != 0)

1302 Mask |= (I % 8);

1303

1304 unsigned PredSize = Mask & -Mask;

1307

1308

1309 for (unsigned I = 0; I < 16; I += PredSize)

1310 if ((PredicateBits & (1 << I)) == 0)

1311 return std::nullopt;

1312

1313 auto *PTruePat =

1314 ConstantInt::get(Type::getInt32Ty(Ctx), AArch64SVEPredPattern::all);

1316 {PredType}, {PTruePat});

1318 Intrinsic::aarch64_sve_convert_to_svbool, {PredType}, {PTrue});

1319 auto *ConvertFromSVBool =

1321 {II.getType()}, {ConvertToSVBool});

1322

1325}

1326

1329 Value *Pg = II.getArgOperand(0);

1330 Value *Vec = II.getArgOperand(1);

1331 auto IntrinsicID = II.getIntrinsicID();

1332 bool IsAfter = IntrinsicID == Intrinsic::aarch64_sve_lasta;

1333

1334

1337

1338

1339

1343 auto *OldBinOp = cast(Vec);

1344 auto OpC = OldBinOp->getOpcode();

1345 auto *NewLHS =

1347 auto *NewRHS =

1350 OpC, NewLHS, NewRHS, OldBinOp, OldBinOp->getName(), II.getIterator());

1352 }

1353 }

1354

1355 auto *C = dyn_cast(Pg);

1356 if (IsAfter && C && C->isNullValue()) {

1357

1360 Extract->insertBefore(&II);

1361 Extract->takeName(&II);

1363 }

1364

1365 auto *IntrPG = dyn_cast(Pg);

1366 if (!IntrPG)

1367 return std::nullopt;

1368

1369 if (IntrPG->getIntrinsicID() != Intrinsic::aarch64_sve_ptrue)

1370 return std::nullopt;

1371

1372 const auto PTruePattern =

1373 cast(IntrPG->getOperand(0))->getZExtValue();

1374

1375

1377 if (!MinNumElts)

1378 return std::nullopt;

1379

1380 unsigned Idx = MinNumElts - 1;

1381

1382

1383 if (IsAfter)

1385

1386

1387

1388

1389 auto *PgVTy = cast(Pg->getType());

1390 if (Idx >= PgVTy->getMinNumElements())

1391 return std::nullopt;

1392

1393

1396 Extract->insertBefore(&II);

1397 Extract->takeName(&II);

1399}

1400

1403

1404

1405

1406

1407

1408

1409

1410 Value *Pg = II.getArgOperand(0);

1412 Value *Vec = II.getArgOperand(2);

1413 Type *Ty = II.getType();

1414

1416 return std::nullopt;

1417

1419 switch (cast(Ty)->getBitWidth()) {

1420 default:

1421 return std::nullopt;

1422 case 16:

1424 break;

1425 case 32:

1427 break;

1428 case 64:

1430 break;

1431 }

1432

1435 FPTy, cast(Vec->getType())->getElementCount());

1438 II.getIntrinsicID(), {FPVec->getType()}, {Pg, FPFallBack, FPVec});

1441}

1442

1446

1447

1448 auto *AllPat =

1449 ConstantInt::get(Type::getInt32Ty(Ctx), AArch64SVEPredPattern::all);

1451 {II.getType()}, {AllPat});

1452 auto *RDFFR =

1456}

1457

1458static std::optional<Instruction *>

1460 const auto Pattern = cast(II.getArgOperand(0))->getZExtValue();

1461

1462 if (Pattern == AArch64SVEPredPattern::all) {

1463 Constant *StepVal = ConstantInt::get(II.getType(), NumElts);

1467 }

1468

1470

1471 return MinNumElts && NumElts >= MinNumElts

1473 II, ConstantInt::get(II.getType(), MinNumElts)))

1474 : std::nullopt;

1475}

1476

1479 Value *PgVal = II.getArgOperand(0);

1480 Value *OpVal = II.getArgOperand(1);

1481

1482

1483

1484 if (PgVal == OpVal &&

1485 (II.getIntrinsicID() == Intrinsic::aarch64_sve_ptest_first ||

1486 II.getIntrinsicID() == Intrinsic::aarch64_sve_ptest_last)) {

1487 Value *Ops[] = {PgVal, OpVal};

1489

1490 auto *PTest =

1493

1495 }

1496

1497 IntrinsicInst *Pg = dyn_cast(PgVal);

1499

1500 if (!Pg || Op)

1501 return std::nullopt;

1502

1504

1505 if (Pg->getIntrinsicID() == Intrinsic::aarch64_sve_convert_to_svbool &&

1506 OpIID == Intrinsic::aarch64_sve_convert_to_svbool &&

1510

1512

1515 }

1516

1517

1518

1519

1520 if ((Pg == Op) && (II.getIntrinsicID() == Intrinsic::aarch64_sve_ptest_any) &&

1521 ((OpIID == Intrinsic::aarch64_sve_brka_z) ||

1522 (OpIID == Intrinsic::aarch64_sve_brkb_z) ||

1523 (OpIID == Intrinsic::aarch64_sve_brkpa_z) ||

1524 (OpIID == Intrinsic::aarch64_sve_brkpb_z) ||

1525 (OpIID == Intrinsic::aarch64_sve_rdffr_z) ||

1526 (OpIID == Intrinsic::aarch64_sve_and_z) ||

1527 (OpIID == Intrinsic::aarch64_sve_bic_z) ||

1528 (OpIID == Intrinsic::aarch64_sve_eor_z) ||

1529 (OpIID == Intrinsic::aarch64_sve_nand_z) ||

1530 (OpIID == Intrinsic::aarch64_sve_nor_z) ||

1531 (OpIID == Intrinsic::aarch64_sve_orn_z) ||

1532 (OpIID == Intrinsic::aarch64_sve_orr_z))) {

1535

1538

1540 }

1541

1542 return std::nullopt;

1543}

1544

1545template <Intrinsic::ID MulOpc, typename Intrinsic::ID FuseOpc>

1546static std::optional<Instruction *>

1548 bool MergeIntoAddendOp) {

1549 Value *P = II.getOperand(0);

1550 Value *MulOp0, *MulOp1, *AddendOp, *Mul;

1551 if (MergeIntoAddendOp) {

1552 AddendOp = II.getOperand(1);

1553 Mul = II.getOperand(2);

1554 } else {

1555 AddendOp = II.getOperand(2);

1556 Mul = II.getOperand(1);

1557 }

1558

1561 return std::nullopt;

1562

1563 if (Mul->hasOneUse())

1564 return std::nullopt;

1565

1567 if (II.getType()->isFPOrFPVectorTy()) {

1569

1570

1571 if (FAddFlags != cast(Mul)->getFastMathFlags())

1572 return std::nullopt;

1574 return std::nullopt;

1576 }

1577

1579 if (MergeIntoAddendOp)

1581 {P, AddendOp, MulOp0, MulOp1}, FMFSource);

1582 else

1584 {P, MulOp0, MulOp1, AddendOp}, FMFSource);

1585

1587}

1588

1589static std::optional<Instruction *>

1591 Value *Pred = II.getOperand(0);

1592 Value *PtrOp = II.getOperand(1);

1593 Type *VecTy = II.getType();

1594

1595

1597 return II_NA;

1598

1601 Load->copyMetadata(II);

1603 }

1604

1610}

1611

1612static std::optional<Instruction *>

1614 Value *VecOp = II.getOperand(0);

1615 Value *Pred = II.getOperand(1);

1616 Value *PtrOp = II.getOperand(2);

1617

1620 Store->copyMetadata(II);

1622 }

1623

1628}

1629

1631 switch (Intrinsic) {

1632 case Intrinsic::aarch64_sve_fmul_u:

1633 return Instruction::BinaryOps::FMul;

1634 case Intrinsic::aarch64_sve_fadd_u:

1635 return Instruction::BinaryOps::FAdd;

1636 case Intrinsic::aarch64_sve_fsub_u:

1637 return Instruction::BinaryOps::FSub;

1638 default:

1639 return Instruction::BinaryOpsEnd;

1640 }

1641}

1642

1643static std::optional<Instruction *>

1645

1646 if (II.isStrictFP())

1647 return std::nullopt;

1648

1649 auto *OpPredicate = II.getOperand(0);

1651 if (BinOpCode == Instruction::BinaryOpsEnd ||

1652 match(OpPredicate, m_IntrinsicIntrinsic::aarch64\_sve\_ptrue(

1653 m_ConstantIntAArch64SVEPredPattern::all())))

1654 return std::nullopt;

1656 BinOpCode, II.getOperand(1), II.getOperand(2), II.getFastMathFlags());

1658}

1659

1660

1661

1664 auto *OpPredicate = II.getOperand(0);

1665 if (match(OpPredicate, m_IntrinsicIntrinsic::aarch64\_sve\_ptrue(

1666 m_ConstantIntAArch64SVEPredPattern::all())))

1667 return std::nullopt;

1668

1669 auto *Mod = II.getModule();

1671 II.setCalledFunction(NewDecl);

1672

1673 return &II;

1674}

1675

1676

1677

1678static std::optional<Instruction *>

1682

1683

1685 }

1687}

1688

1691 if (auto II_U =

1693 return II_U;

1695 Intrinsic::aarch64_sve_mla>(

1696 IC, II, true))

1697 return MLA;

1699 Intrinsic::aarch64_sve_mad>(

1700 IC, II, false))

1701 return MAD;

1702 return std::nullopt;

1703}

1704

1705static std::optional<Instruction *>

1707 if (auto II_U =

1709 return II_U;

1710 if (auto FMLA =

1712 Intrinsic::aarch64_sve_fmla>(IC, II,

1713 true))

1714 return FMLA;

1715 if (auto FMAD =

1717 Intrinsic::aarch64_sve_fmad>(IC, II,

1718 false))

1719 return FMAD;

1720 if (auto FMLA =

1722 Intrinsic::aarch64_sve_fmla>(IC, II,

1723 true))

1724 return FMLA;

1725 return std::nullopt;

1726}

1727

1728static std::optional<Instruction *>

1730 if (auto FMLA =

1732 Intrinsic::aarch64_sve_fmla>(IC, II,

1733 true))

1734 return FMLA;

1735 if (auto FMAD =

1737 Intrinsic::aarch64_sve_fmad>(IC, II,

1738 false))

1739 return FMAD;

1740 if (auto FMLA_U =

1742 Intrinsic::aarch64_sve_fmla_u>(

1743 IC, II, true))

1744 return FMLA_U;

1746}

1747

1748static std::optional<Instruction *>

1750 if (auto II_U =

1752 return II_U;

1753 if (auto FMLS =

1755 Intrinsic::aarch64_sve_fmls>(IC, II,

1756 true))

1757 return FMLS;

1758 if (auto FMSB =

1760 Intrinsic::aarch64_sve_fnmsb>(

1761 IC, II, false))

1762 return FMSB;

1763 if (auto FMLS =

1765 Intrinsic::aarch64_sve_fmls>(IC, II,

1766 true))

1767 return FMLS;

1768 return std::nullopt;

1769}

1770

1771static std::optional<Instruction *>

1773 if (auto FMLS =

1775 Intrinsic::aarch64_sve_fmls>(IC, II,

1776 true))

1777 return FMLS;

1778 if (auto FMSB =

1780 Intrinsic::aarch64_sve_fnmsb>(

1781 IC, II, false))

1782 return FMSB;

1783 if (auto FMLS_U =

1785 Intrinsic::aarch64_sve_fmls_u>(

1786 IC, II, true))

1787 return FMLS_U;

1789}

1790

1793 if (auto II_U =

1795 return II_U;

1797 Intrinsic::aarch64_sve_mls>(

1798 IC, II, true))

1799 return MLS;

1800 return std::nullopt;

1801}

1802

1806 auto *OpPredicate = II.getOperand(0);

1807 auto *OpMultiplicand = II.getOperand(1);

1808 auto *OpMultiplier = II.getOperand(2);

1809

1810

1811 auto IsUnitSplat = [](auto *I) {

1813 if (!SplatValue)

1814 return false;

1816 };

1817

1818

1819

1820 auto IsUnitDup = [](auto *I) {

1821 auto *IntrI = dyn_cast(I);

1822 if (!IntrI || IntrI->getIntrinsicID() != Intrinsic::aarch64_sve_dup)

1823 return false;

1824

1825 auto *SplatValue = IntrI->getOperand(2);

1827 };

1828

1829 if (IsUnitSplat(OpMultiplier)) {

1830

1831 OpMultiplicand->takeName(&II);

1833 } else if (IsUnitDup(OpMultiplier)) {

1834

1835 auto *DupInst = cast(OpMultiplier);

1836 auto *DupPg = DupInst->getOperand(1);

1837

1838

1839 if (OpPredicate == DupPg) {

1840 OpMultiplicand->takeName(&II);

1842 }

1843 }

1844

1846}

1847

1850 Value *UnpackArg = II.getArgOperand(0);

1851 auto *RetTy = cast(II.getType());

1852 bool IsSigned = II.getIntrinsicID() == Intrinsic::aarch64_sve_sunpkhi ||

1853 II.getIntrinsicID() == Intrinsic::aarch64_sve_sunpklo;

1854

1855

1856

1857 if (auto *ScalarArg = getSplatValue(UnpackArg)) {

1858 ScalarArg =

1864 }

1865

1866 return std::nullopt;

1867}

1870 auto *OpVal = II.getOperand(0);

1871 auto *OpIndices = II.getOperand(1);

1872 VectorType *VTy = cast(II.getType());

1873

1874

1875

1876 auto *SplatValue = dyn_cast_or_null(getSplatValue(OpIndices));

1877 if (!SplatValue ||

1878 SplatValue->getValue().uge(VTy->getElementCount().getKnownMinValue()))

1879 return std::nullopt;

1880

1881

1882

1884 auto *VectorSplat =

1886

1889}

1890

1895 constexpr Intrinsic::ID FromSVB = Intrinsic::aarch64_sve_convert_from_svbool;

1896 constexpr Intrinsic::ID ToSVB = Intrinsic::aarch64_sve_convert_to_svbool;

1897

1898

1899

1900 if ((match(II.getArgOperand(0),

1901 m_Intrinsic(m_Intrinsic(m_Value(A)))) &&

1902 match(II.getArgOperand(1),

1903 m_Intrinsic(m_Intrinsic(m_Value(B))))) ||

1904 (match(II.getArgOperand(0), m_Intrinsic(m_Value(A))) &&

1905 match(II.getArgOperand(1), m_Intrinsic(m_Value(B))))) {

1906 auto *TyA = cast(A->getType());

1907 if (TyA == B->getType() &&

1915 }

1916 }

1917

1918 return std::nullopt;

1919}

1920

1923

1924

1926 if (match(II.getArgOperand(0),

1927 m_IntrinsicIntrinsic::aarch64\_sve\_uzp1(m_Value(A), m_Value(B))) &&

1928 match(II.getArgOperand(1), m_IntrinsicIntrinsic::aarch64\_sve\_uzp2(

1931 II, (II.getIntrinsicID() == Intrinsic::aarch64_sve_zip1 ? A : B));

1932

1933 return std::nullopt;

1934}

1935

1936static std::optional<Instruction *>

1938 Value *Mask = II.getOperand(0);

1939 Value *BasePtr = II.getOperand(1);

1940 Value *Index = II.getOperand(2);

1941 Type *Ty = II.getType();

1943

1944

1946 return II_NA;

1947

1948

1949

1950

1951 Value *IndexBase;

1952 if (match(Index, m_IntrinsicIntrinsic::aarch64\_sve\_index(

1954 Align Alignment =

1955 BasePtr->getPointerAlignment(II.getDataLayout());

1956

1958 BasePtr, IndexBase);

1963 }

1964

1965 return std::nullopt;

1966}

1967

1968static std::optional<Instruction *>

1970 Value *Val = II.getOperand(0);

1971 Value *Mask = II.getOperand(1);

1972 Value *BasePtr = II.getOperand(2);

1973 Value *Index = II.getOperand(3);

1975

1976

1977

1978

1979 Value *IndexBase;

1980 if (match(Index, m_IntrinsicIntrinsic::aarch64\_sve\_index(

1982 Align Alignment =

1983 BasePtr->getPointerAlignment(II.getDataLayout());

1984

1986 BasePtr, IndexBase);

1988

1990 }

1991

1992 return std::nullopt;

1993}

1994

1998 Value *Pred = II.getOperand(0);

1999 Value *Vec = II.getOperand(1);

2000 Value *DivVec = II.getOperand(2);

2001

2003 ConstantInt *SplatConstantInt = dyn_cast_or_null(SplatValue);

2004 if (!SplatConstantInt)

2005 return std::nullopt;

2006

2008 const int64_t DivisorValue = Divisor.getSExtValue();

2009 if (DivisorValue == -1)

2010 return std::nullopt;

2011 if (DivisorValue == 1)

2013

2015 Constant *DivisorLog2 = ConstantInt::get(Int32Ty, Divisor.logBase2());

2017 Intrinsic::aarch64_sve_asrd, {II.getType()}, {Pred, Vec, DivisorLog2});

2019 }

2022 Constant *DivisorLog2 = ConstantInt::get(Int32Ty, Divisor.logBase2());

2024 Intrinsic::aarch64_sve_asrd, {II.getType()}, {Pred, Vec, DivisorLog2});

2026 Intrinsic::aarch64_sve_neg, {ASRD->getType()}, {ASRD, Pred, ASRD});

2028 }

2029

2030 return std::nullopt;

2031}

2032

2034 size_t VecSize = Vec.size();

2035 if (VecSize == 1)

2036 return true;

2038 return false;

2039 size_t HalfVecSize = VecSize / 2;

2040

2041 for (auto LHS = Vec.begin(), RHS = Vec.begin() + HalfVecSize;

2043 if (*LHS != nullptr && *RHS != nullptr) {

2045 continue;

2046 else

2047 return false;

2048 }

2049 if (!AllowPoison)

2050 return false;

2051 if (*LHS == nullptr && *RHS != nullptr)

2053 }

2054

2055 Vec.resize(HalfVecSize);

2057 return true;

2058}

2059

2060

2061

2064 Value *CurrentInsertElt = nullptr, *Default = nullptr;

2065 if (match(II.getOperand(0),

2066 m_IntrinsicIntrinsic::vector\_insert(

2068 !isa(CurrentInsertElt->getType()))

2069 return std::nullopt;

2070 auto IIScalableTy = cast(II.getType());

2071

2072

2074 while (auto InsertElt = dyn_cast(CurrentInsertElt)) {

2075 auto Idx = cast(InsertElt->getOperand(2));

2076 Elts[Idx->getValue().getZExtValue()] = InsertElt->getOperand(1);

2077 CurrentInsertElt = InsertElt->getOperand(0);

2078 }

2079

2080 bool AllowPoison =

2081 isa(CurrentInsertElt) && isa(Default);

