1//==-- AArch64ISelLowering.h - AArch64 DAG Lowering Interface ----*- C++ -*-==//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file defines the interfaces that AArch64 uses to lower LLVM code into a
10// selection DAG.
11//
12//===----------------------------------------------------------------------===//
13
14#ifndef LLVM_LIB_TARGET_AARCH64_AARCH64ISELLOWERING_H
15#define LLVM_LIB_TARGET_AARCH64_AARCH64ISELLOWERING_H
16
17#include "llvm/CodeGen/CallingConvLower.h"
18#include "llvm/CodeGen/MachineFunction.h"
19#include "llvm/CodeGen/SelectionDAG.h"
20#include "llvm/CodeGen/TargetLowering.h"
21#include "llvm/IR/CallingConv.h"
22#include "llvm/IR/Instruction.h"
23
24namespace llvm {
25
26class AArch64TargetMachine;
27
28namespace AArch64 {
29/// Possible values of current rounding mode, which is specified in bits
30/// 23:22 of FPCR.
31enum Rounding {
32 RN = 0, // Round to Nearest
33 RP = 1, // Round towards Plus infinity
34 RM = 2, // Round towards Minus infinity
35 RZ = 3, // Round towards Zero
36 rmMask = 3 // Bit mask selecting rounding mode
37};
38
39// Bit position of rounding mode bits in FPCR.
40const unsigned RoundingBitsPos = 22;
41
42// Reserved bits should be preserved when modifying FPCR.
43const uint64_t ReservedFPControlBits = 0xfffffffff80040f8;
44
45// Registers used to pass function arguments.
46ArrayRef<MCPhysReg> getGPRArgRegs();
47ArrayRef<MCPhysReg> getFPRArgRegs();
48
49/// Maximum allowed number of unprobed bytes above SP at an ABI
50/// boundary.
51const unsigned StackProbeMaxUnprobedStack = 1024;
52
53/// Maximum number of iterations to unroll for a constant size probing loop.
54const unsigned StackProbeMaxLoopUnroll = 4;
55
56} // namespace AArch64
57
58namespace ARM64AS {
59enum : unsigned { PTR32_SPTR = 270, PTR32_UPTR = 271, PTR64 = 272 };
60}
61
62class AArch64Subtarget;
63
64class AArch64TargetLowering : public TargetLowering {
65public:
66 explicit AArch64TargetLowering(const TargetMachine &TM,
67 const AArch64Subtarget &STI);
68
69 const AArch64TargetMachine &getTM() const;
70
71 /// Control the following reassociation of operands: (op (op x, c1), y) -> (op
72 /// (op x, y), c1) where N0 is (op x, c1) and N1 is y.
73 bool isReassocProfitable(SelectionDAG &DAG, SDValue N0,
74 SDValue N1) const override;
75
76 /// Selects the correct CCAssignFn for a given CallingConvention value.
77 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool IsVarArg) const;
78
79 /// Selects the correct CCAssignFn for a given CallingConvention value.
80 CCAssignFn *CCAssignFnForReturn(CallingConv::ID CC) const;
81
82 /// Determine which of the bits specified in Mask are known to be either zero
83 /// or one and return them in the KnownZero/KnownOne bitsets.
84 void computeKnownBitsForTargetNode(const SDValue Op, KnownBits &Known,
85 const APInt &DemandedElts,
86 const SelectionDAG &DAG,
87 unsigned Depth = 0) const override;
88
89 unsigned ComputeNumSignBitsForTargetNode(SDValue Op,
90 const APInt &DemandedElts,
91 const SelectionDAG &DAG,
92 unsigned Depth) const override;
93
94 unsigned computeNumSignBitsForTargetInstr(GISelValueTracking &Analysis,
95 Register R,
96 const APInt &DemandedElts,
97 const MachineRegisterInfo &MRI,
98 unsigned Depth = 0) const override;
99
100 MVT getPointerTy(const DataLayout &DL, uint32_t AS = 0) const override {
101 if ((AS == ARM64AS::PTR32_SPTR) || (AS == ARM64AS::PTR32_UPTR)) {
102 // These are 32-bit pointers created using the `__ptr32` extension or
103 // similar. They are handled by marking them as being in a different
104 // address space, and will be extended to 64-bits when used as the target
105 // of a load or store operation, or cast to a 64-bit pointer type.
106 return MVT::i32;
107 } else {
108 // Returning i64 unconditionally here (i.e. even for ILP32) means that the
109 // *DAG* representation of pointers will always be 64-bits. They will be
110 // truncated and extended when transferred to memory, but the 64-bit DAG
111 // allows us to use AArch64's addressing modes much more easily.
112 return MVT::i64;
113 }
114 }
115
116 unsigned getVectorIdxWidth(const DataLayout &DL) const override {
117 // The VectorIdx type is i64, with both normal and ilp32.
118 return 64;
119 }
120
121 bool targetShrinkDemandedConstant(SDValue Op, const APInt &DemandedBits,
122 const APInt &DemandedElts,
123 TargetLoweringOpt &TLO) const override;
124
125 MVT getScalarShiftAmountTy(const DataLayout &DL, EVT) const override;
126
127 /// Returns true if the target allows unaligned memory accesses of the
128 /// specified type.
129 bool allowsMisalignedMemoryAccesses(
130 EVT VT, unsigned AddrSpace = 0, Align Alignment = Align(1),
131 MachineMemOperand::Flags Flags = MachineMemOperand::MONone,
132 unsigned *Fast = nullptr) const override;
133 /// LLT variant.
134 bool allowsMisalignedMemoryAccesses(LLT Ty, unsigned AddrSpace,
135 Align Alignment,
136 MachineMemOperand::Flags Flags,
137 unsigned *Fast = nullptr) const override;
138
139 /// Provide custom lowering hooks for some operations.
140 SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
141
142 SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
143
144 /// This method returns a target specific FastISel object, or null if the
145 /// target does not support "fast" ISel.
146 FastISel *
147 createFastISel(FunctionLoweringInfo &funcInfo,
148 const TargetLibraryInfo *libInfo,
149 const LibcallLoweringInfo *libcallLowering) const override;
150
151 bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const override;
152
153 bool isFPImmLegalAsFMov(const APFloat &Imm, EVT VT) const;
154
155 bool isFPImmLegal(const APFloat &Imm, EVT VT,
156 bool ForCodeSize) const override;
157
158 /// Return true if the given shuffle mask can be codegen'd directly, or if it
159 /// should be stack expanded.
160 bool isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const override;
161
162 /// Similar to isShuffleMaskLegal. Return true is the given 'select with zero'
163 /// shuffle mask can be codegen'd directly.
164 bool isVectorClearMaskLegal(ArrayRef<int> M, EVT VT) const override;
165
166 /// Return the ISD::SETCC ValueType.
167 EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
168 EVT VT) const override;
169
170 SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const;
171
172 MachineBasicBlock *EmitF128CSEL(MachineInstr &MI,
173 MachineBasicBlock *BB) const;
174
175 MachineBasicBlock *EmitLoweredCatchRet(MachineInstr &MI,
176 MachineBasicBlock *BB) const;
177
178 MachineBasicBlock *EmitLoweredSetFpmr(MachineInstr &MI,
179 MachineBasicBlock *MBB) const;
180
181 MachineBasicBlock *EmitDynamicProbedAlloc(MachineInstr &MI,
182 MachineBasicBlock *MBB) const;
183
184 MachineBasicBlock *EmitCheckMatchingVL(MachineInstr &MI,
185 MachineBasicBlock *MBB) const;
186
187 MachineBasicBlock *EmitTileLoad(unsigned Opc, unsigned BaseReg,
188 MachineInstr &MI,
189 MachineBasicBlock *BB) const;
190 MachineBasicBlock *EmitFill(MachineInstr &MI, MachineBasicBlock *BB) const;
191 MachineBasicBlock *EmitZAInstr(unsigned Opc, unsigned BaseReg,
192 MachineInstr &MI, MachineBasicBlock *BB) const;
193 MachineBasicBlock *EmitZTInstr(MachineInstr &MI, MachineBasicBlock *BB,
194 unsigned Opcode, bool Op0IsDef) const;
195 MachineBasicBlock *EmitZero(MachineInstr &MI, MachineBasicBlock *BB) const;
196 MachineBasicBlock *EmitEntryPStateSM(MachineInstr &MI,
197 MachineBasicBlock *BB) const;
198
199 /// Replace (0, vreg) discriminator components with the operands of blend
200 /// or with (immediate, NoRegister) when possible.
201 void fixupPtrauthDiscriminator(MachineInstr &MI, MachineBasicBlock *BB,
202 MachineOperand &IntDiscOp,
203 MachineOperand &AddrDiscOp,
204 const TargetRegisterClass *AddrDiscRC) const;
205
206 MachineBasicBlock *
207 EmitInstrWithCustomInserter(MachineInstr &MI,
208 MachineBasicBlock *MBB) const override;
209
210 void getTgtMemIntrinsic(SmallVectorImpl<IntrinsicInfo> &Infos,
211 const CallBase &I, MachineFunction &MF,
212 unsigned Intrinsic) const override;
213
214 bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy, EVT NewVT,
215 std::optional<unsigned> ByteOffset) const override;
216
217 bool shouldRemoveRedundantExtend(SDValue Op) const override;
218
219 bool isTruncateFree(Type *Ty1, Type *Ty2) const override;
220 bool isTruncateFree(EVT VT1, EVT VT2) const override;
221
222 bool isProfitableToHoist(Instruction *I) const override;
223
224 bool isZExtFree(Type *Ty1, Type *Ty2) const override;
225 bool isZExtFree(EVT VT1, EVT VT2) const override;
226 bool isZExtFree(SDValue Val, EVT VT2) const override;
227
228 bool optimizeExtendOrTruncateConversion(
229 Instruction *I, Loop *L, const TargetTransformInfo &TTI) const override;
230
231 bool hasPairedLoad(EVT LoadedType, Align &RequiredAlignment) const override;
232
233 unsigned getMaxSupportedInterleaveFactor() const override { return 4; }
234
235 bool lowerInterleavedLoad(Instruction *Load, Value *Mask,
236 ArrayRef<ShuffleVectorInst *> Shuffles,
237 ArrayRef<unsigned> Indices, unsigned Factor,
238 const APInt &GapMask) const override;
239 bool lowerInterleavedStore(Instruction *Store, Value *Mask,
240 ShuffleVectorInst *SVI, unsigned Factor,
241 const APInt &GapMask) const override;
242
243 bool lowerDeinterleaveIntrinsicToLoad(Instruction *Load, Value *Mask,
244 IntrinsicInst *DI,
245 const APInt &GapMask) const override;
246
247 bool lowerInterleaveIntrinsicToStore(
248 Instruction *Store, Value *Mask,
249 ArrayRef<Value *> InterleaveValues) const override;
250
251 bool isLegalAddImmediate(int64_t) const override;
252 bool isLegalAddScalableImmediate(int64_t) const override;
253 bool isLegalICmpImmediate(int64_t) const override;
254
255 bool isMulAddWithConstProfitable(SDValue AddNode,
256 SDValue ConstNode) const override;
257
258 bool shouldConsiderGEPOffsetSplit() const override;
259
260 EVT getOptimalMemOpType(LLVMContext &Context, const MemOp &Op,
261 const AttributeList &FuncAttributes) const override;
262
263 LLT getOptimalMemOpLLT(const MemOp &Op,
264 const AttributeList &FuncAttributes) const override;
265
266 bool findOptimalMemOpLowering(LLVMContext &Context, std::vector<EVT> &MemOps,
267 unsigned Limit, const MemOp &Op, unsigned DstAS,
268 unsigned SrcAS,
269 const AttributeList &FuncAttributes,
270 EVT *LargestVT = nullptr) const override;
271
272 /// Return true if the addressing mode represented by AM is legal for this
273 /// target, for a load/store of the specified type.