2083 return std::nullopt;

2084

2085

2087 for (size_t I = 0; I < Elts.size(); I++) {

2088 if (Elts[I] == nullptr)

2089 continue;

2092 }

2093 if (InsertEltChain == nullptr)

2094 return std::nullopt;

2095

2096

2097

2098

2099

2100 unsigned PatternWidth = IIScalableTy->getScalarSizeInBits() * Elts.size();

2101 unsigned PatternElementCount = IIScalableTy->getScalarSizeInBits() *

2102 IIScalableTy->getMinNumElements() /

2103 PatternWidth;

2104

2107 auto *WideShuffleMaskTy =

2109

2113 auto WideBitcast =

2117 WideBitcast, PoisonValue::get(WideScalableTy), WideShuffleMask);

2118 auto NarrowBitcast =

2120

2122}

2123

2126 Value *A = II.getArgOperand(0);

2127 Value *B = II.getArgOperand(1);

2128 if (A == B)

2130

2131 return std::nullopt;

2132}

2133

2136 Value *Pred = II.getOperand(0);

2137 Value *Vec = II.getOperand(1);

2138 Value *Shift = II.getOperand(2);

2139

2140

2141 Value *AbsPred, *MergedValue;

2142 if (match(Vec, m_IntrinsicIntrinsic::aarch64\_sve\_sqabs(

2144 match(Vec, m_IntrinsicIntrinsic::aarch64\_sve\_abs(

2146

2147 return std::nullopt;

2148

2149

2150

2151

2152

2153 if (!isa(MergedValue) && match(MergedValue, m_NonNegative()) &&

2155 return std::nullopt;

2156

2157

2158

2160 return std::nullopt;

2161

2163 {II.getType()}, {Pred, Vec, Shift});

2164

2166}

2167

2170 Value *Vec = II.getOperand(0);

2171

2174

2175 return std::nullopt;

2176}

2177

2180

2181 auto *NI = II.getNextNonDebugInstruction();

2184 return I->mayReadOrWriteMemory() && I->mayHaveSideEffects();

2185 };

2186 while (LookaheadThreshold-- && CanSkipOver(NI)) {

2187 auto *NIBB = NI->getParent();

2188 NI = NI->getNextNonDebugInstruction();

2189 if (!NI) {

2190 if (auto *SuccBB = NIBB->getUniqueSuccessor())

2191 NI = SuccBB->getFirstNonPHIOrDbgOrLifetime();

2192 else

2193 break;

2194 }

2195 }

2196 auto *NextII = dyn_cast_or_null(NI);

2197 if (NextII && II.isIdenticalTo(NextII))

2199

2200 return std::nullopt;

2201}

2202

2203std::optional<Instruction *>

2207 switch (IID) {

2208 default:

2209 break;

2210 case Intrinsic::aarch64_dmb:

2212 case Intrinsic::aarch64_sve_fcvt_bf16f32_v2:

2213 case Intrinsic::aarch64_sve_fcvt_f16f32:

2214 case Intrinsic::aarch64_sve_fcvt_f16f64:

2215 case Intrinsic::aarch64_sve_fcvt_f32f16:

2216 case Intrinsic::aarch64_sve_fcvt_f32f64:

2217 case Intrinsic::aarch64_sve_fcvt_f64f16:

2218 case Intrinsic::aarch64_sve_fcvt_f64f32:

2219 case Intrinsic::aarch64_sve_fcvtlt_f32f16:

2220 case Intrinsic::aarch64_sve_fcvtlt_f64f32:

2221 case Intrinsic::aarch64_sve_fcvtx_f32f64:

2222 case Intrinsic::aarch64_sve_fcvtzs:

2223 case Intrinsic::aarch64_sve_fcvtzs_i32f16:

2224 case Intrinsic::aarch64_sve_fcvtzs_i32f64:

2225 case Intrinsic::aarch64_sve_fcvtzs_i64f16:

2226 case Intrinsic::aarch64_sve_fcvtzs_i64f32:

2227 case Intrinsic::aarch64_sve_fcvtzu:

2228 case Intrinsic::aarch64_sve_fcvtzu_i32f16:

2229 case Intrinsic::aarch64_sve_fcvtzu_i32f64:

2230 case Intrinsic::aarch64_sve_fcvtzu_i64f16:

2231 case Intrinsic::aarch64_sve_fcvtzu_i64f32:

2232 case Intrinsic::aarch64_sve_scvtf:

2233 case Intrinsic::aarch64_sve_scvtf_f16i32:

2234 case Intrinsic::aarch64_sve_scvtf_f16i64:

2235 case Intrinsic::aarch64_sve_scvtf_f32i64:

2236 case Intrinsic::aarch64_sve_scvtf_f64i32:

2237 case Intrinsic::aarch64_sve_ucvtf:

2238 case Intrinsic::aarch64_sve_ucvtf_f16i32:

2239 case Intrinsic::aarch64_sve_ucvtf_f16i64:

2240 case Intrinsic::aarch64_sve_ucvtf_f32i64:

2241 case Intrinsic::aarch64_sve_ucvtf_f64i32:

2243 case Intrinsic::aarch64_sve_fcvtnt_bf16f32_v2:

2244 case Intrinsic::aarch64_sve_fcvtnt_f16f32:

2245 case Intrinsic::aarch64_sve_fcvtnt_f32f64:

2246 case Intrinsic::aarch64_sve_fcvtxnt_f32f64:

2248 case Intrinsic::aarch64_sve_st1_scatter:

2249 case Intrinsic::aarch64_sve_st1_scatter_scalar_offset:

2250 case Intrinsic::aarch64_sve_st1_scatter_sxtw:

2251 case Intrinsic::aarch64_sve_st1_scatter_sxtw_index:

2252 case Intrinsic::aarch64_sve_st1_scatter_uxtw:

2253 case Intrinsic::aarch64_sve_st1_scatter_uxtw_index:

2254 case Intrinsic::aarch64_sve_st1dq:

2255 case Intrinsic::aarch64_sve_st1q_scatter_index:

2256 case Intrinsic::aarch64_sve_st1q_scatter_scalar_offset:

2257 case Intrinsic::aarch64_sve_st1q_scatter_vector_offset:

2258 case Intrinsic::aarch64_sve_st1wq:

2259 case Intrinsic::aarch64_sve_stnt1:

2260 case Intrinsic::aarch64_sve_stnt1_scatter:

2261 case Intrinsic::aarch64_sve_stnt1_scatter_index:

2262 case Intrinsic::aarch64_sve_stnt1_scatter_scalar_offset:

2263 case Intrinsic::aarch64_sve_stnt1_scatter_uxtw:

2265 case Intrinsic::aarch64_sve_st2:

2266 case Intrinsic::aarch64_sve_st2q:

2268 case Intrinsic::aarch64_sve_st3:

2269 case Intrinsic::aarch64_sve_st3q:

2271 case Intrinsic::aarch64_sve_st4:

2272 case Intrinsic::aarch64_sve_st4q:

2274 case Intrinsic::aarch64_sve_addqv:

2275 case Intrinsic::aarch64_sve_and_z:

2276 case Intrinsic::aarch64_sve_bic_z:

2277 case Intrinsic::aarch64_sve_brka_z:

2278 case Intrinsic::aarch64_sve_brkb_z:

2279 case Intrinsic::aarch64_sve_brkn_z:

2280 case Intrinsic::aarch64_sve_brkpa_z:

2281 case Intrinsic::aarch64_sve_brkpb_z:

2282 case Intrinsic::aarch64_sve_cntp:

2283 case Intrinsic::aarch64_sve_compact:

2284 case Intrinsic::aarch64_sve_eor_z:

2285 case Intrinsic::aarch64_sve_eorv:

2286 case Intrinsic::aarch64_sve_eorqv:

2287 case Intrinsic::aarch64_sve_nand_z:

2288 case Intrinsic::aarch64_sve_nor_z:

2289 case Intrinsic::aarch64_sve_orn_z:

2290 case Intrinsic::aarch64_sve_orr_z:

2291 case Intrinsic::aarch64_sve_orv:

2292 case Intrinsic::aarch64_sve_orqv:

2293 case Intrinsic::aarch64_sve_pnext:

2294 case Intrinsic::aarch64_sve_rdffr_z:

2295 case Intrinsic::aarch64_sve_saddv:

2296 case Intrinsic::aarch64_sve_uaddv:

2297 case Intrinsic::aarch64_sve_umaxv:

2298 case Intrinsic::aarch64_sve_umaxqv:

2299 case Intrinsic::aarch64_sve_cmpeq:

2300 case Intrinsic::aarch64_sve_cmpeq_wide:

2301 case Intrinsic::aarch64_sve_cmpge:

2302 case Intrinsic::aarch64_sve_cmpge_wide:

2303 case Intrinsic::aarch64_sve_cmpgt:

2304 case Intrinsic::aarch64_sve_cmpgt_wide:

2305 case Intrinsic::aarch64_sve_cmphi:

2306 case Intrinsic::aarch64_sve_cmphi_wide:

2307 case Intrinsic::aarch64_sve_cmphs:

2308 case Intrinsic::aarch64_sve_cmphs_wide:

2309 case Intrinsic::aarch64_sve_cmple_wide:

2310 case Intrinsic::aarch64_sve_cmplo_wide:

2311 case Intrinsic::aarch64_sve_cmpls_wide:

2312 case Intrinsic::aarch64_sve_cmplt_wide:

2313 case Intrinsic::aarch64_sve_facge:

2314 case Intrinsic::aarch64_sve_facgt:

2315 case Intrinsic::aarch64_sve_fcmpeq:

2316 case Intrinsic::aarch64_sve_fcmpge:

2317 case Intrinsic::aarch64_sve_fcmpgt:

2318 case Intrinsic::aarch64_sve_fcmpne:

2319 case Intrinsic::aarch64_sve_fcmpuo:

2320 case Intrinsic::aarch64_sve_ld1_gather:

2321 case Intrinsic::aarch64_sve_ld1_gather_scalar_offset:

2322 case Intrinsic::aarch64_sve_ld1_gather_sxtw:

2323 case Intrinsic::aarch64_sve_ld1_gather_sxtw_index:

2324 case Intrinsic::aarch64_sve_ld1_gather_uxtw:

2325 case Intrinsic::aarch64_sve_ld1_gather_uxtw_index:

2326 case Intrinsic::aarch64_sve_ld1q_gather_index:

2327 case Intrinsic::aarch64_sve_ld1q_gather_scalar_offset:

2328 case Intrinsic::aarch64_sve_ld1q_gather_vector_offset:

2329 case Intrinsic::aarch64_sve_ld1ro:

2330 case Intrinsic::aarch64_sve_ld1rq:

2331 case Intrinsic::aarch64_sve_ld1udq:

2332 case Intrinsic::aarch64_sve_ld1uwq:

2333 case Intrinsic::aarch64_sve_ld2_sret:

2334 case Intrinsic::aarch64_sve_ld2q_sret:

2335 case Intrinsic::aarch64_sve_ld3_sret:

2336 case Intrinsic::aarch64_sve_ld3q_sret:

2337 case Intrinsic::aarch64_sve_ld4_sret:

2338 case Intrinsic::aarch64_sve_ld4q_sret:

2339 case Intrinsic::aarch64_sve_ldff1:

2340 case Intrinsic::aarch64_sve_ldff1_gather:

2341 case Intrinsic::aarch64_sve_ldff1_gather_index:

2342 case Intrinsic::aarch64_sve_ldff1_gather_scalar_offset:

2343 case Intrinsic::aarch64_sve_ldff1_gather_sxtw:

2344 case Intrinsic::aarch64_sve_ldff1_gather_sxtw_index:

2345 case Intrinsic::aarch64_sve_ldff1_gather_uxtw:

2346 case Intrinsic::aarch64_sve_ldff1_gather_uxtw_index:

2347 case Intrinsic::aarch64_sve_ldnf1:

2348 case Intrinsic::aarch64_sve_ldnt1:

2349 case Intrinsic::aarch64_sve_ldnt1_gather:

2350 case Intrinsic::aarch64_sve_ldnt1_gather_index:

2351 case Intrinsic::aarch64_sve_ldnt1_gather_scalar_offset:

2352 case Intrinsic::aarch64_sve_ldnt1_gather_uxtw:

2354 case Intrinsic::aarch64_sve_prf:

2355 case Intrinsic::aarch64_sve_prfb_gather_index:

2356 case Intrinsic::aarch64_sve_prfb_gather_scalar_offset:

2357 case Intrinsic::aarch64_sve_prfb_gather_sxtw_index:

2358 case Intrinsic::aarch64_sve_prfb_gather_uxtw_index:

2359 case Intrinsic::aarch64_sve_prfd_gather_index:

2360 case Intrinsic::aarch64_sve_prfd_gather_scalar_offset:

2361 case Intrinsic::aarch64_sve_prfd_gather_sxtw_index:

2362 case Intrinsic::aarch64_sve_prfd_gather_uxtw_index:

2363 case Intrinsic::aarch64_sve_prfh_gather_index:

2364 case Intrinsic::aarch64_sve_prfh_gather_scalar_offset:

2365 case Intrinsic::aarch64_sve_prfh_gather_sxtw_index:

2366 case Intrinsic::aarch64_sve_prfh_gather_uxtw_index:

2367 case Intrinsic::aarch64_sve_prfw_gather_index:

2368 case Intrinsic::aarch64_sve_prfw_gather_scalar_offset:

2369 case Intrinsic::aarch64_sve_prfw_gather_sxtw_index:

2370 case Intrinsic::aarch64_sve_prfw_gather_uxtw_index:

2372 case Intrinsic::aarch64_neon_fmaxnm:

2373 case Intrinsic::aarch64_neon_fminnm:

2375 case Intrinsic::aarch64_sve_convert_from_svbool:

2377 case Intrinsic::aarch64_sve_dup:

2379 case Intrinsic::aarch64_sve_dup_x:

2381 case Intrinsic::aarch64_sve_cmpne:

2382 case Intrinsic::aarch64_sve_cmpne_wide:

2384 case Intrinsic::aarch64_sve_rdffr:

2386 case Intrinsic::aarch64_sve_lasta:

2387 case Intrinsic::aarch64_sve_lastb:

2389 case Intrinsic::aarch64_sve_clasta_n:

2390 case Intrinsic::aarch64_sve_clastb_n:

2392 case Intrinsic::aarch64_sve_cntd:

2394 case Intrinsic::aarch64_sve_cntw:

2396 case Intrinsic::aarch64_sve_cnth:

2398 case Intrinsic::aarch64_sve_cntb:

2400 case Intrinsic::aarch64_sve_ptest_any:

2401 case Intrinsic::aarch64_sve_ptest_first:

2402 case Intrinsic::aarch64_sve_ptest_last:

2404 case Intrinsic::aarch64_sve_fabd:

2406 case Intrinsic::aarch64_sve_fadd:

2408 case Intrinsic::aarch64_sve_fadd_u:

2410 case Intrinsic::aarch64_sve_fdiv:

2412 case Intrinsic::aarch64_sve_fmax:

2414 case Intrinsic::aarch64_sve_fmaxnm:

2416 case Intrinsic::aarch64_sve_fmin:

2418 case Intrinsic::aarch64_sve_fminnm:

2420 case Intrinsic::aarch64_sve_fmla:

2422 case Intrinsic::aarch64_sve_fmls:

2424 case Intrinsic::aarch64_sve_fmul:

2425 if (auto II_U =

2427 return II_U;

2429 case Intrinsic::aarch64_sve_fmul_u:

2431 case Intrinsic::aarch64_sve_fmulx:

2433 case Intrinsic::aarch64_sve_fnmla:

2435 case Intrinsic::aarch64_sve_fnmls:

2437 case Intrinsic::aarch64_sve_fsub:

2439 case Intrinsic::aarch64_sve_fsub_u:

2441 case Intrinsic::aarch64_sve_add:

2443 case Intrinsic::aarch64_sve_add_u:

2445 Intrinsic::aarch64_sve_mla_u>(

2446 IC, II, true);

2447 case Intrinsic::aarch64_sve_mla:

2449 case Intrinsic::aarch64_sve_mls:

2451 case Intrinsic::aarch64_sve_mul:

2452 if (auto II_U =

2454 return II_U;

2456 case Intrinsic::aarch64_sve_mul_u:

2458 case Intrinsic::aarch64_sve_sabd:

2460 case Intrinsic::aarch64_sve_smax:

2462 case Intrinsic::aarch64_sve_smin:

2464 case Intrinsic::aarch64_sve_smulh:

2466 case Intrinsic::aarch64_sve_sub:

2468 case Intrinsic::aarch64_sve_sub_u:

2470 Intrinsic::aarch64_sve_mls_u>(

2471 IC, II, true);

2472 case Intrinsic::aarch64_sve_uabd:

2474 case Intrinsic::aarch64_sve_umax:

2476 case Intrinsic::aarch64_sve_umin:

2478 case Intrinsic::aarch64_sve_umulh:

2480 case Intrinsic::aarch64_sve_asr:

2482 case Intrinsic::aarch64_sve_lsl:

2484 case Intrinsic::aarch64_sve_lsr:

2486 case Intrinsic::aarch64_sve_and:

2488 case Intrinsic::aarch64_sve_bic:

2490 case Intrinsic::aarch64_sve_eor:

2492 case Intrinsic::aarch64_sve_orr:

2494 case Intrinsic::aarch64_sve_sqsub:

2496 case Intrinsic::aarch64_sve_uqsub:

2498 case Intrinsic::aarch64_sve_tbl:

2500 case Intrinsic::aarch64_sve_uunpkhi:

2501 case Intrinsic::aarch64_sve_uunpklo:

2502 case Intrinsic::aarch64_sve_sunpkhi:

2503 case Intrinsic::aarch64_sve_sunpklo:

2505 case Intrinsic::aarch64_sve_uzp1:

2507 case Intrinsic::aarch64_sve_zip1:

2508 case Intrinsic::aarch64_sve_zip2:

2510 case Intrinsic::aarch64_sve_ld1_gather_index:

2512 case Intrinsic::aarch64_sve_st1_scatter_index:

2514 case Intrinsic::aarch64_sve_ld1:

2516 case Intrinsic::aarch64_sve_st1:

2518 case Intrinsic::aarch64_sve_sdiv:

2520 case Intrinsic::aarch64_sve_sel:

2522 case Intrinsic::aarch64_sve_srshl:

2524 case Intrinsic::aarch64_sve_dupq_lane:

2526 case Intrinsic::aarch64_sve_insr:

2528 }

2529

2530 return std::nullopt;

2531}

2532

2537 SimplifyAndSetOp) const {

2538 switch (II.getIntrinsicID()) {

2539 default:

2540 break;

2541 case Intrinsic::aarch64_neon_fcvtxn:

2542 case Intrinsic::aarch64_neon_rshrn:

2543 case Intrinsic::aarch64_neon_sqrshrn:

2544 case Intrinsic::aarch64_neon_sqrshrun:

2545 case Intrinsic::aarch64_neon_sqshrn:

2546 case Intrinsic::aarch64_neon_sqshrun:

2547 case Intrinsic::aarch64_neon_sqxtn:

2548 case Intrinsic::aarch64_neon_sqxtun:

2549 case Intrinsic::aarch64_neon_uqrshrn:

2550 case Intrinsic::aarch64_neon_uqshrn:

2551 case Intrinsic::aarch64_neon_uqxtn:

2552 SimplifyAndSetOp(&II, 0, OrigDemandedElts, UndefElts);

2553 break;

2554 }

2555

2556 return std::nullopt;

2557}

2558

2562}

2563

2566 switch (K) {

2576 else

2582 else

2584 }

2586}

2587

2588bool AArch64TTIImpl::isWideningInstruction(Type *DstTy, unsigned Opcode,

2590 Type *SrcOverrideTy) {

2591

2592

2595 cast(DstTy)->getElementCount());

2596 };

2597

2598

2599

2600

2601

2602

2604 if (useNeonVector(DstTy) || Args.size() != 2 ||

2605 (DstEltSize != 16 && DstEltSize != 32 && DstEltSize != 64))

2606 return false;

2607

2608

2609

2610

2611

2612

2613

2614

2615 Type *SrcTy = SrcOverrideTy;

2616 switch (Opcode) {

2617 case Instruction::Add:

2618 case Instruction::Sub:

2619

2620 if (isa(Args[1]) || isa(Args[1])) {

2621 if (!SrcTy)

2622 SrcTy =

2623 toVectorTy(cast(Args[1])->getOperand(0)->getType());

2624 } else

2625 return false;

2626 break;

2627 case Instruction::Mul: {

2628

2629 if ((isa(Args[0]) && isa(Args[1])) ||

2630 (isa(Args[0]) && isa(Args[1]))) {

2631 if (!SrcTy)

2632 SrcTy =

2633 toVectorTy(cast(Args[0])->getOperand(0)->getType());

2634 } else if (isa(Args[0]) || isa(Args[1])) {

2635

2636

2637

2643 return false;

2644 if (!SrcTy)

2647 } else

2648 return false;

2649 break;

2650 }

2651 default:

2652 return false;

2653 }

2654

2655

2656

2658 if (!DstTyL.second.isVector() || DstEltSize != DstTy->getScalarSizeInBits())

2659 return false;

2660

2661

2662

2663 assert(SrcTy && "Expected some SrcTy");

2665 unsigned SrcElTySize = SrcTyL.second.getScalarSizeInBits();

2666 if (!SrcTyL.second.isVector() || SrcElTySize != SrcTy->getScalarSizeInBits())

2667 return false;

2668

2669

2671 DstTyL.first * DstTyL.second.getVectorMinNumElements();

2673 SrcTyL.first * SrcTyL.second.getVectorMinNumElements();

2674

2675

2676

2677 return NumDstEls == NumSrcEls && 2 * SrcElTySize == DstEltSize;

2678}

2679

2680

2681

2682

2683

2684

2685

2687 Type *Src) {

2688

2690 (Src->isScalableTy() && !ST->hasSVE2()))

2691 return false;

2692

2693 if (ExtUser->getOpcode() != Instruction::Add || !ExtUser->hasOneUse())

2694 return false;

2695

2696

2698 auto *AddUser =

2699 dyn_cast_or_null(Add->getUniqueUndroppableUser());

2700 if (AddUser && AddUser->getOpcode() == Instruction::Add)

2701 Add = AddUser;

2702

2703 auto *Shr = dyn_cast_or_null(Add->getUniqueUndroppableUser());

2704 if (!Shr || Shr->getOpcode() != Instruction::LShr)

2705 return false;

2706

2707 auto *Trunc = dyn_cast_or_null(Shr->getUniqueUndroppableUser());

2708 if (!Trunc || Trunc->getOpcode() != Instruction::Trunc ||

2709 Src->getScalarSizeInBits() !=

2710 cast(Trunc)->getDestTy()->getScalarSizeInBits())

2711 return false;

2712

2713

2714

2718 return false;

2719

2720

2723 return true;

2724

2725 return false;

2726}

2727

2734 assert(ISD && "Invalid opcode");

2735

2736

2737 if (I && I->hasOneUser()) {

2738 auto *SingleUser = cast(*I->user_begin());

2740 if (isWideningInstruction(Dst, SingleUser->getOpcode(), Operands, Src)) {

2741

2742

2743

2744 if (SingleUser->getOpcode() == Instruction::Add) {

2745 if (I == SingleUser->getOperand(1) ||

2746 (isa(SingleUser->getOperand(1)) &&

2747 cast(SingleUser->getOperand(1))->getOpcode() == Opcode))

2748 return 0;

2749 } else

2750 return 0;

2751 }

2752

2753

2754 if ((isa(I) || isa(I)) &&

2756 return 0;

2757 }

2758

2759

2762 return Cost == 0 ? 0 : 1;

2763 return Cost;

2764 };

2765

2768

2770 return AdjustCost(

2772

2774 {ISD::FP_ROUND, MVT::bf16, MVT::f32, 1},

2775 {ISD::FP_ROUND, MVT::bf16, MVT::f64, 1},

2776 {ISD::FP_ROUND, MVT::v4bf16, MVT::v4f32, 1},

2777 {ISD::FP_ROUND, MVT::v8bf16, MVT::v8f32, 2},

2778 {ISD::FP_ROUND, MVT::v2bf16, MVT::v2f64, 2},

2779 {ISD::FP_ROUND, MVT::v4bf16, MVT::v4f64, 3},

2780 {ISD::FP_ROUND, MVT::v8bf16, MVT::v8f64, 6},

2781 };

2782

2783 if (ST->hasBF16())

2786 return AdjustCost(Entry->Cost);

2787

2789 {ISD::TRUNCATE, MVT::v2i8, MVT::v2i64, 1},

2790 {ISD::TRUNCATE, MVT::v2i16, MVT::v2i64, 1},

2791 {ISD::TRUNCATE, MVT::v2i32, MVT::v2i64, 1},

2792 {ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 1},

2793 {ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 3},

2794 {ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 1},

2795 {ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2},

2796 {ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1},

2797 {ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 1},

2798 {ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2},

2799 {ISD::TRUNCATE, MVT::v8i8, MVT::v8i64, 4},

2800 {ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 1},

2801 {ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 3},

2802 {ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 2},

2803 {ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 1},

2804 {ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 3},

2805 {ISD::TRUNCATE, MVT::v16i8, MVT::v16i64, 7},

2806 {ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 2},

2807 {ISD::TRUNCATE, MVT::v16i16, MVT::v16i64, 6},

2808 {ISD::TRUNCATE, MVT::v16i32, MVT::v16i64, 4},

2809

2810

2841 {ISD::TRUNCATE, MVT::nxv16i8, MVT::nxv16i16, 1},

2842 {ISD::TRUNCATE, MVT::nxv16i8, MVT::nxv16i32, 3},

2843 {ISD::TRUNCATE, MVT::nxv16i8, MVT::nxv16i64, 7},

2844

2845

2862

2863

2865 {ISD::FP_EXTEND, MVT::v2f64, MVT::v2f32, 1},

2866 {ISD::FP_EXTEND, MVT::v4f64, MVT::v4f32, 2},

2867

2870 {ISD::FP_EXTEND, MVT::v4f32, MVT::v4f16, 1},

2871 {ISD::FP_EXTEND, MVT::v8f32, MVT::v8f16, 2},

2872 {ISD::FP_EXTEND, MVT::v2f64, MVT::v2f16, 2},

2873 {ISD::FP_EXTEND, MVT::v4f64, MVT::v4f16, 3},

2874 {ISD::FP_EXTEND, MVT::v8f64, MVT::v8f16, 6},

2875

2877 {ISD::FP_EXTEND, MVT::f64, MVT::bf16, 2},

2878 {ISD::FP_EXTEND, MVT::v4f32, MVT::v4bf16, 1},

2879 {ISD::FP_EXTEND, MVT::v8f32, MVT::v8bf16, 2},

2880 {ISD::FP_EXTEND, MVT::v2f64, MVT::v2bf16, 2},

2881 {ISD::FP_EXTEND, MVT::v4f64, MVT::v4bf16, 3},

2882 {ISD::FP_EXTEND, MVT::v8f64, MVT::v8bf16, 6},

2883

2885 {ISD::FP_ROUND, MVT::v2f32, MVT::v2f64, 1},

2886 {ISD::FP_ROUND, MVT::v4f32, MVT::v4f64, 2},

2887

2890 {ISD::FP_ROUND, MVT::v4f16, MVT::v4f32, 1},

2891 {ISD::FP_ROUND, MVT::v8f16, MVT::v8f32, 2},

2892 {ISD::FP_ROUND, MVT::v2f16, MVT::v2f64, 2},

2893 {ISD::FP_ROUND, MVT::v4f16, MVT::v4f64, 3},

2894 {ISD::FP_ROUND, MVT::v8f16, MVT::v8f64, 6},

2895

2896 {ISD::FP_ROUND, MVT::bf16, MVT::f32, 8},

2897 {ISD::FP_ROUND, MVT::bf16, MVT::f64, 9},

2904

2905

2912

2913

2920

2921

2926

2927

2932

2933

2936

2937

2944

2945

2948

2949

2956

2957

2964

2965

2970

2971

2980

2981

2988

2989

2998

2999

3008

3009

3014

3015

3022

3023

3028

3029

3038

3039

3048

3049

3058

3059

3063

3064

3068

3069

3073

3074

3078

3079

3083

3084

3088

3089

3090 {ISD::BITCAST, MVT::nxv2f16, MVT::nxv2i16, 0},

3091 {ISD::BITCAST, MVT::nxv4f16, MVT::nxv4i16, 0},

3092 {ISD::BITCAST, MVT::nxv2f32, MVT::nxv2i32, 0},

3093

3094

3095 {ISD::BITCAST, MVT::nxv2i16, MVT::nxv2f16, 0},

3096 {ISD::BITCAST, MVT::nxv4i16, MVT::nxv4f16, 0},

3097 {ISD::BITCAST, MVT::nxv2i32, MVT::nxv2f32, 0},

3098

3099

3100

3101

3108

3115 };

3116

3117

3118

3119

3120 EVT WiderTy = SrcTy.bitsGT(DstTy) ? SrcTy : DstTy;

3124 std::pair<InstructionCost, MVT> LT =

3126 unsigned NumElements =

3128 return AdjustCost(

3129 LT.first *

3134 }

3135

3138 return AdjustCost(Entry->Cost);

3139

3145 {ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f16, 2},

3147 {ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f16, 2},

3151 {ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f16, 4},

3153 {ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f16, 3},

3155 {ISD::FP_TO_SINT, MVT::v16i16, MVT::v16f16, 2},

3157 {ISD::FP_TO_SINT, MVT::v16i32, MVT::v16f16, 8},

3159 {ISD::UINT_TO_FP, MVT::v8f16, MVT::v8i8, 2},

3160 {ISD::SINT_TO_FP, MVT::v8f16, MVT::v8i8, 2},

3161 {ISD::UINT_TO_FP, MVT::v16f16, MVT::v16i8, 4},

3162 {ISD::SINT_TO_FP, MVT::v16f16, MVT::v16i8, 4},

3163 };

3164

3165 if (ST->hasFullFP16())

3168 return AdjustCost(Entry->Cost);

3169

3177

3178

3179

3180

3184 Opcode, LegalTy, Src, CCH, CostKind, I);

3187 return Part1 + Part2;

3188 }

3189

3190

3191

3196

3197 return AdjustCost(

3199}

3200

3204 unsigned Index) {

3205

3206

3207 assert((Opcode == Instruction::SExt || Opcode == Instruction::ZExt) &&

3208 "Invalid opcode");

3209

3210

3211

3213

3214

3215 assert(isa(Dst) && isa(Src) && "Invalid type");

3216

3217

3218

3221 CostKind, Index, nullptr, nullptr);

3222

3223

3227

3228

3229

3230

3231 if (!VecLT.second.isVector() || !TLI->isTypeLegal(DstVT))

3234

3235

3236

3237 if (DstVT.getFixedSizeInBits() < SrcVT.getFixedSizeInBits())

3240

3241 switch (Opcode) {

3242 default:

3244

3245

3246

3247 case Instruction::SExt:

3248 return Cost;

3249

3250

3251

3252 case Instruction::ZExt:

3253 if (DstVT.getSizeInBits() != 64u || SrcVT.getSizeInBits() == 32u)

3254 return Cost;

3255 }

3256

3257

3260}

3261

3266 return Opcode == Instruction::PHI ? 0 : 1;

3268

3269 return 0;

3270}

3271

3272InstructionCost AArch64TTIImpl::getVectorInstrCostHelper(

3273 unsigned Opcode, Type *Val, unsigned Index, bool HasRealUse,

3275 ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) {

3277

3278 if (Index != -1U) {

3279

3281

3282

3283 if (!LT.second.isVector())

3284 return 0;

3285

3286

3287

3288 if (LT.second.isFixedLengthVector()) {

3289 unsigned Width = LT.second.getVectorNumElements();

3290 Index = Index % Width;

3291 }

3292

3293

3294

3295

3296

3297

3298

3300 return 0;

3301

3302

3303

3304

3305

3306 if (I && dyn_cast(I->getOperand(1)))

3308

3309

3310

3313

3314

3315

3316

3317

3318

3319 }

3320

3321

3322

3323

3324

3325

3326

3327

3328

3329

3330

3331

3332

3333

3334

3335

3336 auto ExtractCanFuseWithFmul = [&]() {

3337

3338 if (Index == 0)

3339 return false;

3340

3341

3342

3343 auto IsAllowedScalarTy = [&](const Type *T) {

3344 return T->isFloatTy() || T->isDoubleTy() ||

3345 (T->isHalfTy() && ST->hasFullFP16());

3346 };

3347

3348

3349 auto IsUserFMulScalarTy = [](const Value *EEUser) {

3350

3351 const auto *BO = dyn_cast(EEUser);

3352 return BO && BO->getOpcode() == BinaryOperator::FMul &&

3353 !BO->getType()->isVectorTy();

3354 };

3355

3356

3357

3358 auto IsExtractLaneEquivalentToZero = [&](unsigned Idx, unsigned EltSz) {

3359 auto RegWidth =

3362 return Idx == 0 || (RegWidth != 0 && (Idx * EltSz) % RegWidth == 0);

3363 };

3364

3365

3366

3367 if (!isa(Val) || !IsAllowedScalarTy(Val->getScalarType()))

3368 return false;

3369

3370 if (Scalar) {

3372 for (auto *U : Scalar->users()) {

3373 if (!IsUserFMulScalarTy(U))

3374 return false;

3375

3376

3377 UserToExtractIdx[U];

3378 }

3379 if (UserToExtractIdx.empty())

3380 return false;

3381 for (auto &[S, U, L] : ScalarUserAndIdx) {

3382 for (auto *U : S->users()) {

3383 if (UserToExtractIdx.find(U) != UserToExtractIdx.end()) {

3384 auto *FMul = cast(U);

3385 auto *Op0 = FMul->getOperand(0);

3386 auto *Op1 = FMul->getOperand(1);

3387 if ((Op0 == S && Op1 == S) || Op0 != S || Op1 != S) {

3388 UserToExtractIdx[U] = L;

3389 break;

3390 }

3391 }

3392 }

3393 }

3394 for (auto &[U, L] : UserToExtractIdx) {

3395 if (!IsExtractLaneEquivalentToZero(Index, Val->getScalarSizeInBits()) &&

3397 return false;

3398 }

3399 } else {

3400 const auto *EE = cast(I);

3401

3402 const auto *IdxOp = dyn_cast(EE->getIndexOperand());

3403 if (!IdxOp)

3404 return false;

3405

3406 return !EE->users().empty() && all_of(EE->users(), [&](const User *U) {

3407 if (!IsUserFMulScalarTy(U))

3408 return false;

3409

3410

3411

3412 const auto *BO = cast(U);

3413 const auto *OtherEE = dyn_cast(

3414 BO->getOperand(0) == EE ? BO->getOperand(1) : BO->getOperand(0));

3415 if (OtherEE) {

3416 const auto *IdxOp = dyn_cast(OtherEE->getIndexOperand());

3417 if (!IdxOp)

3418 return false;

3419 return IsExtractLaneEquivalentToZero(

3420 cast(OtherEE->getIndexOperand())

3421 ->getValue()

3422 .getZExtValue(),

3423 OtherEE->getType()->getScalarSizeInBits());

3424 }

3425 return true;

3426 });

3427 }

3428 return true;

3429 };

3430

3431 if (Opcode == Instruction::ExtractElement && (I || Scalar) &&

3432 ExtractCanFuseWithFmul())

3433 return 0;

3434

3435

3436 return ST->getVectorInsertExtractBaseCost();

3437}

3438

3441 unsigned Index, Value *Op0,

3443 bool HasRealUse =

3444 Opcode == Instruction::InsertElement && Op0 && !isa(Op0);

3445 return getVectorInstrCostHelper(Opcode, Val, Index, HasRealUse);

3446}

3447

3451 ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx) {

3452 return getVectorInstrCostHelper(Opcode, Val, Index, false, nullptr, Scalar,

3453 ScalarUserAndIdx);

3454}

3455

3459 unsigned Index) {

3460 return getVectorInstrCostHelper(I.getOpcode(), Val, Index,

3461 true , &I);

3462}

3463

3465 VectorType *Ty, const APInt &DemandedElts, bool Insert, bool Extract,

3467 if (isa(Ty))

3472 return DemandedElts.popcount() * (Insert + Extract) *

3474}

3475

3481

3482

3483

3484

3485

3486 if (auto *VTy = dyn_cast(Ty))

3489

3490

3493 Op2Info, Args, CxtI);

3494

3495

3498

3499 switch (ISD) {

3500 default:

3502 Op2Info);

3505

3506

3507

3508

3510 Instruction::Add, Ty, CostKind,

3515 Instruction::Select, Ty, CostKind,

3519 return Cost;

3520 }

3521 [[fallthrough]];

3526

3527

3528

3535 return MulCost * 2 + AddCost * 2 + ShrCost * 2 + 1;

3536 }

3537 }

3538

3539

3540

3541

3542 if (!VT.isVector() && VT.getSizeInBits() > 64)

3544

3546 Opcode, Ty, CostKind, Op1Info, Op2Info);

3549

3550

3551 if (isa(Ty) && cast(Ty)

3552 ->getPrimitiveSizeInBits()

3553 .getFixedValue() < 128) {

3562

3564 if (nullptr != Entry)

3565 return Entry->Cost;