274 bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, Type *Ty,
275 unsigned AS,
276 Instruction *I = nullptr) const override;
277
278 int64_t getPreferredLargeGEPBaseOffset(int64_t MinOffset,
279 int64_t MaxOffset) const override;
280
281 /// Return true if an FMA operation is faster than a pair of fmul and fadd
282 /// instructions. fmuladd intrinsics will be expanded to FMAs when this method
283 /// returns true, otherwise fmuladd is expanded to fmul + fadd.
284 bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
285 EVT VT) const override;
286 bool isFMAFasterThanFMulAndFAdd(const Function &F, Type *Ty) const override;
287
288 bool generateFMAsInMachineCombiner(EVT VT,
289 CodeGenOptLevel OptLevel) const override;
290
291 /// Return true if the target has native support for
292 /// the specified value type and it is 'desirable' to use the type for the
293 /// given node type.
294 bool isTypeDesirableForOp(unsigned Opc, EVT VT) const override;
295
296 const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const override;
297 ArrayRef<MCPhysReg> getRoundingControlRegisters() const override;
298
299 bool isNarrowingProfitable(SDNode *N, EVT SrcVT, EVT DestVT) const override;
300
301 /// Returns false if N is a bit extraction pattern of (X >> C) & Mask.
302 bool isDesirableToCommuteWithShift(const SDNode *N,
303 CombineLevel Level) const override;
304
305 bool isDesirableToPullExtFromShl(const MachineInstr &MI) const override {
306 return false;
307 }
308
309 /// Returns false if N is a bit extraction pattern of (X >> C) & Mask.
310 bool isDesirableToCommuteXorWithShift(const SDNode *N) const override;
311
312 /// Return true if it is profitable to fold a pair of shifts into a mask.
313 bool shouldFoldConstantShiftPairToMask(const SDNode *N) const override;
314
315 /// Return true if it is profitable to fold a pair of shifts into a mask.
316 bool shouldFoldMaskToVariableShiftPair(SDValue Y) const override {
317 EVT VT = Y.getValueType();
318
319 if (VT.isVector())
320 return false;
321
322 return VT.getScalarSizeInBits() <= 64;
323 }
324
325 bool shouldFoldSelectWithIdentityConstant(unsigned BinOpcode, EVT VT,
326 unsigned SelectOpcode, SDValue X,
327 SDValue Y) const override;
328
329 /// Returns true if it is beneficial to convert a load of a constant
330 /// to just the constant itself.
331 bool shouldConvertConstantLoadToIntImm(const APInt &Imm,
332 Type *Ty) const override;
333
334 /// Return the cost of EXTRACT_SUBVECTOR for this result type with this
335 /// index.
336 ExtractSubvectorCost getExtractSubvectorCost(EVT ResVT, EVT SrcVT,
337 unsigned Index) const override;
338
339 bool shouldFormOverflowOp(unsigned Opcode, EVT VT,
340 bool MathUsed) const override {
341 // Using overflow ops for overflow checks only should beneficial on
342 // AArch64.
343 return TargetLowering::shouldFormOverflowOp(Opcode, VT, MathUsed: true);
344 }
345
346 // Return true if the target wants to optimize the mul overflow intrinsic
347 // for the given \p VT.
348 bool shouldOptimizeMulOverflowWithZeroHighBits(LLVMContext &Context,
349 EVT VT) const override;
350
351 Instruction *emitLeadingFence(IRBuilderBase &Builder, Instruction *Inst,
352 AtomicOrdering Ord) const override;
353 Value *emitLoadLinked(IRBuilderBase &Builder, Type *ValueTy, Value *Addr,
354 AtomicOrdering Ord) const override;
355 Value *emitStoreConditional(IRBuilderBase &Builder, Value *Val, Value *Addr,
356 AtomicOrdering Ord) const override;
357 Value *emitCanLoadSpeculatively(IRBuilderBase &Builder, Value *Ptr,
358 Value *SizeInBytes) const override;
359
360 void emitAtomicCmpXchgNoStoreLLBalance(IRBuilderBase &Builder) const override;
361
362 bool isOpSuitableForLDPSTP(const Instruction *I) const;
363 bool isOpSuitableForLSE128(const Instruction *I) const;
364 bool isOpSuitableForRCPC3(const Instruction *I) const;
365 bool shouldInsertFencesForAtomic(const Instruction *I) const override;
366 bool shouldInsertTrailingSeqCstFenceForAtomicStore(
367 const Instruction *I) const override;
368
369 TargetLoweringBase::AtomicExpansionKind
370 shouldExpandAtomicLoadInIR(LoadInst *LI) const override;
371
372 TargetLoweringBase::AtomicExpansionKind
373 shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
374 TargetLoweringBase::AtomicExpansionKind
375 shouldExpandAtomicRMWInIR(const AtomicRMWInst *AI) const override;
376
377 TargetLoweringBase::AtomicExpansionKind
378 shouldExpandAtomicCmpXchgInIR(const AtomicCmpXchgInst *AI) const override;
379
380 bool shouldIssueAtomicLoadForAtomicEmulationLoop() const override {
381 return false;
382 }
383
384 bool useLoadStackGuardNode(const Module &M) const override;
385 bool useStackGuardMixFP() const override;
386 SDValue emitStackGuardMixFP(SelectionDAG &DAG, SDValue Val,
387 const SDLoc &DL) const override;
388 TargetLoweringBase::LegalizeTypeAction
389 getPreferredVectorAction(MVT VT) const override;
390
391 /// If the target has a standard location for the stack protector cookie,
392 /// returns the address of that location. Otherwise, returns nullptr.
393 Value *getIRStackGuard(IRBuilderBase &IRB,
394 const LibcallLoweringInfo &Libcalls) const override;
395
396 void
397 insertSSPDeclarations(Module &M,
398 const LibcallLoweringInfo &Libcalls) const override;
399
400 /// If the target has a standard location for the unsafe stack pointer,
401 /// returns the address of that location. Otherwise, returns nullptr.
402 Value *getSafeStackPointerLocation(
403 IRBuilderBase &IRB, const LibcallLoweringInfo &Libcalls) const override;
404
405 /// If a physical register, this returns the register that receives the
406 /// exception address on entry to an EH pad.
407 Register
408 getExceptionPointerRegister(ExceptionHandling EH,
409 const Constant *PersonalityFn) const override;
410
411 /// If a physical register, this returns the register that receives the
412 /// exception typeid on entry to a landing pad.
413 Register
414 getExceptionSelectorRegister(ExceptionHandling EH,
415 const Constant *PersonalityFn) const override;
416
417 bool isIntDivCheap(EVT VT, AttributeList Attr) const override;
418
419 bool canMergeStoresTo(unsigned AddressSpace, EVT MemVT,
420 const MachineFunction &MF) const override;
421
422 bool isCheapToSpeculateCttz(Type *) const override {
423 return true;
424 }
425
426 bool isCheapToSpeculateCtlz(Type *) const override {
427 return true;
428 }
429
430 bool isMaskAndCmp0FoldingBeneficial(const Instruction &AndI) const override;
431
432 bool hasAndNotCompare(SDValue V) const override {
433 // We can use bics for any scalar.
434 return V.getValueType().isScalarInteger();
435 }
436
437 bool hasAndNot(SDValue Y) const override {
438 EVT VT = Y.getValueType();
439
440 if (!VT.isVector())
441 return hasAndNotCompare(V: Y);
442
443 if (VT.isScalableVector())
444 return true;
445
446 return VT.getFixedSizeInBits() >= 64; // vector 'bic'
447 }
448
449 bool shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(
450 SDValue X, ConstantSDNode *XC, ConstantSDNode *CC, SDValue Y,
451 unsigned OldShiftOpcode, unsigned NewShiftOpcode,
452 SelectionDAG &DAG) const override;
453
454 ShiftLegalizationStrategy
455 preferredShiftLegalizationStrategy(SelectionDAG &DAG, SDNode *N,
456 unsigned ExpansionFactor) const override;
457
458 CondMergingParams
459 getJumpConditionMergingParams(Instruction::BinaryOps Opc, const Value *Lhs,
460 const Value *Rhs,
461 const Function *F) const override;
462
463 bool shouldTransformSignedTruncationCheck(EVT XVT,
464 unsigned KeptBits) const override {
465 // For vectors, we don't have a preference..
466 if (XVT.isVector())
467 return false;
468
469 auto VTIsOk = [](EVT VT) -> bool {
470 return VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32 ||
471 VT == MVT::i64;
472 };
473
474 // We are ok with KeptBitsVT being byte/word/dword, what SXT supports.
475 // XVT will be larger than KeptBitsVT.
476 MVT KeptBitsVT = MVT::getIntegerVT(BitWidth: KeptBits);
477 return VTIsOk(XVT) && VTIsOk(KeptBitsVT);
478 }
479
480 bool preferIncOfAddToSubOfNot(EVT VT) const override;
481
482 bool shouldConvertFpToSat(unsigned Op, EVT FPVT, EVT VT) const override;
483
484 bool preferSelectsOverBooleanArithmetic(EVT VT) const override;
485
486 bool isComplexDeinterleavingSupported() const override;
487 bool isComplexDeinterleavingOperationSupported(
488 ComplexDeinterleavingOperation Operation, Type *Ty) const override;
489
490 Value *createComplexDeinterleavingIR(
491 IRBuilderBase &B, ComplexDeinterleavingOperation OperationType,
492 ComplexDeinterleavingRotation Rotation, Value *InputA, Value *InputB,
493 Value *Accumulator = nullptr) const override;
494
495 bool supportSplitCSR(MachineFunction *MF) const override {
496 return MF->getFunction().getCallingConv() == CallingConv::CXX_FAST_TLS &&
497 MF->getFunction().hasFnAttribute(Kind: Attribute::NoUnwind);
498 }
499 void initializeSplitCSR(MachineBasicBlock *Entry) const override;
500 void insertCopiesSplitCSR(
501 MachineBasicBlock *Entry,
502 const SmallVectorImpl<MachineBasicBlock *> &Exits) const override;
503
504 bool supportSwiftError() const override {
505 return true;
506 }
507
508 bool supportPtrAuthBundles() const override { return true; }
509
510 bool supportKCFIBundles() const override { return true; }
511
512 MachineInstr *EmitKCFICheck(MachineBasicBlock &MBB,
513 MachineBasicBlock::instr_iterator &MBBI,
514 const TargetInstrInfo *TII) const override;
515
516 bool shallExtractConstSplatVectorElementToStore(
517 Type *VectorTy, unsigned ElemSizeInBits, unsigned &Index) const override;
518
519 /// Enable aggressive FMA fusion on targets that want it.
520 bool enableAggressiveFMAFusion(EVT VT) const override;
521
522 bool aggressivelyPreferBuildVectorSources(EVT VecVT) const override {
523 return true;
524 }
525
526 /// Returns the size of the platform's va_list object.
527 unsigned getVaListSizeInBits(const DataLayout &DL) const override;
528
529 /// Returns true if \p VecTy is a legal interleaved access type. This
530 /// function checks the vector element type and the overall width of the
531 /// vector.
532 bool isLegalInterleavedAccessType(VectorType *VecTy, const DataLayout &DL,
533 bool &UseScalable) const;
534
535 /// Returns the number of interleaved accesses that will be generated when
536 /// lowering accesses of the given type.
537 unsigned getNumInterleavedAccesses(VectorType *VecTy, const DataLayout &DL,
538 bool UseScalable) const;
539
540 MachineMemOperand::Flags getTargetMMOFlags(
541 const Instruction &I) const override;
542
543 bool functionArgumentNeedsConsecutiveRegisters(
544 Type *Ty, CallingConv::ID CallConv, bool isVarArg,
545 const DataLayout &DL) const override;
546
547 /// Used for exception handling on Win64.