3566 }

3567

3568

3569 if (LT.second.getScalarType() == MVT::i8)

3571 else if (LT.second.getScalarType() == MVT::i16)

3573 return Cost;

3574 } else {

3575

3576

3577

3578

3581 if (auto *VTy = dyn_cast(Ty)) {

3584 return (4 + DivCost) * VTy->getNumElements();

3585 }

3586 }

3587

3588

3590 CostKind, Op1Info, Op2Info);

3592 Op1Info, Op2Info);

3593 }

3594

3595

3596

3598 }

3599 return Cost;

3600 }

3602

3603

3604 if (LT.second == MVT::v2i64 && ST->hasSVE())

3605 return LT.first;

3606

3607

3608

3609

3610

3611

3612

3613

3614

3615

3616

3617

3618

3619 if (LT.second != MVT::v2i64 || isWideningInstruction(Ty, Opcode, Args))

3620 return LT.first;

3621 return cast(Ty)->getElementCount().getKnownMinValue() *

3624 nullptr, nullptr) *

3625 2 +

3627 nullptr, nullptr));

3635

3636

3637 return LT.first;

3638

3640

3642 (Ty->isHalfTy() && ST->hasFullFP16())) &&

3643 CxtI &&

3647 return 0;

3648 [[fallthrough]];

3651

3652

3655 return 2 * LT.first;

3657 return LT.first;

3658 [[fallthrough]];

3661

3662

3664 return 2 * LT.first;

3665

3667 Op2Info);

3669

3670

3674 Op2Info);

3675 }

3676}

3677

3681

3682

3683

3684

3686 int MaxMergeDistance = 64;

3687

3690 return NumVectorInstToHideOverhead;

3691

3692

3693

3694 return 1;

3695}

3696

3701

3704 Op1Info, Op2Info, I);

3705

3707

3708

3709 if (isa(ValTy) && ISD == ISD::SELECT) {

3710

3711 const int AmortizationCost = 20;

3712

3713

3714

3719 VecPred = CurrentPred;

3720 }

3721

3722

3727 static const auto ValidMinMaxTys = {

3728 MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16, MVT::v2i32,

3729 MVT::v4i32, MVT::v2i64, MVT::v2f32, MVT::v4f32, MVT::v2f64};

3730 static const auto ValidFP16MinMaxTys = {MVT::v4f16, MVT::v8f16};

3731

3733 if (any_of(ValidMinMaxTys, [&LT](MVT M) { return M == LT.second; }) ||

3734 (ST->hasFullFP16() &&

3735 any_of(ValidFP16MinMaxTys, [&LT](MVT M) { return M == LT.second; })))

3736 return LT.first;

3737 }

3738

3740 VectorSelectTbl[] = {

3741 { ISD::SELECT, MVT::v2i1, MVT::v2f32, 2 },

3742 { ISD::SELECT, MVT::v2i1, MVT::v2f64, 2 },

3743 { ISD::SELECT, MVT::v4i1, MVT::v4f32, 2 },

3744 { ISD::SELECT, MVT::v4i1, MVT::v4f16, 2 },

3745 { ISD::SELECT, MVT::v8i1, MVT::v8f16, 2 },

3746 { ISD::SELECT, MVT::v16i1, MVT::v16i16, 16 },

3747 { ISD::SELECT, MVT::v8i1, MVT::v8i32, 8 },

3748 { ISD::SELECT, MVT::v16i1, MVT::v16i32, 16 },

3749 { ISD::SELECT, MVT::v4i1, MVT::v4i64, 4 * AmortizationCost },

3750 { ISD::SELECT, MVT::v8i1, MVT::v8i64, 8 * AmortizationCost },

3751 { ISD::SELECT, MVT::v16i1, MVT::v16i64, 16 * AmortizationCost }

3752 };

3753

3760 return Entry->Cost;

3761 }

3762 }

3763

3764 if (isa(ValTy) && ISD == ISD::SETCC) {

3766

3767 if (LT.second == MVT::v4f16 && !ST->hasFullFP16())

3768 return LT.first * 4;

3769 }

3770

3771

3772

3773

3779 return 0;

3780

3781

3782

3784 Op1Info, Op2Info, I);

3785}

3786

3790 if (ST->requiresStrictAlign()) {

3791

3792

3794 }

3795 Options.AllowOverlappingLoads = true;

3798

3799

3800

3801 Options.LoadSizes = {8, 4, 2, 1};

3802 Options.AllowedTailExpansions = {3, 5, 6};

3804}

3805

3807 return ST->hasSVE();

3808}

3809

3818 if (!LT.first.isValid())

3820

3821

3822 auto *VT = cast(Src);

3823 if (VT->getElementType()->isIntegerTy(1))

3825

3826

3827

3828

3829

3832

3833 return LT.first;

3834}

3835

3836

3837

3840 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&

3841 "Should be called on only load or stores.");

3842 switch (Opcode) {

3843 case Instruction::Load:

3846 return ST->getGatherOverhead();

3847 break;

3848 case Instruction::Store:

3851 return ST->getScatterOverhead();

3852 break;

3853 default:

3855 }

3856}

3857

3859 unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask,

3864 auto *VT = cast(DataTy);

3866 if (!LT.first.isValid())

3868

3869

3870 if (!LT.second.isVector() ||

3872 VT->getElementType()->isIntegerTy(1))

3874

3875

3876

3877

3878

3881

3882 ElementCount LegalVF = LT.second.getVectorElementCount();

3885 {TTI::OK_AnyValue, TTI::OP_None}, I);

3886

3889}

3890

3893}

3894

3902

3903 if (VT == MVT::Other)

3906

3908 if (!LT.first.isValid())

3910

3911

3912

3913

3914

3915

3916 if (auto *VTy = dyn_cast(Ty))

3918 (VTy->getElementType()->isIntegerTy(1) &&

3919 !VTy->getElementCount().isKnownMultipleOf(

3922

3923

3925 return LT.first;

3926

3928 return 1;

3929

3930 if (ST->isMisaligned128StoreSlow() && Opcode == Instruction::Store &&

3931 LT.second.is128BitVector() && (!Alignment || *Alignment < Align(16))) {

3932

3933

3934

3935

3936

3937 const int AmortizationCost = 6;

3938

3939 return LT.first * 2 * AmortizationCost;

3940 }

3941

3942

3944 return LT.first;

3945

3947

3949

3950 if (VT == MVT::v4i8)

3951 return 2;

3952

3953 return cast(Ty)->getNumElements() * 2;

3954 }

3957 if (isPowerOf2\_32(EltSize) || EltSize < 8 || EltSize > 64 ||

3959 *Alignment != Align(1))

3960 return LT.first;

3961

3962

3964 return LT.first;

3965

3966

3967

3968

3973 while (!TypeWorklist.empty()) {

3978 continue;

3979 }

3980

3981 unsigned PrevPow2 = NextPowerOf2(CurrNumElements) / 2;

3985 }

3986 return Cost;

3987 }

3988

3989 return LT.first;

3990}

3991

3995 bool UseMaskForCond, bool UseMaskForGaps) {

3996 assert(Factor >= 2 && "Invalid interleave factor");

3997 auto *VecVTy = cast(VecTy);

3998

4001

4002

4003

4004 if (!VecTy->isScalableTy() && (UseMaskForCond || UseMaskForGaps))

4006

4007 if (!UseMaskForGaps && Factor <= TLI->getMaxSupportedInterleaveFactor()) {

4008 unsigned MinElts = VecVTy->getElementCount().getKnownMinValue();

4009 auto *SubVecTy =

4011 VecVTy->getElementCount().divideCoefficientBy(Factor));

4012

4013

4014

4015

4016 bool UseScalable;

4017 if (MinElts % Factor == 0 &&

4020 }

4021

4024 UseMaskForCond, UseMaskForGaps);

4025}

4026

4031 for (auto *I : Tys) {

4032 if (I->isVectorTy())

4033 continue;

4034 if (I->getScalarSizeInBits() * cast(I)->getNumElements() ==

4035 128)

4038 }

4039 return Cost;

4040}

4041

4044}

4045

4046

4047

4048

4049

4050static void

4053 enum { MaxStridedLoads = 7 };

4055 int StridedLoads = 0;

4056

4057

4058 for (const auto BB : L->blocks()) {

4059 for (auto &I : *BB) {

4060 LoadInst *LMemI = dyn_cast(&I);

4061 if (!LMemI)

4062 continue;

4063

4065 if (L->isLoopInvariant(PtrValue))

4066 continue;

4067

4068 const SCEV *LSCEV = SE.getSCEV(PtrValue);

4069 const SCEVAddRecExpr *LSCEVAddRec = dyn_cast(LSCEV);

4070 if (!LSCEVAddRec || !LSCEVAddRec->isAffine())

4071 continue;

4072

4073

4074

4075

4076 ++StridedLoads;

4077

4078

4079 if (StridedLoads > MaxStridedLoads / 2)

4080 return StridedLoads;

4081 }

4082 }

4083 return StridedLoads;

4084 };

4085

4086 int StridedLoads = countStridedLoads(L, SE);

4087 LLVM_DEBUG(dbgs() << "falkor-hwpf: detected " << StridedLoads

4088 << " strided loads\n");

4089

4090

4091 if (StridedLoads) {

4092 UP.MaxCount = 1 << Log2_32(MaxStridedLoads / StridedLoads);

4093 LLVM_DEBUG(dbgs() << "falkor-hwpf: setting unroll MaxCount to "

4095 }

4096}

4097

4098

4099

4100static void

4104

4105

4106

4107

4108

4109

4110 if (!L->isInnermost() || !L->getExitBlock() || L->getNumBlocks() > 8)

4111 return;

4112

4114 if (isa(BTC) || isa(BTC) ||

4117 return;

4119 return;

4120

4121 int64_t Size = 0;

4122 for (auto *BB : L->getBlocks()) {

4123 for (auto &I : *BB) {

4124 if (!isa(&I) && isa(&I))

4125 return;

4129 }

4130 }

4131

4132

4134

4135

4136

4137 BasicBlock *Header = L->getHeader();

4138 if (Header == L->getLoopLatch()) {

4139 if (Size > 8)

4140 return;

4141

4144 for (auto *BB : L->blocks()) {

4145 for (auto &I : *BB) {

4147 if (Ptr)

4148 continue;

4151 continue;

4152 if (isa(&I))

4153 LoadedValues.insert(&I);

4154 else

4155 Stores.push_back(cast(&I));

4156 }

4157 }

4158

4159

4160

4161 unsigned MaxInstsPerLine = 16;

4162 unsigned UC = 1;

4163 unsigned BestUC = 1;

4164 unsigned SizeWithBestUC = BestUC * Size;

4165 while (UC <= 8) {

4166 unsigned SizeWithUC = UC * Size;

4167 if (SizeWithUC > 48)

4168 break;

4169 if ((SizeWithUC % MaxInstsPerLine) == 0 ||

4170 (SizeWithBestUC % MaxInstsPerLine) < (SizeWithUC % MaxInstsPerLine)) {

4171 BestUC = UC;

4172 SizeWithBestUC = BestUC * Size;

4173 }

4174 UC++;

4175 }

4176

4177 if (BestUC == 1 || none_of(Stores, [&LoadedValues](StoreInst *SI) {

4178 return LoadedValues.contains(SI->getOperand(0));

4179 }))

4180 return;

4181

4184 return;

4185 }

4186

4187

4188

4189 auto *Term = dyn_cast(Header->getTerminator());

4190 auto *Latch = L->getLoopLatch();

4192 if (!Term || !Term->isConditional() || Preds.size() == 1 ||

4193 none_of(Preds, [Header](BasicBlock *Pred) { return Header == Pred; }) ||

4194 none_of(Preds, [L](BasicBlock *Pred) { return L->contains(Pred); }))

4195 return;

4196

4197 std::function<bool(Instruction *, unsigned)> DependsOnLoopLoad =

4199 if (isa(I) || L->isLoopInvariant(I) || Depth > 8)

4200 return false;

4201

4202 if (isa(I))

4203 return true;

4204

4205 return any_of(I->operands(), [&](Value *V) {

4206 auto *I = dyn_cast(V);

4207 return I && DependsOnLoopLoad(I, Depth + 1);

4208 });

4209 };

4214 DependsOnLoopLoad(I, 0)) {

4216 }

4217}

4218

4222

4224

4226

4227

4228

4229

4230 if (L->getLoopDepth() > 1)

4232

4233

4235

4236

4238 case AArch64Subtarget::AppleA14:

4239 case AArch64Subtarget::AppleA15:

4240 case AArch64Subtarget::AppleA16:

4241 case AArch64Subtarget::AppleM4:

4243 break;

4244 case AArch64Subtarget::Falkor:

4247 break;

4248 default:

4249 break;

4250 }

4251

4252

4253

4254

4255 for (auto *BB : L->getBlocks()) {

4256 for (auto &I : *BB) {

4257

4258 if (I.getType()->isVectorTy())

4259 return;

4260

4261 if (isa(I) || isa(I)) {

4264 continue;

4265 }

4266 return;

4267 }

4268 }

4269 }

4270

4271

4272

4273

4274

4276 !ST->getSchedModel().isOutOfOrder()) {

4281

4284 }

4285}

4286

4290}

4291

4293 Type *ExpectedType) {

4295 default:

4296 return nullptr;

4297 case Intrinsic::aarch64_neon_st2:

4298 case Intrinsic::aarch64_neon_st3:

4299 case Intrinsic::aarch64_neon_st4: {

4300

4301 StructType *ST = dyn_cast(ExpectedType);

4302 if (!ST)

4303 return nullptr;

4304 unsigned NumElts = Inst->arg_size() - 1;

4305 if (ST->getNumElements() != NumElts)

4306 return nullptr;

4307 for (unsigned i = 0, e = NumElts; i != e; ++i) {

4309 return nullptr;

4310 }

4313 for (unsigned i = 0, e = NumElts; i != e; ++i) {

4316 }

4317 return Res;

4318 }

4319 case Intrinsic::aarch64_neon_ld2:

4320 case Intrinsic::aarch64_neon_ld3:

4321 case Intrinsic::aarch64_neon_ld4:

4322 if (Inst->getType() == ExpectedType)

4323 return Inst;

4324 return nullptr;

4325 }

4326}

4327

4331 default:

4332 break;

4333 case Intrinsic::aarch64_neon_ld2:

4334 case Intrinsic::aarch64_neon_ld3:

4335 case Intrinsic::aarch64_neon_ld4:

4336 Info.ReadMem = true;

4337 Info.WriteMem = false;

4339 break;

4340 case Intrinsic::aarch64_neon_st2:

4341 case Intrinsic::aarch64_neon_st3:

4342 case Intrinsic::aarch64_neon_st4:

4343 Info.ReadMem = false;

4344 Info.WriteMem = true;

4346 break;

4347 }

4348

4350 default:

4351 return false;

4352 case Intrinsic::aarch64_neon_ld2:

4353 case Intrinsic::aarch64_neon_st2:

4354 Info.MatchingId = VECTOR_LDST_TWO_ELEMENTS;

4355 break;

4356 case Intrinsic::aarch64_neon_ld3:

4357 case Intrinsic::aarch64_neon_st3:

4358 Info.MatchingId = VECTOR_LDST_THREE_ELEMENTS;

4359 break;

4360 case Intrinsic::aarch64_neon_ld4:

4361 case Intrinsic::aarch64_neon_st4:

4362 Info.MatchingId = VECTOR_LDST_FOUR_ELEMENTS;

4363 break;

4364 }

4365 return true;

4366}

4367

4368

4369

4370

4371

4372

4374 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) {

4375 bool Considerable = false;

4376 AllowPromotionWithoutCommonHeader = false;

4377 if (!isa(&I))

4378 return false;

4379 Type *ConsideredSExtType =

4381 if (I.getType() != ConsideredSExtType)

4382 return false;

4383

4384

4385 for (const User *U : I.users()) {

4386 if (const GetElementPtrInst *GEPInst = dyn_cast(U)) {

4387 Considerable = true;

4388

4389

4390

4391 if (GEPInst->getNumOperands() > 2) {

4392 AllowPromotionWithoutCommonHeader = true;

4393 break;

4394 }

4395 }

4396 }

4397 return Considerable;

4398}

4399

4403 return true;

4404

4407 return false;

4408

4424 return true;

4425 default:

4426 return false;

4427 }

4428}

4429

4434

4435

4436

4437

4438 if (auto *VTy = dyn_cast(Ty))

4441

4443

4444 if (LT.second.getScalarType() == MVT::f16 && !ST->hasFullFP16())

4446

4448 if (LT.first > 1) {

4452 }

4453

4454 return LegalizationCost + 2;

4455}

4456

4461 if (LT.first > 1) {

4464 LegalizationCost *= LT.first - 1;

4465 }

4466

4468 assert(ISD && "Invalid opcode");

4469

4470 switch (ISD) {

4476 return LegalizationCost + 2;

4477 default:

4479 }

4480}

4481

4484 std::optional FMF,

4486

4487

4488

4489

4490 if (auto *VTy = dyn_cast(ValTy))

4493

4495 if (auto *FixedVTy = dyn_cast(ValTy)) {

4498

4499

4500 return BaseCost + FixedVTy->getNumElements();

4501 }

4502

4503 if (Opcode != Instruction::FAdd)

4505

4506 auto *VTy = cast(ValTy);

4510 return Cost;

4511 }

4512

4513 if (isa(ValTy))

4515

4517 MVT MTy = LT.second;

4519 assert(ISD && "Invalid opcode");

4520

4521

4522

4523

4524

4525

4526

4527

4528

4529 static const CostTblEntry CostTblNoPairwise[]{

4536 {ISD::OR, MVT::v8i8, 15},

4537 {ISD::OR, MVT::v16i8, 17},

4538 {ISD::OR, MVT::v4i16, 7},

4539 {ISD::OR, MVT::v8i16, 9},

4540 {ISD::OR, MVT::v2i32, 3},

4541 {ISD::OR, MVT::v4i32, 5},

4542 {ISD::OR, MVT::v2i64, 3},

4557 };

4558 switch (ISD) {

4559 default:

4560 break;

4563

4564

4565 MTy.isVector() && (EltTy->isFloatTy() || EltTy->isDoubleTy() ||

4566 (EltTy->isHalfTy() && ST->hasFullFP16()))) {

4570

4571

4572

4573

4574

4575

4576

4577

4578 return (LT.first - 1) + Log2_32(NElts);

4579 }

4580 break;

4582 if (const auto *Entry = CostTableLookup(CostTblNoPairwise, ISD, MTy))

4583 return (LT.first - 1) + Entry->Cost;

4584 break;

4588 const auto *Entry = CostTableLookup(CostTblNoPairwise, ISD, MTy);

4589 if (!Entry)

4590 break;

4591 auto *ValVTy = cast(ValTy);