548 bool needsFixedCatchObjects() const override;
549
550 bool fallBackToDAGISel(const Instruction &Inst) const override;
551
552 /// SVE code generation for fixed length vectors does not custom lower
553 /// BUILD_VECTOR. This makes BUILD_VECTOR legalisation a source of stores to
554 /// merge. However, merging them creates a BUILD_VECTOR that is just as
555 /// illegal as the original, thus leading to an infinite legalisation loop.
556 /// NOTE: Once BUILD_VECTOR is legal or can be custom lowered for all legal
557 /// vector types this override can be removed.
558 bool mergeStoresAfterLegalization(EVT VT) const override;
559
560 // If the platform/function should have a redzone, return the size in bytes.
561 unsigned getRedZoneSize(const Function &F) const {
562 if (F.hasFnAttribute(Kind: Attribute::NoRedZone))
563 return 0;
564 return 128;
565 }
566
567 bool isAllActivePredicate(const SelectionDAG &DAG, SDValue N) const;
568 EVT getPromotedVTForPredicate(EVT VT) const;
569
570 EVT getAsmOperandValueType(const DataLayout &DL, Type *Ty,
571 bool AllowUnknown = false) const override;
572
573 bool shouldExpandGetActiveLaneMask(EVT VT, EVT OpVT) const override;
574
575 /// If a change in streaming mode is required on entry to/return from a
576 /// function call it emits and returns the corresponding SMSTART or SMSTOP
577 /// node. \p Condition should be one of the enum values from
578 /// AArch64SME::ToggleCondition.
579 SDValue changeStreamingMode(SelectionDAG &DAG, SDLoc DL, bool Enable,
580 SDValue Chain, SDValue InGlue, unsigned Condition,
581 bool InsertVectorLengthCheck = false) const;
582
583 /// Returns true if \p RdxOp should be lowered to a SVE reduction. If a SVE2
584 /// pairwise operation can be used for the reduction \p PairwiseOpIID is set
585 /// to its intrinsic ID.
586 bool
587 shouldLowerReductionToSVE(SDValue RdxOp,
588 std::optional<Intrinsic::ID> &PairwiseOpIID) const;
589
590 // Normally SVE is only used for byte size vectors that do not fit within a
591 // NEON vector. This changes when OverrideNEON is true, allowing SVE to be
592 // used for 64bit and 128bit vectors as well.
593 bool useSVEForFixedLengthVectorVT(EVT VT, bool OverrideNEON = false) const;
594
595 // Follow NEON ABI rules even when using SVE for fixed length vectors.
596 MVT getRegisterTypeForCallingConv(LLVMContext &Context, CallingConv::ID CC,
597 EVT VT) const override;
598 unsigned getNumRegistersForCallingConv(LLVMContext &Context,
599 CallingConv::ID CC,
600 EVT VT) const override;
601 unsigned getVectorTypeBreakdownForCallingConv(LLVMContext &Context,
602 CallingConv::ID CC, EVT VT,
603 EVT &IntermediateVT,
604 unsigned &NumIntermediates,
605 MVT &RegisterVT) const override;
606
607 bool preferVectorizedNonPowerOfTwoTypeBreakdown() const override {
608 return true;
609 }
610
611 /// True if stack clash protection is enabled for this functions.
612 bool hasInlineStackProbe(const MachineFunction &MF) const override;
613
614 /// In AArch64, true if FEAT_CPA is present. Allows pointer arithmetic
615 /// semantics to be preserved for instruction selection.
616 bool shouldPreservePtrArith(const Function &F, EVT PtrVT) const override;
617
618 // Match a register name (e.g. "x5", "d5", "sp") to its register number, with
619 // no validity filtering. This is the single entry point for the generated
620 // register-name matcher, shared with getRegisterByName.
621 Register matchRegisterName(StringRef RegName) const;
622
623private:
624 /// Keep a pointer to the AArch64Subtarget around so that we can
625 /// make the right decision when generating code for different targets.
626 const AArch64Subtarget *Subtarget;
627
628 bool isExtFreeImpl(const Instruction *Ext) const override;
629
630 void addTypeForNEON(MVT VT);
631 void addTypeForFixedLengthSVE(MVT VT);
632 void addDRType(MVT VT);
633 void addQRType(MVT VT);
634
635 bool shouldExpandBuildVectorWithShuffles(EVT, unsigned) const override;
636
637 SDValue lowerEHPadEntry(SDValue Chain, SDLoc const &DL,
638 SelectionDAG &DAG) const override;
639
640 SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
641 bool isVarArg,
642 const SmallVectorImpl<ISD::InputArg> &Ins,
643 const SDLoc &DL, SelectionDAG &DAG,
644 SmallVectorImpl<SDValue> &InVals) const override;
645
646 void AdjustInstrPostInstrSelection(MachineInstr &MI,
647 SDNode *Node) const override;
648
649 SDValue LowerCall(CallLoweringInfo & /*CLI*/,
650 SmallVectorImpl<SDValue> &InVals) const override;
651
652 SDValue LowerCallResult(SDValue Chain, SDValue InGlue,
653 CallingConv::ID CallConv, bool isVarArg,
654 const SmallVectorImpl<CCValAssign> &RVLocs,
655 const SDLoc &DL, SelectionDAG &DAG,
656 SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
657 SDValue ThisVal, bool RequiresSMChange) const;
658
659 SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
660 SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
661 SDValue LowerStore128(SDValue Op, SelectionDAG &DAG) const;
662 SDValue LowerABS(SDValue Op, SelectionDAG &DAG) const;
663 SDValue LowerSMULFIXSAT(SDValue Op, SelectionDAG &DAG) const;
664 SDValue LowerFMUL(SDValue Op, SelectionDAG &DAG) const;
665 SDValue LowerFMA(SDValue Op, SelectionDAG &DAG) const;
666 SDValue LowerCLMUL(SDValue Op, SelectionDAG &DAG) const;
667
668 SDValue LowerMGATHER(SDValue Op, SelectionDAG &DAG) const;
669 SDValue LowerMSCATTER(SDValue Op, SelectionDAG &DAG) const;
670
671 SDValue LowerMLOAD(SDValue Op, SelectionDAG &DAG) const;
672
673 SDValue LowerVECTOR_COMPRESS(SDValue Op, SelectionDAG &DAG) const;
674
675 SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) const;
676 SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
677 SDValue LowerINTRINSIC_VOID(SDValue Op, SelectionDAG &DAG) const;
678
679 bool
680 isEligibleForTailCallOptimization(const CallLoweringInfo &CLI) const;
681
682 /// Finds the incoming stack arguments which overlap the given fixed stack
683 /// object and incorporates their load into the current chain. This prevents
684 /// an upcoming store from clobbering the stack argument before it's used.
685 SDValue addTokenForArgument(SDValue Chain, SelectionDAG &DAG,
686 MachineFrameInfo &MFI, int ClobberedFI) const;
687
688 bool DoesCalleeRestoreStack(CallingConv::ID CallCC, bool TailCallOpt) const;
689
690 void saveVarArgRegisters(CCState &CCInfo, SelectionDAG &DAG, const SDLoc &DL,
691 SDValue &Chain) const;
692
693 bool CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF,
694 bool isVarArg,
695 const SmallVectorImpl<ISD::OutputArg> &Outs,
696 LLVMContext &Context, const Type *RetTy) const override;
697
698 SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
699 const SmallVectorImpl<ISD::OutputArg> &Outs,
700 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
701 SelectionDAG &DAG) const override;
702
703 SDValue getTargetNode(GlobalAddressSDNode *N, EVT Ty, SelectionDAG &DAG,
704 unsigned Flag) const;
705 SDValue getTargetNode(JumpTableSDNode *N, EVT Ty, SelectionDAG &DAG,
706 unsigned Flag) const;
707 SDValue getTargetNode(ConstantPoolSDNode *N, EVT Ty, SelectionDAG &DAG,
708 unsigned Flag) const;
709 SDValue getTargetNode(BlockAddressSDNode *N, EVT Ty, SelectionDAG &DAG,
710 unsigned Flag) const;
711 SDValue getTargetNode(ExternalSymbolSDNode *N, EVT Ty, SelectionDAG &DAG,
712 unsigned Flag) const;
713 template <class NodeTy>
714 SDValue getGOT(NodeTy *N, SelectionDAG &DAG, unsigned Flags = 0) const;
715 template <class NodeTy>
716 SDValue getAddrLarge(NodeTy *N, SelectionDAG &DAG, unsigned Flags = 0) const;
717 template <class NodeTy>
718 SDValue getAddr(NodeTy *N, SelectionDAG &DAG, unsigned Flags = 0) const;
719 template <class NodeTy>
720 SDValue getAddrTiny(NodeTy *N, SelectionDAG &DAG, unsigned Flags = 0) const;
721 SDValue LowerADDROFRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
722 SDValue LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
723 SDValue LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
724 SDValue LowerDarwinGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
725 SDValue LowerELFGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
726 SDValue LowerELFTLSLocalExec(const GlobalValue *GV, SDValue ThreadBase,
727 const SDLoc &DL, SelectionDAG &DAG) const;
728 SDValue LowerELFTLSDescCallSeq(SDValue SymAddr, const SDLoc &DL,
729 SelectionDAG &DAG) const;
730 SDValue LowerWindowsGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const;
731 SDValue LowerPtrAuthGlobalAddress(SDValue Op, SelectionDAG &DAG) const;
732 SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
733 SDValue LowerSETCCCARRY(SDValue Op, SelectionDAG &DAG) const;
734 SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
735 SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
736 SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
737 SDValue LowerSELECT_CC(ISD::CondCode CC, SDValue LHS, SDValue RHS,
738 SDValue TVal, SDValue FVal,
739 iterator_range<SDNode::user_iterator> Users,
740 SDNodeFlags Flags, const SDLoc &dl,
741 SelectionDAG &DAG) const;
742 SDValue LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
743 SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const;
744 SDValue LowerJumpTable(SDValue Op, SelectionDAG &DAG) const;
745 SDValue LowerBR_JT(SDValue Op, SelectionDAG &DAG) const;
746 SDValue LowerBRIND(SDValue Op, SelectionDAG &DAG) const;
747 SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const;
748 SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const;
749 SDValue LowerAAPCS_VASTART(SDValue Op, SelectionDAG &DAG) const;
750 SDValue LowerDarwin_VASTART(SDValue Op, SelectionDAG &DAG) const;
751 SDValue LowerWin64_VASTART(SDValue Op, SelectionDAG &DAG) const;
752 SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const;
753 SDValue LowerVACOPY(SDValue Op, SelectionDAG &DAG) const;
754 SDValue LowerVAARG(SDValue Op, SelectionDAG &DAG) const;
755 SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
756 SDValue LowerSPONENTRY(SDValue Op, SelectionDAG &DAG) const;
757 SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
758 SDValue LowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
759 SDValue LowerSET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
760 SDValue LowerGET_FPMODE(SDValue Op, SelectionDAG &DAG) const;
761 SDValue LowerSET_FPMODE(SDValue Op, SelectionDAG &DAG) const;
762 SDValue LowerRESET_FPMODE(SDValue Op, SelectionDAG &DAG) const;
763 SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
764 SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
765 SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
766 SDValue LowerEXTEND_VECTOR_INREG(SDValue Op, SelectionDAG &DAG) const;
767 SDValue LowerZERO_EXTEND_VECTOR_INREG(SDValue Op, SelectionDAG &DAG) const;
768 SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
769 SDValue LowerSPLAT_VECTOR(SDValue Op, SelectionDAG &DAG) const;
770 SDValue LowerDUPQLane(SDValue Op, SelectionDAG &DAG) const;
771 SDValue LowerToPredicatedOp(SDValue Op, SelectionDAG &DAG,
772 unsigned NewOp) const;
773 SDValue LowerToScalableOp(SDValue Op, SelectionDAG &DAG) const;
774 SDValue LowerVECTOR_SPLICE(SDValue Op, SelectionDAG &DAG) const;
775 SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
776 SDValue LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
777 SDValue LowerVECTOR_DEINTERLEAVE(SDValue Op, SelectionDAG &DAG) const;
778 SDValue LowerVECTOR_INTERLEAVE(SDValue Op, SelectionDAG &DAG) const;
779 SDValue LowerVECTOR_HISTOGRAM(SDValue Op, SelectionDAG &DAG) const;
780 SDValue LowerPARTIAL_REDUCE_MLA(SDValue Op, SelectionDAG &DAG) const;