4595 if (LT.first != 1) {

4596

4597

4601 ExtraCost *= LT.first - 1;

4602 }

4603

4604 auto Cost = ValVTy->getElementType()->isIntegerTy(1) ? 2 : Entry->Cost;

4605 return Cost + ExtraCost;

4606 }

4607 break;

4608 }

4610}

4611

4627 };

4628

4629

4630

4631

4632

4635

4639 EVT PromotedVT = LT.second.getScalarType() == MVT::i1

4641 : LT.second;

4644 if (Index < 0) {

4645 LegalizationCost =

4650 }

4651

4652

4653

4654 if (LT.second.getScalarType() == MVT::i1) {

4655 LegalizationCost +=

4660 }

4661 const auto *Entry =

4663 assert(Entry && "Illegal Type for Splice");

4664 LegalizationCost += Entry->Cost;

4665 return LegalizationCost * LT.first;

4666}

4667

4669 unsigned Opcode, Type *InputTypeA, Type *InputTypeB, Type *AccumType,

4672 std::optional BinOp) const {

4675

4676 if (Opcode != Instruction::Add)

4678

4679 if (InputTypeA != InputTypeB)

4681

4684

4689

4690 if (InputEVT == MVT::i8) {

4692 default:

4694 case 8:

4695 if (AccumEVT == MVT::i32)

4697 else if (AccumEVT != MVT::i64)

4699 break;

4700 case 16:

4701 if (AccumEVT == MVT::i64)

4703 else if (AccumEVT != MVT::i32)

4705 break;

4706 }

4707 } else if (InputEVT == MVT::i16) {

4708

4709

4712 } else

4714

4715

4716

4718 (OpAExtend != OpBExtend && !ST->hasMatMulInt8() &&

4721

4722 if (!BinOp || *BinOp != Instruction::Mul)

4724

4725 return Cost;

4726}

4727

4733

4734

4735

4736 if (!Mask.empty() && isa(Tp) && LT.second.isVector() &&

4738 Mask.size() > LT.second.getVectorNumElements() && !Index && !SubTp) {

4739

4740

4741

4742

4743

4744 if (Args.size() >= 1 && isa(Args[0]) &&

4747 return std::max(1, LT.first / 4);

4748

4749

4750

4751

4752

4758 return LT.first;

4759

4760 unsigned TpNumElts = Mask.size();

4761 unsigned LTNumElts = LT.second.getVectorNumElements();

4762 unsigned NumVecs = (TpNumElts + LTNumElts - 1) / LTNumElts;

4766 for (unsigned N = 0; N < NumVecs; N++) {

4768

4769

4770 unsigned Source1, Source2;

4771 unsigned NumSources = 0;

4772 for (unsigned E = 0; E < LTNumElts; E++) {

4773 int MaskElt = (N * LTNumElts + E < TpNumElts) ? Mask[N * LTNumElts + E]

4775 if (MaskElt < 0) {

4777 continue;

4778 }

4779

4780

4781

4782 unsigned Source = MaskElt / LTNumElts;

4783 if (NumSources == 0) {

4784 Source1 = Source;

4785 NumSources = 1;

4786 } else if (NumSources == 1 && Source != Source1) {

4787 Source2 = Source;

4788 NumSources = 2;

4789 } else if (NumSources >= 2 && Source != Source1 && Source != Source2) {

4790 NumSources++;

4791 }

4792

4793

4794

4795 if (Source == Source1)

4796 NMask.push_back(MaskElt % LTNumElts);

4797 else if (Source == Source2)

4798 NMask.push_back(MaskElt % LTNumElts + LTNumElts);

4799 else

4800 NMask.push_back(MaskElt % LTNumElts);

4801 }

4802

4803

4804

4805 if (NumSources <= 2)

4808 NTp, NMask, CostKind, 0, nullptr, Args, CxtI);

4809 else

4810 Cost += LTNumElts;

4811 }

4812 return Cost;

4813 }

4814

4817

4818

4819

4820 if (IsExtractSubvector && LT.second.isFixedLengthVector()) {

4821 if (LT.second.is128BitVector() &&

4822 cast(SubTp)->getNumElements() ==

4823 LT.second.getVectorNumElements() / 2) {

4824 if (Index == 0)

4825 return 0;

4826 if (Index == (int)LT.second.getVectorNumElements() / 2)

4827 return 1;

4828 }

4830 }

4831

4832

4833

4834

4835

4836

4837

4838

4840 bool IsLoad = !Args.empty() && isa(Args[0]);

4841 if (IsLoad && LT.second.isVector() &&

4843 LT.second.getVectorElementCount()))

4844 return 0;

4845 }

4846

4847

4848

4851 all_of(Mask, [](int E) { return E < 8; }))

4853

4854

4855 if (!Mask.empty() && LT.second.isFixedLengthVector() &&

4858 return M.value() < 0 || M.value() == (int)M.index();

4859 }))

4860 return 0;

4861

4862

4863

4864 unsigned Unused;

4865 if (LT.second.isFixedLengthVector() &&

4866 LT.second.getVectorNumElements() == Mask.size() &&

4868 (isZIPMask(Mask, LT.second.getVectorNumElements(), Unused) ||

4869 isUZPMask(Mask, LT.second.getVectorNumElements(), Unused) ||

4870

4872 [&Mask](int M) { return M < 0 || M == Mask[0]; })))

4873 return 1;

4874

4879

4892

4893

4906

4907

4909 {TTI::SK_Select, MVT::v4i32, 2},

4912 {TTI::SK_Select, MVT::v4f32, 2},

4914

4929

4942

4957

4975

4993 };

4994 if (const auto *Entry = CostTableLookup(ShuffleTbl, Kind, LT.second))

4995 return LT.first * Entry->Cost;

4996 }

4997

4998 if (Kind == TTI::SK_Splice && isa(Tp))

5000

5001

5002

5004 LT.second.getSizeInBits() <= 128 && SubTp) {

5006 if (SubLT.second.isVector()) {

5007 int NumElts = LT.second.getVectorNumElements();

5008 int NumSubElts = SubLT.second.getVectorNumElements();

5009 if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0)

5010 return SubLT.first;

5011 }

5012 }

5013

5014

5015 if (IsExtractSubvector)

5018 CxtI);

5019}

5020

5025

5026

5028 if (isa(&I) || isa(&I)) {

5031 if (getPtrStride(*PSE, AccessTy, Ptr, TheLoop, Strides, true,

5032 false)

5033 .value_or(0) < 0)

5034 return true;

5035 }

5036 }

5037 }

5038 return false;

5039}

5040

5044 return ST->useFixedOverScalableIfEqualCost();

5045}

5046

5049}

5050

5052 if (!ST->hasSVE())

5053 return false;

5054

5055

5056

5057

5059 return false;

5060

5066

5067

5068

5069

5075

5077 Required))

5078 return false;

5079

5080

5081

5082 unsigned NumInsns = 0;

5084 NumInsns += BB->sizeWithoutDebug();

5085 }

5086

5087

5089}

5090

5093 StackOffset BaseOffset, bool HasBaseReg,

5094 int64_t Scale, unsigned AddrSpace) const {

5095

5096

5097

5098

5099

5100

5101

5106 AM.Scale = Scale;

5109

5110

5111 return AM.Scale != 0 && AM.Scale != 1;

5112 return -1;

5113}

5114

5117

5118

5119

5120

5121 if (I->getOpcode() == Instruction::Or &&

5122 isa(I->getNextNode()) &&

5123 cast(I->getNextNode())->isUnconditional())

5124 return true;

5125

5126 if (I->getOpcode() == Instruction::Add ||

5127 I->getOpcode() == Instruction::Sub)

5128 return true;

5129 }

5131}

5132

5135

5136

5137

5138

5139

5145

5147}

5148

5150 if (auto *Shuf = dyn_cast(V))

5151 return all_equal(Shuf->getShuffleMask());

5152 return false;

5153}

5154

5155

5156

5158 bool AllowSplat = false) {

5159

5161 return false;

5162

5163 auto areTypesHalfed = [](Value *FullV, Value *HalfV) {

5164 auto *FullTy = FullV->getType();

5165 auto *HalfTy = HalfV->getType();

5167 2 * HalfTy->getPrimitiveSizeInBits().getFixedValue();

5168 };

5169

5170 auto extractHalf = [](Value *FullV, Value *HalfV) {

5171 auto *FullVT = cast(FullV->getType());

5172 auto *HalfVT = cast(HalfV->getType());

5173 return FullVT->getNumElements() == 2 * HalfVT->getNumElements();

5174 };

5175

5177 Value *S1Op1 = nullptr, *S2Op1 = nullptr;

5180 return false;

5181

5182

5183

5185 S1Op1 = nullptr;

5187 S2Op1 = nullptr;

5188

5189

5190

5191 if ((S1Op1 && (!areTypesHalfed(S1Op1, Op1) || !extractHalf(S1Op1, Op1))) ||

5192 (S2Op1 && (!areTypesHalfed(S2Op1, Op2) || !extractHalf(S2Op1, Op2))))

5193 return false;

5194

5195

5196

5197 int M1Start = 0;

5198 int M2Start = 0;

5199 int NumElements = cast(Op1->getType())->getNumElements() * 2;

5200 if ((S1Op1 &&

5202 (S2Op1 &&

5204 return false;

5205

5206 if ((M1Start != 0 && M1Start != (NumElements / 2)) ||

5207 (M2Start != 0 && M2Start != (NumElements / 2)))

5208 return false;

5209 if (S1Op1 && S2Op1 && M1Start != M2Start)

5210 return false;

5211

5212 return true;

5213}

5214

5215

5216

5218 auto areExtDoubled = [](Instruction *Ext) {

5219 return Ext->getType()->getScalarSizeInBits() ==

5220 2 * Ext->getOperand(0)->getType()->getScalarSizeInBits();

5221 };

5222

5225 !areExtDoubled(cast(Ext1)) ||

5226 !areExtDoubled(cast(Ext2)))

5227 return false;

5228

5229 return true;

5230}

5231

5232

5234 Value *VectorOperand = nullptr;

5238 ElementIndex->getValue() == 1 &&

5239 isa(VectorOperand->getType()) &&

5240 cast(VectorOperand->getType())->getNumElements() == 2;

5241}

5242

5243

5246}

5247

5249

5250 auto *GEP = dyn_cast(Ptrs);

5251 if (GEP || GEP->getNumOperands() != 2)

5252 return false;

5253

5255 Value *Offsets = GEP->getOperand(1);

5256

5257

5258 if (Base->getType()->isVectorTy() || !Offsets->getType()->isVectorTy())

5259 return false;

5260

5261

5262 if (isa(Offsets) || isa(Offsets)) {

5263 auto *OffsetsInst = cast(Offsets);

5264 if (OffsetsInst->getType()->getScalarSizeInBits() > 32 &&

5265 OffsetsInst->getOperand(0)->getType()->getScalarSizeInBits() <= 32)

5267 }

5268

5269

5270 return true;

5271}

5272

5273

5274

5275

5278 return true;

5281 Ops.push_back(&cast(Op)->getOperandUse(0));

5282 return true;

5283 }

5286 Value *ZExtOp = cast(Op)->getOperand(0);

5287 Ops.push_back(&cast(ZExtOp)->getOperandUse(0));

5288 Ops.push_back(&cast(Op)->getOperandUse(0));

5289 return true;

5290 }

5291 return false;

5292}

5293

5294

5295

5296

5300 switch (II->getIntrinsicID()) {

5301 case Intrinsic::aarch64_neon_smull:

5302 case Intrinsic::aarch64_neon_umull:

5304 true)) {

5307 return true;

5308 }

5309 [[fallthrough]];

5310

5311 case Intrinsic::fma:

5312 case Intrinsic::fmuladd:

5313 if (isa(I->getType()) &&

5314 cast(I->getType())->getElementType()->isHalfTy() &&

5315 !ST->hasFullFP16())

5316 return false;

5317 [[fallthrough]];

5318 case Intrinsic::aarch64_neon_sqdmull:

5319 case Intrinsic::aarch64_neon_sqdmulh:

5320 case Intrinsic::aarch64_neon_sqrdmulh:

5321

5326 return !Ops.empty();

5327 case Intrinsic::aarch64_neon_fmlal:

5328 case Intrinsic::aarch64_neon_fmlal2:

5329 case Intrinsic::aarch64_neon_fmlsl:

5330 case Intrinsic::aarch64_neon_fmlsl2:

5331

5336 return !Ops.empty();

5337 case Intrinsic::aarch64_sve_ptest_first:

5338 case Intrinsic::aarch64_sve_ptest_last:

5339 if (auto *IIOp = dyn_cast(II->getOperand(0)))

5340 if (IIOp->getIntrinsicID() == Intrinsic::aarch64_sve_ptrue)

5342 return !Ops.empty();

5343 case Intrinsic::aarch64_sme_write_horiz:

5344 case Intrinsic::aarch64_sme_write_vert:

5345 case Intrinsic::aarch64_sme_writeq_horiz:

5346 case Intrinsic::aarch64_sme_writeq_vert: {

5347 auto *Idx = dyn_cast(II->getOperand(1));

5348 if (Idx || Idx->getOpcode() != Instruction::Add)

5349 return false;

5351 return true;

5352 }

5353 case Intrinsic::aarch64_sme_read_horiz:

5354 case Intrinsic::aarch64_sme_read_vert:

5355 case Intrinsic::aarch64_sme_readq_horiz:

5356 case Intrinsic::aarch64_sme_readq_vert:

5357 case Intrinsic::aarch64_sme_ld1b_vert:

5358 case Intrinsic::aarch64_sme_ld1h_vert:

5359 case Intrinsic::aarch64_sme_ld1w_vert:

5360 case Intrinsic::aarch64_sme_ld1d_vert:

5361 case Intrinsic::aarch64_sme_ld1q_vert:

5362 case Intrinsic::aarch64_sme_st1b_vert:

5363 case Intrinsic::aarch64_sme_st1h_vert:

5364 case Intrinsic::aarch64_sme_st1w_vert:

5365 case Intrinsic::aarch64_sme_st1d_vert:

5366 case Intrinsic::aarch64_sme_st1q_vert:

5367 case Intrinsic::aarch64_sme_ld1b_horiz:

5368 case Intrinsic::aarch64_sme_ld1h_horiz:

5369 case Intrinsic::aarch64_sme_ld1w_horiz:

5370 case Intrinsic::aarch64_sme_ld1d_horiz:

5371 case Intrinsic::aarch64_sme_ld1q_horiz:

5372 case Intrinsic::aarch64_sme_st1b_horiz:

5373 case Intrinsic::aarch64_sme_st1h_horiz:

5374 case Intrinsic::aarch64_sme_st1w_horiz:

5375 case Intrinsic::aarch64_sme_st1d_horiz:

5376 case Intrinsic::aarch64_sme_st1q_horiz: {

5377 auto *Idx = dyn_cast(II->getOperand(3));

5378 if (Idx || Idx->getOpcode() != Instruction::Add)

5379 return false;

5381 return true;

5382 }

5383 case Intrinsic::aarch64_neon_pmull:

5385 return false;

5388 return true;

5389 case Intrinsic::aarch64_neon_pmull64:

5391 II->getArgOperand(1)))

5392 return false;

5393 Ops.push_back(&II->getArgOperandUse(0));

5394 Ops.push_back(&II->getArgOperandUse(1));

5395 return true;

5396 case Intrinsic::masked_gather:

5398 return false;

5399 Ops.push_back(&II->getArgOperandUse(0));

5400 return true;

5401 case Intrinsic::masked_scatter:

5403 return false;

5404 Ops.push_back(&II->getArgOperandUse(1));

5405 return true;

5406 default:

5407 return false;

5408 }

5409 }

5410

5411 auto ShouldSinkCondition = [](Value *Cond) -> bool {

5412 auto *II = dyn_cast(Cond);

5413 return II && II->getIntrinsicID() == Intrinsic::vector_reduce_or &&

5414 isa(II->getOperand(0)->getType());

5415 };

5416

5417 switch (I->getOpcode()) {

5418 case Instruction::GetElementPtr:

5419 case Instruction::Add:

5420 case Instruction::Sub:

5421

5422 for (unsigned Op = 0; Op < I->getNumOperands(); ++Op) {

5425 return true;

5426 }

5427 }

5428 break;

5429 case Instruction::Select: {

5430 if (!ShouldSinkCondition(I->getOperand(0)))

5431 return false;

5432

5433 Ops.push_back(&I->getOperandUse(0));

5434 return true;

5435 }

5436 case Instruction::Br: {

5437 if (cast(I)->isUnconditional())

5438 return false;

5439

5440 if (!ShouldSinkCondition(cast(I)->getCondition()))

5441 return false;

5442

5443 Ops.push_back(&I->getOperandUse(0));

5444 return true;

5445 }

5446 default:

5447 break;

5448 }

5449

5450 if (I->getType()->isVectorTy())

5451 return false;

5452

5453 switch (I->getOpcode()) {

5454 case Instruction::Sub:

5455 case Instruction::Add: {

5457 return false;

5458

5459

5460

5461 auto Ext1 = cast(I->getOperand(0));

5462 auto Ext2 = cast(I->getOperand(1));

5464 Ops.push_back(&Ext1->getOperandUse(0));

5465 Ops.push_back(&Ext2->getOperandUse(0));

5466 }

5467

5468 Ops.push_back(&I->getOperandUse(0));

5469 Ops.push_back(&I->getOperandUse(1));

5470

5471 return true;

5472 }

5473 case Instruction::Or: {

5474

5475

5476 if (ST->hasNEON()) {

5478 Value *MaskValue;

5479

5483 if (match(OtherAnd,

5485 Instruction *MainAnd = I->getOperand(0) == OtherAnd

5486 ? cast(I->getOperand(1))

5487 : cast(I->getOperand(0));

5488

5489

5490 if (I->getParent() != MainAnd->getParent() ||

5491 I->getParent() != OtherAnd->getParent())

5492 return false;

5493

5494

5495 if (I->getParent() != IA->getParent() ||

5496 I->getParent() != IB->getParent())

5497 return false;

5498

5501 Ops.push_back(&I->getOperandUse(0));

5502 Ops.push_back(&I->getOperandUse(1));

5503

5504 return true;

5505 }

5506 }

5507 }

5508

5509 return false;

5510 }

5511 case Instruction::Mul: {

5512 auto ShouldSinkSplatForIndexedVariant = [](Value *V) {

5513 auto *Ty = cast(V->getType());

5514

5515 if (Ty->isScalableTy())

5516 return false;

5517

5518

5519 return Ty->getScalarSizeInBits() == 16 || Ty->getScalarSizeInBits() == 32;

5520 };

5521

5522 int NumZExts = 0, NumSExts = 0;

5523 for (auto &Op : I->operands()) {

5524

5525 if (any_of(Ops, [&](Use *U) { return U->get() == Op; }))

5526 continue;