781 SDValue LowerGET_ACTIVE_LANE_MASK(SDValue Op, SelectionDAG &DAG) const;
782 SDValue LowerDIV(SDValue Op, SelectionDAG &DAG) const;
783 SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
784 SDValue LowerVectorSRA_SRL_SHL(SDValue Op, SelectionDAG &DAG) const;
785 SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) const;
786 SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) const;
787 SDValue LowerCTPOP_PARITY(SDValue Op, SelectionDAG &DAG) const;
788 SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) const;
789 SDValue LowerBitreverse(SDValue Op, SelectionDAG &DAG) const;
790 SDValue LowerMinMax(SDValue Op, SelectionDAG &DAG) const;
791 SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const;
792 SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
793 SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
794 SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) const;
795 SDValue LowerVectorFP_TO_INT_SAT(SDValue Op, SelectionDAG &DAG) const;
796 SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const;
797 SDValue LowerFP_TO_INT_SAT(SDValue Op, SelectionDAG &DAG) const;
798 SDValue LowerVectorXRINT(SDValue Op, SelectionDAG &DAG) const;
799 SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
800 SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
801 SDValue LowerVectorOR(SDValue Op, SelectionDAG &DAG) const;
802 SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) const;
803 SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const;
804 SDValue LowerLOOP_DEPENDENCE_MASK(SDValue Op, SelectionDAG &DAG) const;
805 SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) const;
806 SDValue LowerVSCALE(SDValue Op, SelectionDAG &DAG) const;
807 SDValue LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const;
808 SDValue LowerVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
809 SDValue LowerVECREDUCE_MUL(SDValue Op, SelectionDAG &DAG) const;
810 SDValue LowerATOMIC_LOAD_AND(SDValue Op, SelectionDAG &DAG) const;
811 SDValue LowerWindowsDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
812 SDValue LowerInlineDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
813 SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
814 SDValue LowerMSTORE(SDValue Op, SelectionDAG &DAG) const;
815 SDValue LowerFCANONICALIZE(SDValue Op, SelectionDAG &DAG) const;
816 SDValue LowerAVG(SDValue Op, SelectionDAG &DAG, unsigned NewOp) const;
817
818 SDValue LowerFPToIntToSVE(SDValue Op, SelectionDAG &DAG) const;
819
820 SDValue LowerFixedLengthVectorIntDivideToSVE(SDValue Op,
821 SelectionDAG &DAG) const;
822 SDValue LowerFixedLengthVectorIntExtendToSVE(SDValue Op,
823 SelectionDAG &DAG) const;
824 SDValue LowerFixedLengthVectorLoadToSVE(SDValue Op, SelectionDAG &DAG) const;
825 SDValue LowerFixedLengthVectorMLoadToSVE(SDValue Op, SelectionDAG &DAG) const;
826 SDValue LowerVECREDUCE_SEQ_FADD(SDValue ScalarOp, SelectionDAG &DAG) const;
827 SDValue LowerPredReductionToSVE(SDValue ScalarOp, SelectionDAG &DAG) const;
828 SDValue LowerReductionToSVE(SDValue Op, SelectionDAG &DAG) const;
829 SDValue LowerFixedLengthVectorSelectToSVE(SDValue Op, SelectionDAG &DAG) const;
830 SDValue LowerFixedLengthVectorSetccToSVE(SDValue Op, SelectionDAG &DAG) const;
831 SDValue LowerFixedLengthVectorStoreToSVE(SDValue Op, SelectionDAG &DAG) const;
832 SDValue LowerFixedLengthVectorMStoreToSVE(SDValue Op,
833 SelectionDAG &DAG) const;
834 SDValue LowerFixedLengthVectorTruncateToSVE(SDValue Op,
835 SelectionDAG &DAG) const;
836 SDValue LowerFixedLengthExtractVectorElt(SDValue Op, SelectionDAG &DAG) const;
837 SDValue LowerFixedLengthInsertVectorElt(SDValue Op, SelectionDAG &DAG) const;
838 SDValue LowerFixedLengthBitcastToSVE(SDValue Op, SelectionDAG &DAG) const;
839 SDValue LowerFixedLengthConcatVectorsToSVE(SDValue Op,
840 SelectionDAG &DAG) const;
841 SDValue LowerFixedLengthFPExtendToSVE(SDValue Op, SelectionDAG &DAG) const;
842 SDValue LowerFixedLengthFPRoundToSVE(SDValue Op, SelectionDAG &DAG) const;
843 SDValue LowerFixedLengthIntToFPToSVE(SDValue Op, SelectionDAG &DAG) const;
844 SDValue LowerFixedLengthFPToIntToSVE(SDValue Op, SelectionDAG &DAG) const;
845 SDValue LowerFixedLengthVECTOR_SHUFFLEToSVE(SDValue Op,
846 SelectionDAG &DAG) const;
847 SDValue LowerFixedLengthBuildVectorToSVE(SDValue Op, SelectionDAG &DAG) const;
848 SDValue LowerFixedLengthVectorCompressToSVE(SDValue Op,
849 SelectionDAG &DAG) const;
850
851 SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
852 SmallVectorImpl<SDNode *> &Created) const override;
853 SDValue BuildSREMPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
854 SmallVectorImpl<SDNode *> &Created) const override;
855 SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
856 int &ExtraSteps, bool &UseOneConst,
857 bool Reciprocal) const override;
858 SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
859 int &ExtraSteps) const override;
860 SDValue getSqrtInputTest(SDValue Operand, SelectionDAG &DAG,
861 const DenormalMode &Mode,
862 SDNodeFlags Flags = {}) const override;
863 SDValue getSqrtResultForDenormInput(SDValue Operand,
864 SelectionDAG &DAG) const override;
865 unsigned combineRepeatedFPDivisors() const override;
866
867 ConstraintType getConstraintType(StringRef Constraint) const override;
868 Register getRegisterByName(const char* RegName, LLT VT,
869 const MachineFunction &MF) const override;
870
871private:
872 /// Examine constraint string and operand type and determine a weight value.