5527

5529 auto *Ext = cast(Op);

5530 auto *ExtOp = Ext->getOperand(0);

5531 if (isSplatShuffle(ExtOp) && ShouldSinkSplatForIndexedVariant(ExtOp))

5532 Ops.push_back(&Ext->getOperandUse(0));

5534

5535 if (isa(Ext))

5536 NumSExts++;

5537 else

5538 NumZExts++;

5539

5540 continue;

5541 }

5542

5544 if (!Shuffle)

5545 continue;

5546

5547

5548

5549

5555 NumSExts++;

5556 else

5557 NumZExts++;

5558 continue;

5559 }

5560

5562 InsertElementInst *Insert = dyn_cast(ShuffleOperand);

5563 if (!Insert)

5564 continue;

5565

5566 Instruction *OperandInstr = dyn_cast(Insert->getOperand(1));

5567 if (!OperandInstr)

5568 continue;

5569

5571 dyn_cast(Insert->getOperand(2));

5572

5573 if (!ElementConstant || !ElementConstant->isZero())

5574 continue;

5575

5576 unsigned Opcode = OperandInstr->getOpcode();

5577 if (Opcode == Instruction::SExt)

5578 NumSExts++;

5579 else if (Opcode == Instruction::ZExt)

5580 NumZExts++;

5581 else {

5582

5583

5584 unsigned Bitwidth = I->getType()->getScalarSizeInBits();

5588 continue;

5589 NumZExts++;

5590 }

5591

5592

5593

5595 Ops.push_back(&Insert->getOperandUse(1));

5598 }

5599

5600

5601 if (!Ops.empty() && (NumSExts == 2 || NumZExts == 2))

5602 return true;

5603

5604

5605 if (!ShouldSinkSplatForIndexedVariant(I))

5606 return false;

5607

5610 Ops.push_back(&I->getOperandUse(0));

5612 Ops.push_back(&I->getOperandUse(1));

5613

5614 return !Ops.empty();

5615 }

5616 case Instruction::FMul: {

5617

5618 if (I->getType()->isScalableTy())

5619 return false;

5620

5621 if (cast(I->getType())->getElementType()->isHalfTy() &&

5622 !ST->hasFullFP16())

5623 return false;

5624

5625

5627 Ops.push_back(&I->getOperandUse(0));

5629 Ops.push_back(&I->getOperandUse(1));

5630 return !Ops.empty();

5631 }

5632 default:

5633 return false;

5634 }

5635 return false;

5636}

static bool isAllActivePredicate(SelectionDAG &DAG, SDValue N)

SmallVector< AArch64_IMM::ImmInsnModel, 4 > Insn

static std::optional< Instruction * > instCombineSVEVectorMul(InstCombiner &IC, IntrinsicInst &II, Intrinsic::ID IID)

TailFoldingOption TailFoldingOptionLoc

static std::optional< Instruction * > instCombineSVEVectorFAdd(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineSVEVectorFuseMulAddSub(InstCombiner &IC, IntrinsicInst &II, bool MergeIntoAddendOp)

static void getFalkorUnrollingPreferences(Loop *L, ScalarEvolution &SE, TargetTransformInfo::UnrollingPreferences &UP)

bool SimplifyValuePattern(SmallVector< Value * > &Vec, bool AllowPoison)

static std::optional< Instruction * > instCombineSVESel(InstCombiner &IC, IntrinsicInst &II)

static bool shouldSinkVScale(Value *Op, SmallVectorImpl< Use * > &Ops)

We want to sink following cases: (add|sub|gep) A, ((mul|shl) vscale, imm); (add|sub|gep) A,...

static std::optional< Instruction * > tryCombineFromSVBoolBinOp(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineSVEUnpack(InstCombiner &IC, IntrinsicInst &II)

static cl::opt< unsigned > SVETailFoldInsnThreshold("sve-tail-folding-insn-threshold", cl::init(15), cl::Hidden)

static cl::opt< bool > EnableFixedwidthAutovecInStreamingMode("enable-fixedwidth-autovec-in-streaming-mode", cl::init(false), cl::Hidden)

static std::optional< Instruction * > instCombineSVEAllOrNoActive(InstCombiner &IC, IntrinsicInst &II, Intrinsic::ID IID)

static std::optional< Instruction * > instCombineSVEVectorFAddU(InstCombiner &IC, IntrinsicInst &II)

static bool areExtractExts(Value *Ext1, Value *Ext2)

Check if Ext1 and Ext2 are extends of the same type, doubling the bitwidth of the vector elements.

static cl::opt< bool > EnableLSRCostOpt("enable-aarch64-lsr-cost-opt", cl::init(true), cl::Hidden)

static bool shouldSinkVectorOfPtrs(Value *Ptrs, SmallVectorImpl< Use * > &Ops)

static std::optional< Instruction * > instCombineSVEVectorSub(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineLD1GatherIndex(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineSVEVectorFSub(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > processPhiNode(InstCombiner &IC, IntrinsicInst &II)

The function will remove redundant reinterprets casting in the presence of the control flow.

static std::optional< Instruction * > instCombineSVENoActiveUnaryErase(InstCombiner &IC, IntrinsicInst &II, int PredPos)

static std::optional< Instruction * > instCombineSVEInsr(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineST1ScatterIndex(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineSVESDIV(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineSVEST1(InstCombiner &IC, IntrinsicInst &II, const DataLayout &DL)

static bool containsDecreasingPointers(Loop *TheLoop, PredicatedScalarEvolution *PSE)

static std::optional< Instruction * > instCombineSVEAllActive(IntrinsicInst &II, Intrinsic::ID IID)

static std::optional< Instruction * > instCombineSVENoActiveZero(InstCombiner &IC, IntrinsicInst &II)

static cl::opt< bool > SVEPreferFixedOverScalableIfEqualCost("sve-prefer-fixed-over-scalable-if-equal", cl::Hidden)

static bool isUnpackedVectorVT(EVT VecVT)

static std::optional< Instruction * > instCombineSVEDupX(InstCombiner &IC, IntrinsicInst &II)

static void getAppleRuntimeUnrollPreferences(Loop *L, ScalarEvolution &SE, TargetTransformInfo::UnrollingPreferences &UP, AArch64TTIImpl &TTI)

For Apple CPUs, we want to runtime-unroll loops to make better use if the OOO engine's wide instructi...

static std::optional< Instruction * > instCombineSVECmpNE(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineDMB(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineSVEVectorFSubU(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineRDFFR(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineMaxMinNM(InstCombiner &IC, IntrinsicInst &II)

static InstructionCost getHistogramCost(const IntrinsicCostAttributes &ICA)

static cl::opt< unsigned > SVEGatherOverhead("sve-gather-overhead", cl::init(10), cl::Hidden)

static std::optional< Instruction * > instCombineSVECondLast(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineSVEPTest(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineSVEZip(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineSVEDup(InstCombiner &IC, IntrinsicInst &II)

static cl::opt< unsigned > BaseHistCntCost("aarch64-base-histcnt-cost", cl::init(8), cl::Hidden, cl::desc("The cost of a histcnt instruction"))

static std::optional< Instruction * > instCombineConvertFromSVBool(InstCombiner &IC, IntrinsicInst &II)

static cl::opt< unsigned > CallPenaltyChangeSM("call-penalty-sm-change", cl::init(5), cl::Hidden, cl::desc("Penalty of calling a function that requires a change to PSTATE.SM"))

static std::optional< Instruction * > instCombineSVEUzp1(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineSVEVectorBinOp(InstCombiner &IC, IntrinsicInst &II)

static cl::opt< bool > EnableScalableAutovecInStreamingMode("enable-scalable-autovec-in-streaming-mode", cl::init(false), cl::Hidden)

static std::optional< Instruction * > instCombineSVETBL(InstCombiner &IC, IntrinsicInst &II)

static bool areOperandsOfVmullHighP64(Value *Op1, Value *Op2)

Check if Op1 and Op2 could be used with vmull_high_p64 intrinsic.

static Instruction::BinaryOps intrinsicIDToBinOpCode(unsigned Intrinsic)

static bool isSplatShuffle(Value *V)

static cl::opt< unsigned > InlineCallPenaltyChangeSM("inline-call-penalty-sm-change", cl::init(10), cl::Hidden, cl::desc("Penalty of inlining a call that requires a change to PSTATE.SM"))

static std::optional< Instruction * > instCombineSVEAllOrNoActiveUnary(InstCombiner &IC, IntrinsicInst &II)

static std::optional< Instruction * > instCombineSVELD1(InstCombiner &IC, IntrinsicInst &II, const DataLayout &DL)

static std::optional< Instruction * > instCombineSVENoActiveReplace(InstCombiner &IC, IntrinsicInst &II, bool hasInactiveVector)

static std::optional< Instruction * > instCombineSVESrshl(InstCombiner &IC, IntrinsicInst &II)

static bool isSMEABIRoutineCall(const CallInst &CI)

static cl::opt< unsigned > DMBLookaheadThreshold("dmb-lookahead-threshold", cl::init(10), cl::Hidden, cl::desc("The number of instructions to search for a redundant dmb"))

static unsigned getSVEGatherScatterOverhead(unsigned Opcode, const AArch64Subtarget *ST)

static bool isOperandOfVmullHighP64(Value *Op)

Check if Op could be used with vmull_high_p64 intrinsic.

static std::optional< Instruction * > instCombineSVELast(InstCombiner &IC, IntrinsicInst &II)

static cl::opt< unsigned > NeonNonConstStrideOverhead("neon-nonconst-stride-overhead", cl::init(10), cl::Hidden)

static cl::opt< bool > EnableFalkorHWPFUnrollFix("enable-falkor-hwpf-unroll-fix", cl::init(true), cl::Hidden)

static std::optional< Instruction * > instCombineSVECntElts(InstCombiner &IC, IntrinsicInst &II, unsigned NumElts)

static bool hasPossibleIncompatibleOps(const Function *F)

Returns true if the function has explicit operations that can only be lowered using incompatible inst...

cl::opt< TailFoldingOption, true, cl::parser< std::string > > SVETailFolding("sve-tail-folding", cl::desc("Control the use of vectorisation using tail-folding for SVE where the" " option is specified in the form (Initial)[+(Flag1|Flag2|...)]:" "\ndisabled (Initial) No loop types will vectorize using " "tail-folding" "\ndefault (Initial) Uses the default tail-folding settings for " "the target CPU" "\nall (Initial) All legal loop types will vectorize using " "tail-folding" "\nsimple (Initial) Use tail-folding for simple loops (not " "reductions or recurrences)" "\nreductions Use tail-folding for loops containing reductions" "\nnoreductions Inverse of above" "\nrecurrences Use tail-folding for loops containing fixed order " "recurrences" "\nnorecurrences Inverse of above" "\nreverse Use tail-folding for loops requiring reversed " "predicates" "\nnoreverse Inverse of above"), cl::location(TailFoldingOptionLoc))

static bool areExtractShuffleVectors(Value *Op1, Value *Op2, bool AllowSplat=false)

Check if both Op1 and Op2 are shufflevector extracts of either the lower or upper half of the vector ...

static std::optional< Instruction * > instCombineSVEVectorAdd(InstCombiner &IC, IntrinsicInst &II)

static cl::opt< bool > EnableOrLikeSelectOpt("enable-aarch64-or-like-select", cl::init(true), cl::Hidden)

static cl::opt< unsigned > SVEScatterOverhead("sve-scatter-overhead", cl::init(10), cl::Hidden)

static std::optional< Instruction * > instCombineSVEDupqLane(InstCombiner &IC, IntrinsicInst &II)

This file a TargetTransformInfo::Concept conforming object specific to the AArch64 target machine.

AMDGPU Register Bank Select

MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL

This file provides a helper that implements much of the TTI interface in terms of the target-independ...

static Error reportError(StringRef Message)

static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")

static GCRegistry::Add< ErlangGC > A("erlang", "erlang-compatible garbage collector")

Analysis containing CSE Info

static cl::opt< TargetTransformInfo::TargetCostKind > CostKind("cost-kind", cl::desc("Target cost kind"), cl::init(TargetTransformInfo::TCK_RecipThroughput), cl::values(clEnumValN(TargetTransformInfo::TCK_RecipThroughput, "throughput", "Reciprocal throughput"), clEnumValN(TargetTransformInfo::TCK_Latency, "latency", "Instruction latency"), clEnumValN(TargetTransformInfo::TCK_CodeSize, "code-size", "Code size"), clEnumValN(TargetTransformInfo::TCK_SizeAndLatency, "size-latency", "Code size and latency")))

Cost tables and simple lookup functions.

Returns the sub type a function will return at a given Idx Should correspond to the result type of an ExtractValue instruction executed with just that one unsigned Idx

This file defines the DenseMap class.

This file provides the interface for the instcombine pass implementation.

This file defines the LoopVectorizationLegality class.

mir Rename Register Operands

static const Function * getCalledFunction(const Value *V)

uint64_t IntrinsicInst * II

const SmallVectorImpl< MachineOperand > & Cond

static uint64_t getBits(uint64_t Val, int Start, int End)

assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())

static unsigned getScalarSizeInBits(Type *Ty)

static SymbolRef::Type getType(const Symbol *Sym)

This file describes how to lower LLVM code to machine code.

This pass exposes codegen information to IR-level passes.

static unsigned getBitWidth(Type *Ty, const DataLayout &DL)

Returns the bitwidth of the given scalar or pointer type.

bool isNeonAvailable() const

Returns true if the target has NEON and the function at runtime is known to have NEON enabled (e....

unsigned getVectorInsertExtractBaseCost() const

ARMProcFamilyEnum getProcFamily() const

Returns ARM processor family.

unsigned getMaxInterleaveFactor() const

bool isSVEorStreamingSVEAvailable() const

Returns true if the target has access to either the full range of SVE instructions,...

TailFoldingOpts getSVETailFoldingDefaultOpts() const

bool useSVEForFixedLengthVectors() const

unsigned getEpilogueVectorizationMinVF() const

unsigned getMinSVEVectorSizeInBits() const

bool isSVEAvailable() const

Returns true if the target has SVE and can use the full range of SVE instructions,...

InstructionCost getSpliceCost(VectorType *Tp, int Index)

InstructionCost getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, TTI::CastContextHint CCH, TTI::TargetCostKind CostKind, const Instruction *I=nullptr)

InstructionCost getIntImmCostInst(unsigned Opcode, unsigned Idx, const APInt &Imm, Type *Ty, TTI::TargetCostKind CostKind, Instruction *Inst=nullptr)

InstructionCost getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, StackOffset BaseOffset, bool HasBaseReg, int64_t Scale, unsigned AddrSpace) const

Return the cost of the scaling factor used in the addressing mode represented by AM for this target,...

InstructionCost getMaskedMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind)

bool shouldTreatInstructionLikeSelect(const Instruction *I)

InstructionCost getAddressComputationCost(Type *Ty, ScalarEvolution *SE, const SCEV *Ptr)

bool prefersVectorizedAddressing() const

InstructionCost getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, TTI::TargetCostKind CostKind)

InstructionCost getArithmeticReductionCost(unsigned Opcode, VectorType *Ty, std::optional< FastMathFlags > FMF, TTI::TargetCostKind CostKind)

InstructionCost getPartialReductionCost(unsigned Opcode, Type *InputTypeA, Type *InputTypeB, Type *AccumType, ElementCount VF, TTI::PartialReductionExtendKind OpAExtend, TTI::PartialReductionExtendKind OpBExtend, std::optional< unsigned > BinOp) const

InstructionCost getExtractWithExtendCost(unsigned Opcode, Type *Dst, VectorType *VecTy, unsigned Index)

bool isProfitableToSinkOperands(Instruction *I, SmallVectorImpl< Use * > &Ops) const

Check if sinking I's operands to I's basic block is profitable, because the operands can be folded in...

unsigned getInlineCallPenalty(const Function *F, const CallBase &Call, unsigned DefaultCallPenalty) const

bool isLegalToVectorizeReduction(const RecurrenceDescriptor &RdxDesc, ElementCount VF) const

unsigned getEpilogueVectorizationMinVF() const

Value * getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst, Type *ExpectedType)

bool isLegalBroadcastLoad(Type *ElementTy, ElementCount NumElements) const

bool shouldConsiderAddressTypePromotion(const Instruction &I, bool &AllowPromotionWithoutCommonHeader)

See if I should be considered for address type promotion.

InstructionCost getArithmeticReductionCostSVE(unsigned Opcode, VectorType *ValTy, TTI::TargetCostKind CostKind)

InstructionCost getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef< unsigned > Indices, Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, bool UseMaskForCond=false, bool UseMaskForGaps=false)

InstructionCost getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred, TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info={TTI::OK_AnyValue, TTI::OP_None}, TTI::OperandValueInfo Op2Info={TTI::OK_AnyValue, TTI::OP_None}, const Instruction *I=nullptr)

uint64_t getFeatureMask(const Function &F) const

std::optional< Instruction * > instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const

bool shouldMaximizeVectorBandwidth(TargetTransformInfo::RegisterKind K) const

bool isLSRCostLess(const TargetTransformInfo::LSRCost &C1, const TargetTransformInfo::LSRCost &C2)

InstructionCost getMemoryOpCost(unsigned Opcode, Type *Src, MaybeAlign Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, TTI::OperandValueInfo OpInfo={TTI::OK_AnyValue, TTI::OP_None}, const Instruction *I=nullptr)

bool isElementTypeLegalForScalableVector(Type *Ty) const

InstructionCost getCostOfKeepingLiveOverCall(ArrayRef< Type * > Tys)

bool preferFixedOverScalableIfEqualCost() const

bool areInlineCompatible(const Function *Caller, const Function *Callee) const

bool enableScalableVectorization() const

bool useNeonVector(const Type *Ty) const

InstructionCost getVectorInstrCost(unsigned Opcode, Type *Val, TTI::TargetCostKind CostKind, unsigned Index, Value *Op0, Value *Op1)

TTI::PopcntSupportKind getPopcntSupport(unsigned TyWidth)

bool areTypesABICompatible(const Function *Caller, const Function *Callee, const ArrayRef< Type * > &Types) const

TypeSize getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const

InstructionCost getGatherScatterOpCost(unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask, Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I=nullptr)

bool isMultiversionedFunction(const Function &F) const

unsigned getMaxNumElements(ElementCount VF) const

Try to return an estimate cost factor that can be used as a multiplier when scalarizing an operation ...