873 /// The operand object must already have been set up with the operand type.
874 ConstraintWeight
875 getSingleConstraintMatchWeight(AsmOperandInfo &info,
876 const char *constraint) const override;
877
878 std::pair<unsigned, const TargetRegisterClass *>
879 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
880 StringRef Constraint, MVT VT) const override;
881
882 const char *LowerXConstraint(EVT ConstraintVT) const override;
883
884 void LowerAsmOperandForConstraint(SDValue Op, StringRef Constraint,
885 std::vector<SDValue> &Ops,
886 SelectionDAG &DAG) const override;
887
888 InlineAsm::ConstraintCode
889 getInlineAsmMemConstraint(StringRef ConstraintCode) const override {
890 if (ConstraintCode == "Q")
891 return InlineAsm::ConstraintCode::Q;
892 // FIXME: clang has code for 'Ump', 'Utf', 'Usa', and 'Ush' but these are
893 // followed by llvm_unreachable so we'll leave them unimplemented in
894 // the backend for now.
895 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
896 }
897
898 /// Handle Lowering flag assembly outputs.
899 SDValue LowerAsmOutputForConstraint(SDValue &Chain, SDValue &Flag,
900 const SDLoc &DL,
901 const AsmOperandInfo &Constraint,
902 SelectionDAG &DAG) const override;
903
904 bool shouldExtendGSIndex(EVT VT, EVT &EltTy) const override;
905 bool shouldRemoveExtendFromGSIndex(SDValue Extend, EVT DataVT) const override;
906 bool isVectorLoadExtDesirable(SDValue ExtVal) const override;
907 bool isUsedByReturnOnly(SDNode *N, SDValue &Chain) const override;
908 bool mayBeEmittedAsTailCall(const CallInst *CI) const override;
909 bool getIndexedAddressParts(SDNode *N, SDNode *Op, SDValue &Base,
910 SDValue &Offset, SelectionDAG &DAG,
911 ISD::MemIndexedMode AM) const;
912 bool getPreIndexedAddressParts(SDNode *N, SDValue &Base, SDValue &Offset,
913 ISD::MemIndexedMode &AM,
914 SelectionDAG &DAG) const override;
915 bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, SDValue &Base,
916 SDValue &Offset, ISD::MemIndexedMode &AM,
917 SelectionDAG &DAG) const override;
918 bool isIndexingLegal(MachineInstr &MI, Register Base, Register Offset,
919 bool IsPre, MachineRegisterInfo &MRI) const override;
920
921 void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
922 SelectionDAG &DAG) const override;
923 void ReplaceBITCASTResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
924 SelectionDAG &DAG) const;
925 void ReplaceExtractSubVectorResults(SDNode *N,
926 SmallVectorImpl<SDValue> &Results,
927 SelectionDAG &DAG) const;
928 void ReplaceGetActiveLaneMaskResults(SDNode *N,
929 SmallVectorImpl<SDValue> &Results,
930 SelectionDAG &DAG) const;
931
932 bool shouldNormalizeToSelectSequence(LLVMContext &, EVT, EVT) const override;
933
934 void finalizeLowering(MachineFunction &MF) const override;
935
936 bool shouldLocalize(const MachineInstr &MI,
937 const TargetTransformInfo *TTI) const override;
938
939 bool SimplifyDemandedBitsForTargetNode(SDValue Op,
940 const APInt &OriginalDemandedBits,
941 const APInt &OriginalDemandedElts,
942 KnownBits &Known,
943 TargetLoweringOpt &TLO,
944 unsigned Depth) const override;
945
946 bool canCreateUndefOrPoisonForTargetNode(
947 SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
948 UndefPoisonKind Kind, bool ConsiderFlags, unsigned Depth) const override;
949
950 bool isTargetCanonicalConstantNode(SDValue Op) const override;
951
952 // With the exception of data-predicate transitions, no instructions are
953 // required to cast between legal scalable vector types. However:
954 // 1. Packed and unpacked types have different bit lengths, meaning BITCAST
955 // is not universally useable.
956 // 2. Most unpacked integer types are not legal and thus integer extends
957 // cannot be used to convert between unpacked and packed types.
958 // These can make "bitcasting" a multiphase process. REINTERPRET_CAST is used
959 // to transition between unpacked and packed types of the same element type,
960 // with BITCAST used otherwise.
961 // This function does not handle predicate bitcasts.
962 SDValue getSVESafeBitCast(EVT VT, SDValue Op, SelectionDAG &DAG) const;
963
964 // Returns the runtime value for PSTATE.SM by generating a call to
965 // __arm_sme_state.
966 SDValue getRuntimePStateSM(SelectionDAG &DAG, SDValue Chain, SDLoc DL,
967 EVT VT) const;
968
969 bool preferScalarizeSplat(SDNode *N) const override;
970
971 unsigned getMinimumJumpTableEntries() const override;
972
973 bool shouldScalarizeBinop(SDValue VecOp) const override {
974 return VecOp.getOpcode() == ISD::SETCC;
975 }
976
977 bool hasMultipleConditionRegisters(EVT VT) const override {
978 return VT.isScalableVector();
979 }
980};
981
982namespace AArch64 {
983FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
984 const TargetLibraryInfo *libInfo,
985 const LibcallLoweringInfo *libcallLowering);
986
987// Determine the effective TLS model for an ELF global, applying
988// AArch64-specific restrictions and configuration.
989TLSModel::Model getELFTLSModel(const GlobalValue *GV, const TargetMachine &TM,
990 bool HasELFSignedGOT);
991} // end namespace AArch64
992
993} // end namespace llvm
994
995#endif
996