InstructionCost getScalarizationOverhead(VectorType *Ty, const APInt &DemandedElts, bool Insert, bool Extract, TTI::TargetCostKind CostKind, ArrayRef< Value * > VL={})

InstructionCost getArithmeticInstrCost(unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info={TTI::OK_AnyValue, TTI::OP_None}, TTI::OperandValueInfo Op2Info={TTI::OK_AnyValue, TTI::OP_None}, ArrayRef< const Value * > Args={}, const Instruction *CxtI=nullptr)

bool preferPredicateOverEpilogue(TailFoldingInfo *TFI)

bool isLegalMaskedGatherScatter(Type *DataType) const

unsigned getMaxInterleaveFactor(ElementCount VF)

InstructionCost getShuffleCost(TTI::ShuffleKind Kind, VectorType *Tp, ArrayRef< int > Mask, TTI::TargetCostKind CostKind, int Index, VectorType *SubTp, ArrayRef< const Value * > Args={}, const Instruction *CxtI=nullptr)

void getPeelingPreferences(Loop *L, ScalarEvolution &SE, TTI::PeelingPreferences &PP)

InstructionCost getCFInstrCost(unsigned Opcode, TTI::TargetCostKind CostKind, const Instruction *I=nullptr)

InstructionCost getIntImmCost(int64_t Val)

Calculate the cost of materializing a 64-bit value.

std::optional< Value * > simplifyDemandedVectorEltsIntrinsic(InstCombiner &IC, IntrinsicInst &II, APInt DemandedElts, APInt &UndefElts, APInt &UndefElts2, APInt &UndefElts3, std::function< void(Instruction *, unsigned, APInt, APInt &)> SimplifyAndSetOp) const

void getUnrollingPreferences(Loop *L, ScalarEvolution &SE, TTI::UnrollingPreferences &UP, OptimizationRemarkEmitter *ORE)

bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info)

TTI::MemCmpExpansionOptions enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const

InstructionCost getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty, FastMathFlags FMF, TTI::TargetCostKind CostKind)

bool isExtPartOfAvgExpr(const Instruction *ExtUser, Type *Dst, Type *Src)

InstructionCost getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, const APInt &Imm, Type *Ty, TTI::TargetCostKind CostKind)

EVT getPromotedVTForPredicate(EVT VT) const

unsigned getNumInterleavedAccesses(VectorType *VecTy, const DataLayout &DL, bool UseScalable) const

Returns the number of interleaved accesses that will be generated when lowering accesses of the given...

bool isLegalInterleavedAccessType(VectorType *VecTy, const DataLayout &DL, bool &UseScalable) const

Returns true if VecTy is a legal interleaved access type.

Class for arbitrary precision integers.

bool isNegatedPowerOf2() const

Check if this APInt's negated value is a power of two greater than zero.

unsigned popcount() const

Count the number of bits set.

unsigned countLeadingOnes() const

void negate()

Negate this APInt in place.

APInt sextOrTrunc(unsigned width) const

Sign extend or truncate to width.

unsigned logBase2() const

APInt ashr(unsigned ShiftAmt) const

Arithmetic right-shift function.

bool isPowerOf2() const

Check if this APInt's value is a power of two greater than zero.

static APInt getHighBitsSet(unsigned numBits, unsigned hiBitsSet)

Constructs an APInt value that has the top hiBitsSet bits set.

int64_t getSExtValue() const

Get sign extended value.

ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...

LLVM Basic Block Representation.

bool isTypeLegal(Type *Ty)

InstructionCost getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA, TTI::TargetCostKind CostKind)

Get intrinsic cost based on arguments.

InstructionCost getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef< unsigned > Indices, Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, bool UseMaskForCond=false, bool UseMaskForGaps=false)

void getUnrollingPreferences(Loop *L, ScalarEvolution &SE, TTI::UnrollingPreferences &UP, OptimizationRemarkEmitter *ORE)

InstructionCost getMaskedMemoryOpCost(unsigned Opcode, Type *DataTy, Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind)

TTI::ShuffleKind improveShuffleKindFromMask(TTI::ShuffleKind Kind, ArrayRef< int > Mask, VectorType *Ty, int &Index, VectorType *&SubTy) const

InstructionCost getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty, FastMathFlags FMF, TTI::TargetCostKind CostKind)

Try to calculate op costs for min/max reduction operations.

InstructionCost getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred, TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info={TTI::OK_AnyValue, TTI::OP_None}, TTI::OperandValueInfo Op2Info={TTI::OK_AnyValue, TTI::OP_None}, const Instruction *I=nullptr)

InstructionCost getMemoryOpCost(unsigned Opcode, Type *Src, MaybeAlign Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind, TTI::OperandValueInfo OpInfo={TTI::OK_AnyValue, TTI::OP_None}, const Instruction *I=nullptr)

InstructionCost getGatherScatterOpCost(unsigned Opcode, Type *DataTy, const Value *Ptr, bool VariableMask, Align Alignment, TTI::TargetCostKind CostKind, const Instruction *I=nullptr)

InstructionCost getShuffleCost(TTI::ShuffleKind Kind, VectorType *Tp, ArrayRef< int > Mask, TTI::TargetCostKind CostKind, int Index, VectorType *SubTp, ArrayRef< const Value * > Args={}, const Instruction *CxtI=nullptr)

bool areInlineCompatible(const Function *Caller, const Function *Callee) const

void getPeelingPreferences(Loop *L, ScalarEvolution &SE, TTI::PeelingPreferences &PP)

InstructionCost getCallInstrCost(Function *F, Type *RetTy, ArrayRef< Type * > Tys, TTI::TargetCostKind CostKind)

Compute a cost of the given call instruction.

InstructionCost getArithmeticReductionCost(unsigned Opcode, VectorType *Ty, std::optional< FastMathFlags > FMF, TTI::TargetCostKind CostKind)

std::pair< InstructionCost, MVT > getTypeLegalizationCost(Type *Ty) const

Estimate the cost of type-legalization and the legalized type.

InstructionCost getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, TTI::CastContextHint CCH, TTI::TargetCostKind CostKind, const Instruction *I=nullptr)

InstructionCost getScalarizationOverhead(VectorType *InTy, const APInt &DemandedElts, bool Insert, bool Extract, TTI::TargetCostKind CostKind, ArrayRef< Value * > VL={})

Estimate the overhead of scalarizing an instruction.

InstructionCost getArithmeticInstrCost(unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, TTI::OperandValueInfo Opd1Info={TTI::OK_AnyValue, TTI::OP_None}, TTI::OperandValueInfo Opd2Info={TTI::OK_AnyValue, TTI::OP_None}, ArrayRef< const Value * > Args={}, const Instruction *CxtI=nullptr)

bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg, int64_t Scale, unsigned AddrSpace, Instruction *I=nullptr, int64_t ScalableOffset=0)

static BinaryOperator * CreateWithCopiedFlags(BinaryOps Opc, Value *V1, Value *V2, Value *CopyO, const Twine &Name="", InsertPosition InsertBefore=nullptr)

Base class for all callable instructions (InvokeInst and CallInst) Holds everything related to callin...

Function * getCalledFunction() const

Returns the function called, or null if this is an indirect function invocation or the function signa...

Value * getArgOperand(unsigned i) const

unsigned arg_size() const

This class represents a function call, abstracting a target machine's calling convention.

Predicate

This enumeration lists the possible predicates for CmpInst subclasses.

@ FCMP_OEQ

0 0 0 1 True if ordered and equal

@ FCMP_OLT

0 1 0 0 True if ordered and less than

@ FCMP_OGT

0 0 1 0 True if ordered and greater than

@ FCMP_OGE

0 0 1 1 True if ordered and greater than or equal

@ FCMP_OLE

0 1 0 1 True if ordered and less than or equal

@ FCMP_UNE

1 1 1 0 True if unordered or not equal

@ FCMP_UNO

1 0 0 0 True if unordered: isnan(X) | isnan(Y)

bool isIntPredicate() const

An abstraction over a floating-point predicate, and a pack of an integer predicate with samesign info...

static ConstantAggregateZero * get(Type *Ty)

This is the shared class of boolean and integer constants.

bool isZero() const

This is just a convenience method to make client code smaller for a common code.

const APInt & getValue() const

Return the constant as an APInt value reference.

static Constant * get(StructType *T, ArrayRef< Constant * > V)

This is an important base class in LLVM.

static Constant * getNullValue(Type *Ty)

Constructor to create a '0' constant of arbitrary type.

This class represents an Operation in the Expression.

A parsed version of the target data layout string in and methods for querying it.

iterator find(const_arg_type_t< KeyT > Val)

static constexpr ElementCount getScalable(ScalarTy MinVal)

static constexpr ElementCount getFixed(ScalarTy MinVal)

This provides a helper for copying FMF from an instruction or setting specified flags.

Convenience struct for specifying and reasoning about fast-math flags.

bool allowContract() const

static FixedVectorType * get(Type *ElementType, unsigned NumElts)

an instruction for type-safe pointer arithmetic to access elements of arrays and structs

bool isEquality() const

Return true if this predicate is either EQ or NE.

Value * CreateVScale(Constant *Scaling, const Twine &Name="")

Create a call to llvm.vscale, multiplied by Scaling.

Value * CreateInsertElement(Type *VecTy, Value *NewElt, Value *Idx, const Twine &Name="")

Value * CreateInsertValue(Value *Agg, Value *Val, ArrayRef< unsigned > Idxs, const Twine &Name="")

CallInst * CreateInsertVector(Type *DstType, Value *SrcVec, Value *SubVec, Value *Idx, const Twine &Name="")

Create a call to the vector.insert intrinsic.

Value * CreateExtractElement(Value *Vec, Value *Idx, const Twine &Name="")

IntegerType * getIntNTy(unsigned N)

Fetch the type representing an N-bit integer.

Type * getDoubleTy()

Fetch the type representing a 64-bit floating point value.

Value * CreateVectorSplat(unsigned NumElts, Value *V, const Twine &Name="")

Return a vector value that contains.

CallInst * CreateMaskedLoad(Type *Ty, Value *Ptr, Align Alignment, Value *Mask, Value *PassThru=nullptr, const Twine &Name="")

Create a call to Masked Load intrinsic.

Value * CreateSelect(Value *C, Value *True, Value *False, const Twine &Name="", Instruction *MDFrom=nullptr)

IntegerType * getInt32Ty()

Fetch the type representing a 32-bit integer.

Type * getHalfTy()

Fetch the type representing a 16-bit floating point value.

IntegerType * getInt64Ty()

Fetch the type representing a 64-bit integer.

Value * CreateGEP(Type *Ty, Value *Ptr, ArrayRef< Value * > IdxList, const Twine &Name="", GEPNoWrapFlags NW=GEPNoWrapFlags::none())

ConstantInt * getInt64(uint64_t C)

Get a constant 64-bit value.

CallInst * CreateIntrinsic(Intrinsic::ID ID, ArrayRef< Type * > Types, ArrayRef< Value * > Args, FMFSource FMFSource={}, const Twine &Name="")

Create a call to intrinsic ID with Args, mangled using Types.

Value * CreateBitOrPointerCast(Value *V, Type *DestTy, const Twine &Name="")

PHINode * CreatePHI(Type *Ty, unsigned NumReservedValues, const Twine &Name="")

Value * CreateBinOpFMF(Instruction::BinaryOps Opc, Value *LHS, Value *RHS, FMFSource FMFSource, const Twine &Name="", MDNode *FPMathTag=nullptr)

Value * CreateBitCast(Value *V, Type *DestTy, const Twine &Name="")

LoadInst * CreateLoad(Type *Ty, Value *Ptr, const char *Name)

Provided to resolve 'CreateLoad(Ty, Ptr, "...")' correctly, instead of converting the string to 'bool...

Value * CreateShuffleVector(Value *V1, Value *V2, Value *Mask, const Twine &Name="")

StoreInst * CreateStore(Value *Val, Value *Ptr, bool isVolatile=false)

CallInst * CreateMaskedStore(Value *Val, Value *Ptr, Align Alignment, Value *Mask)

Create a call to Masked Store intrinsic.

Type * getFloatTy()

Fetch the type representing a 32-bit floating point value.

Value * CreateIntCast(Value *V, Type *DestTy, bool isSigned, const Twine &Name="")

void SetInsertPoint(BasicBlock *TheBB)

This specifies that created instructions should be appended to the end of the specified block.

This provides a uniform API for creating instructions and inserting them into a basic block: either a...

This instruction inserts a single (scalar) element into a VectorType value.

static InsertElementInst * Create(Value *Vec, Value *NewElt, Value *Idx, const Twine &NameStr="", InsertPosition InsertBefore=nullptr)

The core instruction combiner logic.

virtual Instruction * eraseInstFromFunction(Instruction &I)=0

Combiner aware instruction erasure.

Instruction * replaceInstUsesWith(Instruction &I, Value *V)

A combiner-aware RAUW-like routine.

Instruction * replaceOperand(Instruction &I, unsigned OpNum, Value *V)

Replace operand of instruction and add old operand to the worklist.

static InstructionCost getInvalid(CostType Val=0)

std::optional< CostType > getValue() const

This function is intended to be used as sparingly as possible, since the class provides the full rang...

unsigned getOpcode() const

Returns a member of one of the enums like Instruction::Add.

void copyMetadata(const Instruction &SrcInst, ArrayRef< unsigned > WL=ArrayRef< unsigned >())

Copy metadata from SrcInst to this instruction.

Class to represent integer types.

bool hasGroups() const

Returns true if we have any interleave groups.

const SmallVectorImpl< Type * > & getArgTypes() const

Type * getReturnType() const

const SmallVectorImpl< const Value * > & getArgs() const

Intrinsic::ID getID() const

A wrapper class for inspecting calls to intrinsic functions.

Intrinsic::ID getIntrinsicID() const

Return the intrinsic ID of this intrinsic.

This is an important class for using LLVM in a threaded context.

An instruction for reading from memory.

Value * getPointerOperand()

iterator_range< block_iterator > blocks() const

RecurrenceSet & getFixedOrderRecurrences()

Return the fixed-order recurrences found in the loop.

PredicatedScalarEvolution * getPredicatedScalarEvolution() const

const ReductionList & getReductionVars() const

Returns the reduction variables found in the loop.

Represents a single loop in the control flow graph.

uint64_t getScalarSizeInBits() const

unsigned getVectorNumElements() const

bool isVector() const

Return true if this is a vector value type.

void addIncoming(Value *V, BasicBlock *BB)

Add an incoming value to the end of the PHI list.

static PoisonValue * get(Type *T)

Static factory methods - Return an 'poison' object of the specified type.

An interface layer with SCEV used to manage how we see SCEV expressions for values in the context of ...

The RecurrenceDescriptor is used to identify recurrences variables in a loop.

Type * getRecurrenceType() const

Returns the type of the recurrence.

RecurKind getRecurrenceKind() const

This node represents a polynomial recurrence on the trip count of the specified loop.

bool isAffine() const

Return true if this represents an expression A + B*x where A and B are loop invariant values.

This class represents an analyzed expression in the program.

SMEAttrs is a utility class to parse the SME ACLE attributes on functions.

bool requiresSMChange(const SMEAttrs &Callee) const

void set(unsigned M, bool Enable=true)

bool hasStreamingBody() const

static ScalableVectorType * get(Type *ElementType, unsigned MinNumElts)

static ScalableVectorType * getDoubleElementsVectorType(ScalableVectorType *VTy)

The main scalar evolution driver.

const SCEV * getBackedgeTakenCount(const Loop *L, ExitCountKind Kind=Exact)

If the specified loop has a predictable backedge-taken count, return it, otherwise return a SCEVCould...

const SCEV * getSCEV(Value *V)

Return a SCEV expression for the full generality of the specified expression.

unsigned getSmallConstantMaxTripCount(const Loop *L, SmallVectorImpl< const SCEVPredicate * > *Predicates=nullptr)

Returns the upper bound of the loop trip count as a normal unsigned value.

bool isLoopInvariant(const SCEV *S, const Loop *L)

Return true if the value of the given SCEV is unchanging in the specified loop.

This instruction constructs a fixed permutation of two input vectors.

static bool isDeInterleaveMaskOfFactor(ArrayRef< int > Mask, unsigned Factor, unsigned &Index)

Check if the mask is a DE-interleave mask of the given factor Factor like: <Index,...

static bool isExtractSubvectorMask(ArrayRef< int > Mask, int NumSrcElts, int &Index)

Return true if this shuffle mask is an extract subvector mask.

static bool isInterleaveMask(ArrayRef< int > Mask, unsigned Factor, unsigned NumInputElts, SmallVectorImpl< unsigned > &StartIndexes)

Return true if the mask interleaves one or more input vectors together.

std::pair< iterator, bool > insert(PtrType Ptr)

Inserts Ptr if and only if there is no element in the container equal to Ptr.

bool contains(ConstPtrType Ptr) const

SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements.

This class consists of common code factored out of the SmallVector class to reduce code duplication b...

iterator insert(iterator I, T &&Elt)

void push_back(const T &Elt)

This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.

StackOffset holds a fixed and a scalable offset in bytes.

static StackOffset getScalable(int64_t Scalable)

static StackOffset getFixed(int64_t Fixed)

An instruction for storing to memory.

StringRef - Represent a constant reference to a string, i.e.

std::pair< StringRef, StringRef > split(char Separator) const

Split into two substrings around the first occurrence of a separator character.

A switch()-like statement whose cases are string literals.

StringSwitch & Case(StringLiteral S, T Value)

Class to represent struct types.

int InstructionOpcodeToISD(unsigned Opcode) const

Get the ISD node that corresponds to the Instruction class opcode.

EVT getValueType(const DataLayout &DL, Type *Ty, bool AllowUnknown=false) const

Return the EVT corresponding to this LLVM type.

unsigned getMaxExpandSizeMemcmp(bool OptSize) const

Get maximum # of load operations permitted for memcmp.

bool isTypeLegal(EVT VT) const

Return true if the target has native support for the specified value type.

bool isOperationLegalOrCustom(unsigned Op, EVT VT, bool LegalOnly=false) const

Return true if the specified operation is legal on this target or can be made legal with custom lower...

LegalizeKind getTypeConversion(LLVMContext &Context, EVT VT) const

Return pair that represents the legalization kind (first) that needs to happen to EVT (second) in ord...

LegalizeTypeAction getTypeAction(LLVMContext &Context, EVT VT) const

Return how we should legalize values of this type, either it is already legal (return 'Legal') or we ...

std::pair< LegalizeTypeAction, EVT > LegalizeKind

LegalizeKind holds the legalization kind that needs to happen to EVT in order to type-legalize it.

bool shouldTreatInstructionLikeSelect(const Instruction *I)

bool areTypesABICompatible(const Function *Caller, const Function *Callee, const ArrayRef< Type * > &Types) const

bool isLSRCostLess(const TTI::LSRCost &C1, const TTI::LSRCost &C2) const

bool isConstantStridedAccessLessThan(ScalarEvolution *SE, const SCEV *Ptr, int64_t MergeDistance) const

bool isLoweredToCall(const Function *F) const

This pass provides access to the codegen interfaces that are needed for IR-level transformations.

static OperandValueInfo getOperandInfo(const Value *V)

Collect properties of V used in cost analysis, e.g. OP_PowerOf2.

TargetCostKind

The kind of cost model.

@ TCK_RecipThroughput

Reciprocal throughput.

@ TCK_CodeSize

Instruction code size.

@ TCK_SizeAndLatency

The weighted sum of size and latency.

static bool requiresOrderedReduction(std::optional< FastMathFlags > FMF)

A helper function to determine the type of reduction algorithm used for a given Opcode and set of Fas...

PopcntSupportKind

Flags indicating the kind of support for population count.

PartialReductionExtendKind

@ TCC_Free

Expected to fold away in lowering.

@ TCC_Basic

The cost of a typical 'add' instruction.

InstructionCost getInstructionCost(const User *U, ArrayRef< const Value * > Operands, TargetCostKind CostKind) const

Estimate the cost of a given IR user when lowered.

ShuffleKind

The various kinds of shuffle patterns for vector queries.

@ SK_InsertSubvector

InsertSubvector. Index indicates start offset.

@ SK_Select

Selects elements from the corresponding lane of either source operand.

@ SK_PermuteSingleSrc

Shuffle elements of single source vector with any shuffle mask.

@ SK_Transpose

Transpose two vectors.

@ SK_Splice

Concatenates elements from the first input vector with elements of the second input vector.

@ SK_Broadcast

Broadcast element 0 to all other elements.

@ SK_PermuteTwoSrc

Merge elements from two source vectors into one with any shuffle mask.

@ SK_Reverse

Reverse the order of the vector.

@ SK_ExtractSubvector

ExtractSubvector Index indicates start offset.

CastContextHint

Represents a hint about the context in which a cast is used.

@ Masked

The cast is used with a masked load/store.

@ None

The cast is not used with a load/store of any kind.

@ Normal

The cast is used with a normal load/store.

static constexpr TypeSize getFixed(ScalarTy ExactSize)

static constexpr TypeSize getScalable(ScalarTy MinimumSize)

The instances of the Type class are immutable: once they are created, they are never changed.

bool isVectorTy() const

True if this is an instance of VectorType.

bool isPointerTy() const

True if this is an instance of PointerType.

static IntegerType * getInt1Ty(LLVMContext &C)

bool isFloatTy() const

Return true if this is 'float', a 32-bit IEEE fp type.

bool isBFloatTy() const

Return true if this is 'bfloat', a 16-bit bfloat type.

static IntegerType * getIntNTy(LLVMContext &C, unsigned N)

bool isFP128Ty() const

Return true if this is 'fp128'.

unsigned getScalarSizeInBits() const LLVM_READONLY

If this is a vector type, return the getPrimitiveSizeInBits value for the element type.

bool isHalfTy() const

Return true if this is 'half', a 16-bit IEEE fp type.

bool isScalableTy(SmallPtrSetImpl< const Type * > &Visited) const

Return true if this is a type whose size is a known multiple of vscale.

LLVMContext & getContext() const

Return the LLVMContext in which this type was uniqued.

bool isDoubleTy() const

Return true if this is 'double', a 64-bit IEEE fp type.

bool isFloatingPointTy() const

Return true if this is one of the floating-point types.

bool isPtrOrPtrVectorTy() const

Return true if this is a pointer type or a vector of pointer types.

static IntegerType * getInt32Ty(LLVMContext &C)

static IntegerType * getInt64Ty(LLVMContext &C)

static Type * getFloatTy(LLVMContext &C)

bool isIntegerTy() const

True if this is an instance of IntegerType.

TypeSize getPrimitiveSizeInBits() const LLVM_READONLY

Return the basic size of this type if it is a primitive type.

Type * getScalarType() const

If this is a vector type, return the element type, otherwise return 'this'.

static UndefValue * get(Type *T)

Static factory methods - Return an 'undef' object of the specified type.

A Use represents the edge between a Value definition and its users.

const Use & getOperandUse(unsigned i) const

Value * getOperand(unsigned i) const

LLVM Value Representation.

Type * getType() const

All values are typed, get the type of this value.

user_iterator user_begin()

bool hasOneUse() const

Return true if there is exactly one use of this value.

Align getPointerAlignment(const DataLayout &DL) const

Returns an alignment of the pointer value.

void takeName(Value *V)

Transfer the name from V to this value.

Base class of all SIMD vector types.

ElementCount getElementCount() const

Return an ElementCount instance to represent the (possibly scalable) number of elements in the vector...

static VectorType * get(Type *ElementType, ElementCount EC)

This static method is the primary way to construct an VectorType.

Type * getElementType() const

int getNumOccurrences() const

constexpr ScalarTy getFixedValue() const

constexpr bool isScalable() const

Returns whether the quantity is scaled by a runtime quantity (vscale).

constexpr bool isFixed() const

Returns true if the quantity is not scaled by vscale.

constexpr ScalarTy getKnownMinValue() const

Returns the minimum value this quantity can represent.

const ParentTy * getParent() const

#define llvm_unreachable(msg)

Marks that the current location is not supposed to be reachable.

static bool isLogicalImmediate(uint64_t imm, unsigned regSize)

isLogicalImmediate - Return true if the immediate is valid for a logical immediate instruction of the...

void expandMOVImm(uint64_t Imm, unsigned BitSize, SmallVectorImpl< ImmInsnModel > &Insn)

Expand a MOVi32imm or MOVi64imm pseudo instruction to one or more real move-immediate instructions to...

uint64_t getFMVPriority(ArrayRef< StringRef > Features)

static constexpr unsigned SVEBitsPerBlock

@ C

The default llvm calling convention, compatible with C.

@ SETCC

SetCC operator - This evaluates to a true value iff the condition is true.

@ ADD

Simple integer binary arithmetic operators.

@ SINT_TO_FP

[SU]INT_TO_FP - These operators convert integers (whose interpreted sign depends on the first letter)...

@ FADD

Simple binary floating point operators.

@ BITCAST

BITCAST - This operator converts between integer, vector and FP values, as if the value was stored to...

@ SIGN_EXTEND

Conversion operators.

@ FNEG

Perform various unary floating-point operations inspired by libm.

@ SELECT

Select(COND, TRUEVAL, FALSEVAL).

@ MULHU

MULHU/MULHS - Multiply high - Multiply two integers of type iN, producing an unsigned/signed value of...

@ SHL

Shift and rotation operations.

@ ZERO_EXTEND

ZERO_EXTEND - Used for integer types, zeroing the new bits.

@ FP_EXTEND

X = FP_EXTEND(Y) - Extend a smaller FP type into a larger FP type.

@ FP_TO_SINT

FP_TO_[US]INT - Convert a floating point value to a signed or unsigned integer.

@ AND

Bitwise operators - logical and, logical or, logical xor.

@ FP_ROUND

X = FP_ROUND(Y, TRUNC) - Rounding 'Y' from a larger floating point type down to the precision of the ...

@ TRUNCATE

TRUNCATE - Completely drop the high bits.

Function * getOrInsertDeclaration(Module *M, ID id, ArrayRef< Type * > Tys={})

Look up the Function declaration of the intrinsic id in the Module M.

BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)

class_match< BinaryOperator > m_BinOp()

Match an arbitrary binary operation and ignore it.

BinaryOp_match< LHS, RHS, Instruction::And, true > m_c_And(const LHS &L, const RHS &R)

Matches an And with LHS and RHS in either order.

specific_intval< false > m_SpecificInt(const APInt &V)

Match a specific integer value or vector with all elements equal to the value.

BinaryOp_match< LHS, RHS, Instruction::FMul > m_FMul(const LHS &L, const RHS &R)

bool match(Val *V, const Pattern &P)

bind_ty< Instruction > m_Instruction(Instruction *&I)

Match an instruction, capturing it if we match.

specificval_ty m_Specific(const Value *V)

Match if we have a specific specified value.

TwoOps_match< Val_t, Idx_t, Instruction::ExtractElement > m_ExtractElt(const Val_t &Val, const Idx_t &Idx)

Matches ExtractElementInst.

cst_pred_ty< is_nonnegative > m_NonNegative()

Match an integer or vector of non-negative values.

class_match< ConstantInt > m_ConstantInt()

Match an arbitrary ConstantInt and ignore it.

cst_pred_ty< is_one > m_One()

Match an integer 1 or a vector with all elements equal to 1.

ThreeOps_match< Cond, LHS, RHS, Instruction::Select > m_Select(const Cond &C, const LHS &L, const RHS &R)

Matches SelectInst.

BinaryOp_match< LHS, RHS, Instruction::Mul > m_Mul(const LHS &L, const RHS &R)

cst_pred_ty< is_zero_int > m_ZeroInt()

Match an integer 0 or a vector with all elements equal to 0.

OneUse_match< T > m_OneUse(const T &SubPattern)

TwoOps_match< V1_t, V2_t, Instruction::ShuffleVector > m_Shuffle(const V1_t &v1, const V2_t &v2)

Matches ShuffleVectorInst independently of mask value.

OneOps_match< OpTy, Instruction::Load > m_Load(const OpTy &Op)

Matches LoadInst.

CastInst_match< OpTy, ZExtInst > m_ZExt(const OpTy &Op)

Matches ZExt.

class_match< CmpInst > m_Cmp()

Matches any compare instruction and ignore it.

brc_match< Cond_t, bind_ty< BasicBlock >, bind_ty< BasicBlock > > m_Br(const Cond_t &C, BasicBlock *&T, BasicBlock *&F)

specific_fpval m_FPOne()

Match a float 1.0 or vector with all elements equal to 1.0.

BinaryOp_match< LHS, RHS, Instruction::Add, true > m_c_Add(const LHS &L, const RHS &R)

Matches a Add with LHS and RHS in either order.

VScaleVal_match m_VScale()

class_match< Value > m_Value()

Match an arbitrary value and ignore it.

CmpClass_match< LHS, RHS, ICmpInst > m_ICmp(CmpPredicate &Pred, const LHS &L, const RHS &R)

match_combine_or< CastInst_match< OpTy, ZExtInst >, CastInst_match< OpTy, SExtInst > > m_ZExtOrSExt(const OpTy &Op)

BinaryOp_match< LHS, RHS, Instruction::Shl > m_Shl(const LHS &L, const RHS &R)

auto m_Undef()

Match an arbitrary undef constant.

BinaryOp_match< cst_pred_ty< is_all_ones >, ValTy, Instruction::Xor, true > m_Not(const ValTy &V)

Matches a 'Not' as 'xor V, -1' or 'xor -1, V'.

CastInst_match< OpTy, SExtInst > m_SExt(const OpTy &Op)

Matches SExt.

is_zero m_Zero()

Match any null constant or a vector with all elements equal to 0.

BinaryOp_match< LHS, RHS, Instruction::Or, true > m_c_Or(const LHS &L, const RHS &R)

Matches an Or with LHS and RHS in either order.

initializer< Ty > init(const Ty &Val)

LocationClass< Ty > location(Ty &L)

This is an optimization pass for GlobalISel generic memory operations.

auto drop_begin(T &&RangeOrContainer, size_t N=1)

Return a range covering RangeOrContainer with the first N elements excluded.

bool all_of(R &&range, UnaryPredicate P)

Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly.

const CostTblEntryT< CostType > * CostTableLookup(ArrayRef< CostTblEntryT< CostType > > Tbl, int ISD, MVT Ty)

Find in cost table.

TailFoldingOpts

An enum to describe what types of loops we should attempt to tail-fold: Disabled: None Reductions: Lo...

auto enumerate(FirstRange &&First, RestRanges &&...Rest)

Given two or more input ranges, returns a new range whose values are tuples (A, B,...

std::optional< const MDOperand * > findStringMetadataForLoop(const Loop *TheLoop, StringRef Name)

Find string metadata for loop.

const Value * getLoadStorePointerOperand(const Value *V)

A helper function that returns the pointer operand of a load or store instruction.

constexpr bool isPowerOf2_64(uint64_t Value)

Return true if the argument is a power of two > 0 (64 bit edition.)

Value * getSplatValue(const Value *V)

Get splat value if the input is a splat vector or return nullptr.

bool MaskedValueIsZero(const Value *V, const APInt &Mask, const SimplifyQuery &SQ, unsigned Depth=0)

Return true if 'V & Mask' is known to be zero.

unsigned M1(unsigned Val)

bool any_of(R &&range, UnaryPredicate P)

Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly.

bool isSplatValue(const Value *V, int Index=-1, unsigned Depth=0)

Return true if each element of the vector value V is poisoned or equal to every other non-poisoned el...

unsigned getPerfectShuffleCost(llvm::ArrayRef< int > M)

unsigned Log2_32(uint32_t Value)

Return the floor log base 2 of the specified value, -1 if the value is zero.

constexpr bool isPowerOf2_32(uint32_t Value)

Return true if the argument is a power of two > 0.

raw_ostream & dbgs()

dbgs() - This returns a reference to a raw_ostream for debugging messages.

bool none_of(R &&Range, UnaryPredicate P)

Provide wrappers to std::none_of which take ranges instead of having to pass begin/end explicitly.

void report_fatal_error(Error Err, bool gen_crash_diag=true)

Report a serious error, calling any installed error handler.

std::optional< int64_t > getPtrStride(PredicatedScalarEvolution &PSE, Type *AccessTy, Value *Ptr, const Loop *Lp, const DenseMap< Value *, const SCEV * > &StridesMap=DenseMap< Value *, const SCEV * >(), bool Assume=false, bool ShouldCheckWrap=true)

If the pointer has a constant stride return it in units of the access type size.

bool isUZPMask(ArrayRef< int > M, unsigned NumElts, unsigned &WhichResultOut)

Return true for uzp1 or uzp2 masks of the form: <0, 2, 4, 6, 8, 10, 12, 14> or <1,...

constexpr int PoisonMaskElem

raw_fd_ostream & errs()

This returns a reference to a raw_ostream for standard error.

@ Mod

The access may modify the value stored in memory.

bool isZIPMask(ArrayRef< int > M, unsigned NumElts, unsigned &WhichResultOut)

Return true for zip1 or zip2 masks of the form: <0, 8, 1, 9, 2, 10, 3, 11> or <4, 12,...

@ UMin

Unsigned integer min implemented in terms of select(cmp()).

@ FAnyOf

Any_of reduction with select(fcmp(),x,y) where one of (x,y) is loop invariant, and both x and y are i...

@ Or

Bitwise or logical OR of integers.

@ Mul

Product of integers.

@ Xor

Bitwise or logical XOR of integers.

@ FMax

FP max implemented in terms of select(cmp()).

@ FMulAdd

Sum of float products with llvm.fmuladd(a * b + sum).

@ SMax

Signed integer max implemented in terms of select(cmp()).

@ And

Bitwise or logical AND of integers.

@ SMin

Signed integer min implemented in terms of select(cmp()).

@ FMin

FP min implemented in terms of select(cmp()).

@ IAnyOf

Any_of reduction with select(icmp(),x,y) where one of (x,y) is loop invariant, and both x and y are i...

@ UMax

Unsigned integer max implemented in terms of select(cmp()).

void computeKnownBits(const Value *V, KnownBits &Known, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, bool UseInstrInfo=true)

Determine which bits of V are known to be either zero or one and return them in the KnownZero/KnownOn...

DWARFExpression::Operation Op

unsigned getNumElementsFromSVEPredPattern(unsigned Pattern)

Return the number of active elements for VL1 to VL256 predicate pattern, zero for all other patterns.

auto predecessors(const MachineBasicBlock *BB)

Type * getLoadStoreType(const Value *I)

A helper function that returns the type of a load or store instruction.

bool all_equal(std::initializer_list< T > Values)

Returns true if all Values in the initializer lists are equal or the list.

Type * toVectorTy(Type *Scalar, ElementCount EC)

A helper function for converting Scalar types to vector types.

@ Default

The result values are uniform if and only if all operands are uniform.

const TypeConversionCostTblEntryT< CostType > * ConvertCostTableLookup(ArrayRef< TypeConversionCostTblEntryT< CostType > > Tbl, int ISD, MVT Dst, MVT Src)

Find in type conversion cost table.

constexpr uint64_t NextPowerOf2(uint64_t A)

Returns the next power of two (in 64-bits) that is strictly greater than A.

This struct is a compact representation of a valid (non-zero power of two) alignment.

bool isSimple() const

Test if the given EVT is simple (as opposed to being extended).

static EVT getVectorVT(LLVMContext &Context, EVT VT, unsigned NumElements, bool IsScalable=false)

Returns the EVT that represents a vector NumElements in length, where each element is of type VT.

bool bitsGT(EVT VT) const

Return true if this has more bits than VT.

TypeSize getSizeInBits() const

Return the size of the specified value type in bits.

uint64_t getScalarSizeInBits() const

static EVT getEVT(Type *Ty, bool HandleUnknown=false)

Return the value type corresponding to the specified type.

MVT getSimpleVT() const

Return the SimpleValueType held in the specified simple EVT.

bool isFixedLengthVector() const

Type * getTypeForEVT(LLVMContext &Context) const

This method returns an LLVM type corresponding to the specified EVT.

bool isScalableVector() const

Return true if this is a vector type where the runtime length is machine dependent.

EVT getVectorElementType() const

Given a vector type, return the type of each element.

unsigned getVectorNumElements() const

Given a vector type, return the number of elements it contains.

This struct is a compact representation of a valid (power of two) or undefined (0) alignment.

Information about a load/store intrinsic defined by the target.

InterleavedAccessInfo * IAI

LoopVectorizationLegality * LVL

This represents an addressing mode of: BaseGV + BaseOffs + BaseReg + Scale*ScaleReg + ScalableOffset*...

unsigned Insns

TODO: Some of these could be merged.

Returns options for expansion of memcmp. IsZeroCmp is.

OperandValueInfo getNoProps() const

Parameters that control the generic loop unrolling transformation.

bool UpperBound

Allow using trip count upper bound to unroll loops.

unsigned PartialOptSizeThreshold

The cost threshold for the unrolled loop when optimizing for size, like OptSizeThreshold,...

unsigned DefaultUnrollRuntimeCount

Default unroll count for loops with run-time trip count.

unsigned SCEVExpansionBudget

Don't allow runtime unrolling if expanding the trip count takes more than SCEVExpansionBudget.

unsigned UnrollAndJamInnerLoopThreshold

Threshold for unroll and jam, for inner loop size.

bool UnrollAndJam

Allow unroll and jam. Used to enable unroll and jam for the target.

bool UnrollRemainder

Allow unrolling of all the iterations of the runtime loop remainder.

unsigned PartialThreshold

The cost threshold for the unrolled loop, like Threshold, but used for partial/runtime unrolling (set...

bool Runtime

Allow runtime unrolling (unrolling of loops to expand the size of the loop body even when the number ...

bool Partial

Allow partial unrolling (unrolling of loops to expand the size of the loop body, not only to eliminat...

Type Conversion Cost Table.